Research Article Comparison of the Manual, Semiautomatic...
Transcript of Research Article Comparison of the Manual, Semiautomatic...
![Page 1: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/1.jpg)
Research ArticleComparison of the Manual Semiautomatic and AutomaticSelection and Leveling of Hot Spots in Whole Slide Images forKi-67 Quantification in Meningiomas
Zaneta Swiderska1 Anna Korzynska2 Tomasz Markiewicz13
Malgorzata Lorent3 Jakub Zak12 Anna Wesolowska2 Lukasz Roszkowiak2
Janina Slodkowska3 and Bartlomiej Grala3
1Warsaw University of Technology Pl Politechniki 1 00-661 Warsaw Poland2Nalecz Institute of Biocybernetics and Biomedical Engineering PAS Trojdena 4 02-109 Warsaw Poland3Military Institute of Medicine Szaserow 128 04-141 Warsaw Poland
Correspondence should be addressed to Zaneta Swiderska swiderszeepwedupl
Received 29 April 2015 Revised 10 June 2015 Accepted 14 June 2015
Academic Editor Sebastian Wachsmann-Hogiu
Copyright copy 2015 Zaneta Swiderska et al This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited
Background This paper presents the study concerning hot-spot selection in the assessment of whole slide images of tissue sectionscollected frommeningioma patientsThe samples were immunohistochemically stained to determine theKi-67MIB-1 proliferationindex used for prognosis and treatment planning Objective The observer performance was examined by comparing results of theproposed method of automatic hot-spot selection in whole slide images results of traditional scoring under a microscope andresults of a pathologistrsquos manual hot-spot selection Methods The results of scoring the Ki-67 index using optical scoring under amicroscope software for Ki-67 index quantification based on hot spots selected by two pathologists (resp once and three times)and the same software but on hot spots selected by proposed automatic methods were compared using Kendallrsquos tau-b statisticsResults Results show intra- and interobserver agreement The agreement between Ki-67 scoring with manual and automatic hot-spot selection is high while agreement between Ki-67 index scoring results in whole slide images and traditional microscopicexamination is lower Conclusions The agreement observed for the three scoring methods shows that automation of area selectionis an effective tool in supporting physicians and in increasing the reliability of Ki-67 scoring in meningioma
1 Introduction
Immunohistochemistry (IHC) has become an importanttechnique to both diagnostic pathology and clinical researchas it can help in the process of diagnosis prognosis andgrading [1] Furthermore during a personalized cancertreatment various molecular markers coupled with specificantibodies allow the pattern of the growth of certain tumorsand their response to the particular treatment to be pre-dicted For example the proliferation marker Ki-67 is usedin meningiomas to differentiate cancer into meningothelial(WHO I) atypical (WHO II) and anaplastic (WHO III)and correlates with tumor recurrences [1ndash5] This is because
the immunopositive signal expression is a surrogate measureof Ki-67 expression inside cellsrsquo nuclei According to theWorld Health Organization (WHO) rules the quantitativeevaluation of the proliferation index is performed on a setof high power areas of hot spots selected in various placesinside a whole specimen observed under a microscope Foreach chosen area of selection the number of immunopositiveand immunonegative cell nuclei is counted to establish theKi-67 index as the ratio of immunopositive cell nuclei to thewhole number of cell nucleiThis routine practice lacks repro-ducibility from observer to observer because this definition ishighly flexible By definition selected areas should representfields of high Ki-67 index in different tumor localizations
Hindawi Publishing CorporationAnalytical Cellular PathologyVolume 2015 Article ID 498746 15 pageshttpdxdoiorg1011552015498746
2 Analytical Cellular Pathology
The significant variability of possible selection leads to inter-and intraobserver variability in quantitative results whichshould be investigated in observer based assessment [6ndash16]
There have been many attempts to help histopathologistsin Ki-67 index quantification involving computers and digitalversions of the glass slide called the whole slide image (WSI)A review of papers concerning this subject published bothin the days when only small images could be handled bycomputers [17ndash24] and nowadays when WSIs are availableand computers or clusters of computers have the necessarycomputing power to manipulate them [25ndash34] shows thatinvestigators propose the use of computers on at least 3 levelsof the process of proliferation index quantification (1) inregion selection (2) in immunopositive and immunonegativecell nuclei selection and (3) in proliferation and other indexcounting While the third level is obvious and the second iswidely explored the first level is still poorly represented in theliterature There are methods of region selection concerningHematoxylin and Eosin staining [35ndash37] while for Ki-67stained with DAB and counterstained by Hematoxylin thereare the studies published by Potts [34] and coworkers Lu andcoworkers [35] and Gavrielides and coworkers [7 8] Thethird group of investigators performed a pooling study andconcluded that ldquo for validation study should be focusedon specific pathology tasks to eliminate sources of variabilitythat might dilute findingsrdquo So a validation study of a specificuse of Digital Pathology that is in the quantification of theproliferation index based on Ki-67 used in meningiomas ispresented in this paper
2 Materials and Methods
21 Glass Slide Preparation The glass slides used in thisstudy came from meningioma patients diagnosed or gradedat theDepartment of Pathomorphology theMilitary Instituteof Medicine in Warsaw Poland They were divided intotwo sets of data according to two methods of preparationIn set A there were twenty-three glass slides (57 13patients in grade I 30 7 patients in grade II and 133 patients in grade III according to WHO scores) preparedfrom paraffin blocks which had been randomly chosen withrespect to quality from the hospital archivesThe Ki-67MIB-1 immunohistochemical stained procedure was performedusing a Dako Autostainer Link and the following chemicalFLEXMonoclonal Mouse Anti-Human Ki-67 Antigen CloneMIB-1 Ready-to-Use (Link) reference number IR626 fromDako The staining was visualized using EnVision FLEX Tar-get Retrieval Solution from Dako according to the proceduredescribed in the user manual All manual and mechanicalactivities were performed very carefully because the sampleswere supposed to bemodel quality in comparison to the slidesfrom set B
In set B twenty-seven glass slides (70 19 patients ingrade I and 30 8 patients in grade II) from routine hospitalprognoses and grading using Ki-67MIB were chosen to beinvolved in the study All these slides had been preparedbetween 2011 and 2014 with or without Autostainer Linkin a manual procedure using various chemicals purchasedfrom Dako Set B contained inhomogeneousWSI in terms of
both the manner of preparation and the chemicals used Theoverall quality of glass slides from set B was worse than thatof glass slides from set A
22 Microscope and Monitor Review of the Digitalized GlassSlides The sets of glass slides were both scored by an expe-rienced pathologist henceforth known as expert using anOlympus BX40 optical microscope with PlanApo objectiveThen the slides were digitalized using an Aperio ScanScopescanner for set A and a 3DHISTECH Panoramic II for setB These were then reviewed on a calibrated EIZO FlexS-can 22-inch monitor The WSIs were acquired under 400xmagnification with a resolution of 0279 120583m and 038895 120583mper pixel for sets A and B respectively Digital images werereviewed using dedicated software prepared according toproject requirements which allowed panning around with amousetrackball to view the WSI in various magnificationsand to mark fields of quantification This software wasprepared in MATLAB using library Open Slide [38] to readWSI files
To ensure comparability of an area examined by anexpert under a microscope as one field of view and areaof quantification chosen from digital WSI the size of therectangle which covered the same area as the microscopiccircular field of view was determined It was assumed thatthemicroscopic field of view at 400xmagnification representsaround 012mm2 of a tissue the size of the digitized field ofview was 1424 times 1064 pixels in set A and 1024 times 766 pixels inset B
23 Observer Training and Environmental Adjustments Twopathologists with 7 and 3 years of practice in meningiomasections quantification were asked to support this study Tominimize sources of variability both observers were trainedon the software they were to use and their environments werecontrolled they used the same computer monitor and lightin the room in order to eliminate environmental influenceson the pathologists work
The pathologists had an introductory session to becomefamiliar with all the controls and interfaces which werenecessary in the selection of hot spots and proper size areasfor quantification by automatic software Pathologists hadbeen instructed the following
(i) The interpretation of Ki-67 does not include theclassification of the intensity of staining but thepercentage of tumor cells with positive staining
(ii) They should find 20 areas of the size mentionedabove with high populations of brown objects incomparison to the nearest neighborhoods but theseareas should be distributed among all hot spots whichcould be found in WSI
(iii) Each area should be at least 80 covered by tumorlesion and without any artifacts
Cases where even one of pathologists was unable to score(because of a lack or inadequacy of the region of a hotspot) were removed from the analysis During the areaselection the leader of the project assisted the pathologists by
Analytical Cellular Pathology 3
offering hardware and software support but did not make anysuggestions as to how to gather information about hot spotsor how to choose areas for quantification
24 Textural FeaturesApplied in the ProposedMethod Tofindhot-spot localizations a texture analysis was performed onWSI The normalized probabilities
119875
119904(119894) and
119875
119889(119894) of the 119894th
intensity on the basis of histograms of the sum and differenceimages [37] were used These images were formed from theoriginal image by applying the relative translation (1198891 1198892) Let119891
119896119897mean the intensity of a pixel at (119896 119897)th position in the gray
scale (each of RGB channels) and the image was translated bya fixed displacement (1198891 1198892)
119904
119896119897= 119891
119896119897+119891
119896+1198891 119897+1198892
119889
119896119897= 119891
119896119897minus119891
119896+1198891 119897+1198892
(1)
where 119904 and 119889 represent the sum and difference images Thenormalized sum and difference probabilities were estimatedby
119875
119904(119894) =
ℎ
119904(119894)
119873
119875
119889(119894) =
ℎ
119889(119894)
119873
(2)
where 119873 is the total number of pixels in the image Weused the modified formulas of Unser features [37] which arepresented in Table 1They were applied over a given regionΩ
associated with each pixel of the image In our notation 119873Ω
represents the total number of pixels in Ω region and 119904(x)and 119889(x) represent the pixel values of the sum and differenceimages
The determination of the image resolution and Ω radiuswhich allow the best characterization of the local structuresin images was achieved
The texture analysis was performed in the following stepsThe sum and difference images on the basis of the originalimage and the original image translated by 3 pixels werecalculated for each of the RGB channelsThen the diskmaskswith a radius of 10 pixels selected the set of the neighborhoodregion masks for each pixel location For a neighborhoodof size of 5 8 10 12 15 and 20 pixels the radius size of 10pixels appears to be the best and this was used in furtherexperiments
For the texture features defined in Table 1 the com-putation complexity problems were obvious These wereassociated with the traveling location of the central pixel andits neighboring region Ω This was solved by applying thearray operations The process of adding the pixel values insum and difference images was realized quickly by applyingthe average filtering of the image (embedded imfilter functionin MATLAB) Thereby the mean mask for the whole imagecould be calculated in only one analysis The (119896 119897)th coor-dinate of this mask represented the region center located inthis point To efficiently implement this method of featurecalculation the array form of operations was applied Forexample the variance feature (the second row in Table 1)
Table 1 Modified definitions of Unser features
Name Modified computational formula
Mean 1198911 =
sumxisinΩ 119904 (x)2119873Ω
= 120583
Ω
Variance 1198912 =
12(
sumxisinΩ (119904 (x) minus 2120583Ω)
2+ sumxisinΩ 119889 (x)2
119873
Ω
)
Energy 1198913 =
sumxisinΩ 119904 (x)2sdot sumxisinΩ 119889(x)
2
119873
2Ω
Correlation 1198914 =
12(
sumxisinΩ (119904 (x) minus 2 sdot 120583
Ω)
2minus sumxisinΩ 119889 (x)2
119873
Ω
)
Contrast 1198915 =
sumxisinΩ 119889(x)2
119873
Ω
Homogeneity 1198916 =
sumxisinΩ (11 + 119889(x)2)119873
Ω
Cluster shade 1198917 =
sumxisinΩ (119904 (x) minus 2120583Ω)
3
119873
Ω
Clusterprominence 1198918 =
sumxisinΩ (119904 (x) minus 2120583Ω)
4
119873
Ω
could be computed according to the following (modified)expression
1198912
=
sumxisinΩ 119904 (x)2minus 4120583ΩsumxisinΩ 119904 (x) + 4120583
Ω
2119873
Ω+ sumxisinΩ 119889 (x)2
2119873Ω
(3)
The first term of this relation was calculated by applyingthe filtering of the array-squared sum image (the Hadamardproduct) and the second by array-fashion multiplication ofthemean of the image and the filtered sum image In the sameway the other terms were calculated Thereby the texturefeature computation time has been significantly decreased
25 Automatic Hot Spot and Area of Quantification Selec-tion The proposed method for the hot-spot localizationand area of quantification selection based on mathematicalmorphology texture classification and controlled dispersionwas described in this section
An analysis of the information contained in WSI after aresolution decrease on various scales showed that the texturein the original image is redundant and the resolution can bedecreased To localize hot spots information about the ratioof brown (red) to blue pixels as a basic feature and some otherfeatures described below were needed All features were alsovisible in images with the resolution decreased by up to 8xwhile at a 16x decrease they were not visibleThis is presentedin Figure 1
It appears that an eightfold reduction of the resolutiondoes not disturb the required further textural features (sizeof object-cell nuclei is decreased from 128 plusmn 51 for brownand 102 plusmn 73 for blue in original image to 18 plusmn 9 and 10 plusmn
6 for the selected 8x decreased resolution resp) and enablesthe evaluation to be performed by a computer and by apathologist with a direct visual examination
4 Analytical Cellular Pathology
1x 2x 4x 8x 16x
Figure 1 Fragments of WSI in original and decreased resolution 2x 4x 8x and 16x
The proposed method of analysis of WSI in decreasedresolutions uses the following steps (1) the specimen map isestablished (2) the texture quantification and classificationare done to eliminate hemorrhage areas from the specimenmap (3) the hot spots are detected and finally (4) basedon the proposed penalty function selection of the area ofquantification inside selected hot spots is performed Thegeneral schema of the algorithm for steps 1ndash3 is presented inFigure 2
In the first step a map of the specimen was created usingthe thresholding procedure and morphological filtering [39ndash42] To do this a whole slide image was used to produce asupported image by the morphological operation of openingand brightness equalization This was performed using astructuring element shaped like a disk with a large radius(100 pixels) The operation of the division of each RGB colorcomponent of the image by its version after morphologicalopening was performed independently for each channelAfterwards components from channels B and R were pro-cessed with the Otsu thresholdingmethod [43] Additionallymorphological operations such as erosion dilatation andhole filling were performed to filter the specimen map
The next step which eliminates hemorrhage areas fromthe specimen map was performed by differentiating thetumor area from hemorrhage areas using texture analysis andclassification The local textural descriptors came from theUnser features [33 40] and were applied independently forRGB and CMYK color channels and also for the combined u(from CIE Luv) and C (form CMYK) representation A setof 64 textures was created as 8 features defined in Table 1by 8 color channels or sums of channels as presented in
Table 2 Next based on Fisherrsquos linear discriminant the mostsignificant 25were selected on a teaching phase and then usedin the classification phases (see Table 2)
Finally the Support Vector Machine (SVM) with Gaus-sian kernel function [41ndash46] was applied as a classifier torecognize the hemorrhage areas and to eliminate them fromthe specimen map
The third step of the algorithmwas an estimate of the localdensity of immunopositive cells using the reduced resolutionWSI The local maxima of the immunopositive cell densitiesare hot spots To select these the mathematical morphologyand proportion of the color components were used It wasfound that u of the CIE Luv representation of colors is strictlyassociated with the red color and can be used to differentiatethe immunopositive cells from the remainder of the imageThe extended regional minima transformation is applied toevaluate the spatial relation of the stained brown objects totheir neighboring environmentThe densitymapwas createdbased on the isolatedmarks representing the immunoreactivetumor cells
The fourth and final step of the proposedmethod focusedon the fields of quantification selection based on an artificialmodel of field spatial dispersion To prevent all fields ofquantification being chosen the penalty functionwas definedfrom one large dominant hot spot with a high Ki-67 index bythe following formula
penalty = 1minus120588sum
119894
1
(
radic
(119909 minus 119909
119894)
2+ (119910 minus 119910
119894)
2)
05 (4)
Analytical Cellular Pathology 5
Table2Sign
ificant
texturefeatures(features
in8colorc
hann
elso
rsum
ofchannels(a)colorc
hann
elso
rsum
ofchannelsin
features
(b))
(a)
Color
compo
nent
R(1ndash
8)G
(9ndash16)
B(17ndash24)
u(lu
v)+C
(CMYK
)(25ndash32)
C(33ndash40
)M
(41ndash48)
Y(49ndash
56)
K(57ndash64
)
Num
bero
ffeatures
35
1113
mdash252627293032
333435373840
4142434546
48
4954
63
Featuren
ame
(i)En
ergy
(ii)C
ontrast
(i)En
ergy
(ii)C
ontrast
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)
Hom
ogeneity
Clustershade
(b)
12
34
56
78
PSM
Varia
nce
Energy
Correlatio
nCon
trast
Hom
ogeneity
Clustershade
Clusterp
rominence
(u+C)
CM
Y(u
+C)
CM
RG(u+C)
CM
mdashR
G(u+C)
CM
(u+C)
CM
YK
(u+C)
CM
6 Analytical Cellular Pathology
WSI image (8-fold reduction) Channel R
Channel B
Difference map
Binary map
Otsu thresholding
Specimen map
Morphological
filterin
g
RGB CMYK Luv
Texture Texture Texture
Feature selection
SVM classifier
Hemorrhage and
Map without hemorrhages Immunopositive cells inside Hot-spot selection
Cell recognition for one hot-
WSI file with full resolutionreading the selected hot-spot
field of view
Select
ed field
locat
ion
(RGB)
(RGB)
the map spot
background map
Division of f by 120601(f)
Division of f by 120601(f)
Figure 2 The schema of the algorithm for hot-spot localization
which was based on information about the distance betweenthe designated areas and the position of another candidate forhot spot
An increase in the 120588 value shows an increase in the scat-tering of the areas of quantification selectionThe 120588 value hadbeen chosen experimentally (see Section 3) The proposedfunction combined selection of fields of quantification fromdifferent localization in the specimen according to a gradualreduction of the concentration of immunopositive cellsHowever when hot-spot areas other than the dominant oneshow a significantly lower density of immunopositive nucleithe candidates from dominant region will still be selectedfirstThe final analysis of theKi-67 index in all chosen areas ofquantification was performed on full resolution images withthe method published earlier and described in [45]
26 Evaluation of the Concordance of Selected Hot-Spot FieldsTo evaluate the concordance of hot-spot field localizationbetween the expertsrsquo and automatic results the localizationconcordance measure (LCM) was proposed This measureassumed that (1) those fields at a shorter distance shouldhave a reduced impact on the LCM and (2) the significanceof fields should relate to their Ki-67 index The localization
concordancemeasure was calculated according to the follow-ing formula
LCM
= sum
119894
(119908
119894lowast sigm(min
119895
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
dist [(119909119895 119910
119895) (119909
119894 119910
119894)]
(4FOVsize)
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
))
where 119908119894=
119871
119864119894
119871
119864
(5)
in which 119871
119864is the level of the Ki-67 index for the expert
and FOVsize is a one field of view size A low value ofthe LCM shows a similarity in the areas of quantificationselection by algorithm and expert This means that theexpertrsquos selected fields of quantification are represented ornear the fields selected by the proposed method for examplethey represent the same tumor area If expert and algorithmselect fields of view from different virtual slide areas thelocalization measure LCM is high The proposed measureallows both the evaluation of the similarity of choice of hot-spot fields and the identification of the best penalty factorThe proposedmeasure can be used in cases of both inter- andintraobservation variability
Analytical Cellular Pathology 7
27 Study Design Both the proposed automated methodof area of quantification selection and the two pathologistswere used to review all the samples (A and B sets) usingdigital representation of glass slides with an 8x reducedWSI resolution while expert quantification was done infull resolution and for 10 areas of unknown location Theoutcomes were then averaged to give the final result
Each of the pathologists chose 20 fields of quantificationfor eachWSI One pathologist had two additional sessions fortheWSI from set A to estimate interobserver variability Eachadditional session was performed with at least a one-monthdelay between sessionsThe order of samples was randomizedfor each session
Then automated hot-spot selection software was used toselect 20 fields of quantification as an area which fulfill twocriteria
(i) Biggest number of immunopositive nuclei in compar-ison to the others
(ii) That distances between new and previously foundareas are large enough tomeet the requirements of theabove defined penalty function that is (4)
The scores of the Ki-67 index from the areas of quantifi-cation chosen by the pathologists and the automatic methodwere produced using software which segments nuclei insubimages from WSI in full resolution and then classifiesthem into immunopositive and immunonegative classes andestimates the Ki-67 index This software was published in2009 [45] by the principal investigator of the project who iscoauthor of this paperThe ratio of immunopositive nuclei tothe number of all cells in each area of quantification and themean of these ratios for each WSI were sent for computer-enabled statistical analysis
28 Statistical Analysis The scores from the expert micro-scopic examination and all automatic scores from areaschosen by both pathologists and scores from the proposedautomatic method of area of quantification selection wereanalyzed using agreement analysis since IHC interpretationis a subjective process of evaluation For this process a truescore is not available Besides agreement between digital andoptical scores was not the primary objective of the study butrather this is considered the reference value while agreementbetween the automatic and pathologistsrsquo hot spots and area ofquantification selection in IHC assessment was the main aimof the study
The primary objective of the investigationwas to find pat-terns of agreement between manual human and automatedselection of the area of quantification inWSIThe commonlyused concordance measure Kendallrsquos tau-b was used as inGavrielides and coworkers [7 8] The test was calculatedseparately for setsA andB in pooled and categorizedgroupedversions in both pairwise and cumulated versions
Kendallrsquos tau-b is a rank-based correlation metric whichcalculates the difference between the rate of concordanceand discordance [46ndash48] The range of Kendallrsquos tau-b is minus1to 1 where 1 indicates perfect agreement minus1 indicates dataare inverted (perfect agreement inversion) and 0 indicates
no relationship Kendallrsquos tau-b was computed according toBalboaca and Jantschi [48] using dedicated software preparedin MATLAB
Kendallrsquos tau-b values were utilized to quantify interob-server and intraobserver agreements The interobserver agree-ment was estimated between all pairs (1) between patholo-gists themselves ((2) and (3)) between each pathologist andclassical expert microscopic reviewing ((4) and (5)) betweeneach pathologist and the proposed automated method andadditionally in a grouped version between (6) the mean ofpathologist and classical expert microscopic reviewing and(7) the mean of the pathologist and proposed automatedmethod applied to WSI and (8) between classical expertmicroscopic reviewing and proposed automated methodapplied to WSI The intraobserver agreement (agreementbetween the scores of the same observer in various sessionsof area selection) was estimated between all pairs of threeindependent scorings from one pathologist (1) the first andthe second scores (2) the second and the third scores and(3) the first and the third scores both in pooled data and incategorized data Because of the small number of WSIs frompatients in grade III the results relate to only two categoriesgrade I and grade II in diagnosed meningioma patientsConfidence intervals for the overall agreement measureswere calculated applying bootstrap analysis using a proceduredescribed in detail in the study by Gavrielides et al [7]
Software for bootstrap was implemented using MATLAB(MathWorks Natick MA USA) functions
3 Results
First the influence of the 120588 value on the penalty factorwas examined A subset of twelve WSIs from set A waschosen for this analysis The hot-spot localization and areasof quantification selections performed by pathologists andthe automated proposed method were compared using theLCM measure The results of Ki-67 index estimations andLCM measures for 120588 from 01 to 05 with an increment of005 are presented in Figure 3The best concordance betweenthe automatically selected areas of quantification and thoseselected by pathologists is for a 120588 value equal to 02 For this 120588value LCM is minimal for a relatively high value of the Ki-67index
The dispersion of areas of quantification chosen bythe proposed automatic method and pathologists can beobserved in Figures 4 and 5
Figure 4 presents the distribution of the areas of quan-tification selected in hot spots found for two WSIs bytwo pathologists (red and yellow rectangles) and by theproposed automatic method (black rectangles) In the topline (Figures 4(a) and 4(b)) it can be observed that thereis no agreement between both pathologists and that thedistribution of regions is different and inhomogeneous so themeasure of concordance LCM is 9 The distance betweenthe proposed automatic method and the mean measurefor both pathologists is 86 The bottom line (Figures 4(c)and 4(d)) shows good agreement in areas of quantificationdistribution Their measure of concordance LCM is 29
8 Analytical Cellular Pathology
4
5
6
7
8
9
10
202224262830323436
01 015 02 025 03 04 05
LCM
Ki-6
7 in
dex
()
Mean for AUMean for humanLCM
120588
Figure 3 Examination of Ki-67 index and LCM in respect to the 120588
factor of the penalty function
between pathologists while between the proposed automaticmethod and the mean measure for both pathologists it is 34
Figure 5 presents results of three repetitions of the selec-tions of the areas of quantification from one pathologist(Figures 5(a) 5(b) and 5(c) blind trial) and from automatedmethod (Figure 5(d)) using one of the WSIs from set A It isvisible that this one person chose a region of quantificationin various parts of specimen The third trial is significantlydifferent from the previous two rather similar trials but theKi-67 indexes for each of them are similar (107 106 and119 for pathologist and 132 for automated method)
As the selected fields of quantification were previouslyquantified by the software the Ki-67 index for each area andfor each specimen becomes the data for statistical analysis
In Figure 6 where all results for Ki-67 quantificationusing all early interdicted methods of its estimation (man-ually experts by means of two pathologistsrsquo semiauto-matic approach and fully automatic approach) are shownthe general tendency for a relationship between them canbe observed Quantification with the manual microscopicmethod produces the lowest Ki-67 proliferation index valueswhile the automatic methods produce the highest values ofthis index This pattern is biased by one specimen from setA and for 3 specimens from set B In the first case thelower result for the pathologists is caused by an undervaluedscore from one pathologist In the other cases from set Bthe pattern is reversed and the highest values for this indexappear for the manual microscopic method Expose controlof WSI shows that there are very small hot spots in each ofthese three specimens It seems that when hot spots do notcover the whole area of quantification (although they fulfillthe criterion that about 80 of the area of quantificationshould be covered by a hot spot) it causes various resultsfrom the pathologists and automatic method In such casesan expert performing microscopic scoring used to deal withpart of field of quantification restricted to hot spots while theautomatic method diluted the score by counting the numberof cells from the whole area of the rectangle
The results of inter- and intraobserver variability mea-surements are presented in Tables 3 4 and 5
Table 3 presents the results of pairwise agreement usingKendallrsquos tau-b analysis for interobserver variability as acoefficient of concordance along with confidence intervals(95 confidence level) constructed using bootstrap analysisof the samples (100 order changes) It can be observed that allconcordance between all pairs calculated for set A is biggerthan the analogous values for set B This can be explained bythe fact that WSIs in set A were prepared using autostainerand the same set of new chemicals while WSIs in set B wereregular glass slides prepared earlier some with and somewithout autostainer and using the chemicals available at thetime Visual examination shows that glasses in set A are ofreally good quality and homogenous in performance whilethe glasses from set B are not This inhomogeneity amongglasses from set B led to differing interpretations by the twopathologists which is seen as a decrease in agreement betweenthem (from 092 to 086) and between each of them andthe automated method (from 082 and 081 to 078 and 076resp)
Both parts of Table 3 A and B show an overall tendencyfor the highest correlation to be between both pathologistsand the lowest agreement to be between a classical micro-scopic expert scoring and the proposed automated methodemployed on WSI The concordance between both pathol-ogists and the other two methods of scoring are betweenthese two extremes but the concordance is greater betweenthe pathologist and the proposed automated method thanbetween the pathologists and the classical microscopic score
Table 4 presents the results of pooled expert agreement intwo categories grade I and grade II (WHO categorization ofmeningioma) usingKendallrsquos tau-b analysis for interobservervariability The coefficient of concordance is presented withconfidence intervals (95 confidence level) calculated usingbootstrap analysis of the samples (100 order changes) It canbe observed that concordance is higher in category grade IIthan in grade I in both subsets A and B This fact can beexplained as the reason that grade I patientsrsquo scores are usuallylower (up to 8 see Figure 4) than grade II patientsrsquo scores(up to 20 see Figure 4) This means that the hot spots aremore intensive and visible in comparison to the surroundingspace in those specimens from patients diagnosed as gradeII This visibility is more important for the pathologiststhan for the automated method So coincidence betweenpathologistsrsquo digital scoring and the proposed automatedmethod is very high (086 for set A and 08 for set B) ifgrade II patientsrsquo sections are analyzed while for grade Ipatientsrsquo sections the coincidence is lower (079 and 078)but it still shows coincidence For the coincidence betweenexpertsrsquo digital scoring and the manual microscopic expertscoring the results show similar patterns but the numbers arelower (087 043 and 06 and 03 resp)
The coincidence between the results of the proposedautomatic method and manual microscopic expert scoring isambiguousThis coincidence is rather low except for grade IIof set A
Table 5 presents the results of pairwise agreement analysisfor uncategorized and categorized data using Kendallrsquos tau-bfor intraobserver variability One of the pathologists repeatedthe scoring procedure 3 times with a delay long enough
Analytical Cellular Pathology 9
966 1088
372 165
176 147
(a)
990
310
179
(b)
1185 1212
743 924
655 578
(c)
1308
860
577
(d)
Figure 4 TwoWSIs with an area of quantification inside a hot spot marked by two experts ((a) and (c) red and yellow rectangles) and chosenby the proposed automatic method ((b) and (d) black rectangles) with examples of 3 areas of quantification for each method and for bothWSIs with large and small variation in Ki-67 index value calculated as percentage of immunopositive nuclei to the whole number of nucleiin presented area
10 Analytical Cellular Pathology
1420
1253
703
(a)
1573
1274
655
(b)
1522
1518
703
(c)
1891
1500
880
(d)
Figure 5The results of hot spot and area of quantification selections by one pathologist ((a) (b) and (c)) and by proposed automatedmethod(d) with examples of 3 areas of quantification chosen approximately in the same region (hot spot) for each selection The Ki-67 index valuecalculated as percentage of immunopositive nuclei to the whole number of nuclei in presented area differs slightly
Table 3The results of the pairwise agreement analysis of data without taking into account categorical information produced using Kendallrsquostau-b analysis for interobserver variability on the WSI from sets A and B separately TM means classical expert microscopic review P1 andP2 mean pathologists and AU means proposed automatic method
TM P1 P2 AUA
TM 085281 (078355 087879) 079221 (071429 082684) 068831 (060173 068831)P1 085281 (078355 087619) 092208 (088745 093074) 081818 (076623 081818)P2 079221 (071429 082684) 092208 (088745 093074) 080952 (075758 080952)AU 068831 (060173 068831) 081818 (076623 081818) 080952 (075758 080952)
BTM 067687 (065 070017) 058503 (055629 060697) 054762 (051888 056871)P1 067687 (064966 070034) 086735 (085782 087398) 078231 (077262 078912)P2 058503 (055629 060714) 086735 (085748 087381) 07551 (075 076531)AU 054762 (051871 056871) 078231 (077228 078912) 07551 (075 076531)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 2: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/2.jpg)
2 Analytical Cellular Pathology
The significant variability of possible selection leads to inter-and intraobserver variability in quantitative results whichshould be investigated in observer based assessment [6ndash16]
There have been many attempts to help histopathologistsin Ki-67 index quantification involving computers and digitalversions of the glass slide called the whole slide image (WSI)A review of papers concerning this subject published bothin the days when only small images could be handled bycomputers [17ndash24] and nowadays when WSIs are availableand computers or clusters of computers have the necessarycomputing power to manipulate them [25ndash34] shows thatinvestigators propose the use of computers on at least 3 levelsof the process of proliferation index quantification (1) inregion selection (2) in immunopositive and immunonegativecell nuclei selection and (3) in proliferation and other indexcounting While the third level is obvious and the second iswidely explored the first level is still poorly represented in theliterature There are methods of region selection concerningHematoxylin and Eosin staining [35ndash37] while for Ki-67stained with DAB and counterstained by Hematoxylin thereare the studies published by Potts [34] and coworkers Lu andcoworkers [35] and Gavrielides and coworkers [7 8] Thethird group of investigators performed a pooling study andconcluded that ldquo for validation study should be focusedon specific pathology tasks to eliminate sources of variabilitythat might dilute findingsrdquo So a validation study of a specificuse of Digital Pathology that is in the quantification of theproliferation index based on Ki-67 used in meningiomas ispresented in this paper
2 Materials and Methods
21 Glass Slide Preparation The glass slides used in thisstudy came from meningioma patients diagnosed or gradedat theDepartment of Pathomorphology theMilitary Instituteof Medicine in Warsaw Poland They were divided intotwo sets of data according to two methods of preparationIn set A there were twenty-three glass slides (57 13patients in grade I 30 7 patients in grade II and 133 patients in grade III according to WHO scores) preparedfrom paraffin blocks which had been randomly chosen withrespect to quality from the hospital archivesThe Ki-67MIB-1 immunohistochemical stained procedure was performedusing a Dako Autostainer Link and the following chemicalFLEXMonoclonal Mouse Anti-Human Ki-67 Antigen CloneMIB-1 Ready-to-Use (Link) reference number IR626 fromDako The staining was visualized using EnVision FLEX Tar-get Retrieval Solution from Dako according to the proceduredescribed in the user manual All manual and mechanicalactivities were performed very carefully because the sampleswere supposed to bemodel quality in comparison to the slidesfrom set B
In set B twenty-seven glass slides (70 19 patients ingrade I and 30 8 patients in grade II) from routine hospitalprognoses and grading using Ki-67MIB were chosen to beinvolved in the study All these slides had been preparedbetween 2011 and 2014 with or without Autostainer Linkin a manual procedure using various chemicals purchasedfrom Dako Set B contained inhomogeneousWSI in terms of
both the manner of preparation and the chemicals used Theoverall quality of glass slides from set B was worse than thatof glass slides from set A
22 Microscope and Monitor Review of the Digitalized GlassSlides The sets of glass slides were both scored by an expe-rienced pathologist henceforth known as expert using anOlympus BX40 optical microscope with PlanApo objectiveThen the slides were digitalized using an Aperio ScanScopescanner for set A and a 3DHISTECH Panoramic II for setB These were then reviewed on a calibrated EIZO FlexS-can 22-inch monitor The WSIs were acquired under 400xmagnification with a resolution of 0279 120583m and 038895 120583mper pixel for sets A and B respectively Digital images werereviewed using dedicated software prepared according toproject requirements which allowed panning around with amousetrackball to view the WSI in various magnificationsand to mark fields of quantification This software wasprepared in MATLAB using library Open Slide [38] to readWSI files
To ensure comparability of an area examined by anexpert under a microscope as one field of view and areaof quantification chosen from digital WSI the size of therectangle which covered the same area as the microscopiccircular field of view was determined It was assumed thatthemicroscopic field of view at 400xmagnification representsaround 012mm2 of a tissue the size of the digitized field ofview was 1424 times 1064 pixels in set A and 1024 times 766 pixels inset B
23 Observer Training and Environmental Adjustments Twopathologists with 7 and 3 years of practice in meningiomasections quantification were asked to support this study Tominimize sources of variability both observers were trainedon the software they were to use and their environments werecontrolled they used the same computer monitor and lightin the room in order to eliminate environmental influenceson the pathologists work
The pathologists had an introductory session to becomefamiliar with all the controls and interfaces which werenecessary in the selection of hot spots and proper size areasfor quantification by automatic software Pathologists hadbeen instructed the following
(i) The interpretation of Ki-67 does not include theclassification of the intensity of staining but thepercentage of tumor cells with positive staining
(ii) They should find 20 areas of the size mentionedabove with high populations of brown objects incomparison to the nearest neighborhoods but theseareas should be distributed among all hot spots whichcould be found in WSI
(iii) Each area should be at least 80 covered by tumorlesion and without any artifacts
Cases where even one of pathologists was unable to score(because of a lack or inadequacy of the region of a hotspot) were removed from the analysis During the areaselection the leader of the project assisted the pathologists by
Analytical Cellular Pathology 3
offering hardware and software support but did not make anysuggestions as to how to gather information about hot spotsor how to choose areas for quantification
24 Textural FeaturesApplied in the ProposedMethod Tofindhot-spot localizations a texture analysis was performed onWSI The normalized probabilities
119875
119904(119894) and
119875
119889(119894) of the 119894th
intensity on the basis of histograms of the sum and differenceimages [37] were used These images were formed from theoriginal image by applying the relative translation (1198891 1198892) Let119891
119896119897mean the intensity of a pixel at (119896 119897)th position in the gray
scale (each of RGB channels) and the image was translated bya fixed displacement (1198891 1198892)
119904
119896119897= 119891
119896119897+119891
119896+1198891 119897+1198892
119889
119896119897= 119891
119896119897minus119891
119896+1198891 119897+1198892
(1)
where 119904 and 119889 represent the sum and difference images Thenormalized sum and difference probabilities were estimatedby
119875
119904(119894) =
ℎ
119904(119894)
119873
119875
119889(119894) =
ℎ
119889(119894)
119873
(2)
where 119873 is the total number of pixels in the image Weused the modified formulas of Unser features [37] which arepresented in Table 1They were applied over a given regionΩ
associated with each pixel of the image In our notation 119873Ω
represents the total number of pixels in Ω region and 119904(x)and 119889(x) represent the pixel values of the sum and differenceimages
The determination of the image resolution and Ω radiuswhich allow the best characterization of the local structuresin images was achieved
The texture analysis was performed in the following stepsThe sum and difference images on the basis of the originalimage and the original image translated by 3 pixels werecalculated for each of the RGB channelsThen the diskmaskswith a radius of 10 pixels selected the set of the neighborhoodregion masks for each pixel location For a neighborhoodof size of 5 8 10 12 15 and 20 pixels the radius size of 10pixels appears to be the best and this was used in furtherexperiments
For the texture features defined in Table 1 the com-putation complexity problems were obvious These wereassociated with the traveling location of the central pixel andits neighboring region Ω This was solved by applying thearray operations The process of adding the pixel values insum and difference images was realized quickly by applyingthe average filtering of the image (embedded imfilter functionin MATLAB) Thereby the mean mask for the whole imagecould be calculated in only one analysis The (119896 119897)th coor-dinate of this mask represented the region center located inthis point To efficiently implement this method of featurecalculation the array form of operations was applied Forexample the variance feature (the second row in Table 1)
Table 1 Modified definitions of Unser features
Name Modified computational formula
Mean 1198911 =
sumxisinΩ 119904 (x)2119873Ω
= 120583
Ω
Variance 1198912 =
12(
sumxisinΩ (119904 (x) minus 2120583Ω)
2+ sumxisinΩ 119889 (x)2
119873
Ω
)
Energy 1198913 =
sumxisinΩ 119904 (x)2sdot sumxisinΩ 119889(x)
2
119873
2Ω
Correlation 1198914 =
12(
sumxisinΩ (119904 (x) minus 2 sdot 120583
Ω)
2minus sumxisinΩ 119889 (x)2
119873
Ω
)
Contrast 1198915 =
sumxisinΩ 119889(x)2
119873
Ω
Homogeneity 1198916 =
sumxisinΩ (11 + 119889(x)2)119873
Ω
Cluster shade 1198917 =
sumxisinΩ (119904 (x) minus 2120583Ω)
3
119873
Ω
Clusterprominence 1198918 =
sumxisinΩ (119904 (x) minus 2120583Ω)
4
119873
Ω
could be computed according to the following (modified)expression
1198912
=
sumxisinΩ 119904 (x)2minus 4120583ΩsumxisinΩ 119904 (x) + 4120583
Ω
2119873
Ω+ sumxisinΩ 119889 (x)2
2119873Ω
(3)
The first term of this relation was calculated by applyingthe filtering of the array-squared sum image (the Hadamardproduct) and the second by array-fashion multiplication ofthemean of the image and the filtered sum image In the sameway the other terms were calculated Thereby the texturefeature computation time has been significantly decreased
25 Automatic Hot Spot and Area of Quantification Selec-tion The proposed method for the hot-spot localizationand area of quantification selection based on mathematicalmorphology texture classification and controlled dispersionwas described in this section
An analysis of the information contained in WSI after aresolution decrease on various scales showed that the texturein the original image is redundant and the resolution can bedecreased To localize hot spots information about the ratioof brown (red) to blue pixels as a basic feature and some otherfeatures described below were needed All features were alsovisible in images with the resolution decreased by up to 8xwhile at a 16x decrease they were not visibleThis is presentedin Figure 1
It appears that an eightfold reduction of the resolutiondoes not disturb the required further textural features (sizeof object-cell nuclei is decreased from 128 plusmn 51 for brownand 102 plusmn 73 for blue in original image to 18 plusmn 9 and 10 plusmn
6 for the selected 8x decreased resolution resp) and enablesthe evaluation to be performed by a computer and by apathologist with a direct visual examination
4 Analytical Cellular Pathology
1x 2x 4x 8x 16x
Figure 1 Fragments of WSI in original and decreased resolution 2x 4x 8x and 16x
The proposed method of analysis of WSI in decreasedresolutions uses the following steps (1) the specimen map isestablished (2) the texture quantification and classificationare done to eliminate hemorrhage areas from the specimenmap (3) the hot spots are detected and finally (4) basedon the proposed penalty function selection of the area ofquantification inside selected hot spots is performed Thegeneral schema of the algorithm for steps 1ndash3 is presented inFigure 2
In the first step a map of the specimen was created usingthe thresholding procedure and morphological filtering [39ndash42] To do this a whole slide image was used to produce asupported image by the morphological operation of openingand brightness equalization This was performed using astructuring element shaped like a disk with a large radius(100 pixels) The operation of the division of each RGB colorcomponent of the image by its version after morphologicalopening was performed independently for each channelAfterwards components from channels B and R were pro-cessed with the Otsu thresholdingmethod [43] Additionallymorphological operations such as erosion dilatation andhole filling were performed to filter the specimen map
The next step which eliminates hemorrhage areas fromthe specimen map was performed by differentiating thetumor area from hemorrhage areas using texture analysis andclassification The local textural descriptors came from theUnser features [33 40] and were applied independently forRGB and CMYK color channels and also for the combined u(from CIE Luv) and C (form CMYK) representation A setof 64 textures was created as 8 features defined in Table 1by 8 color channels or sums of channels as presented in
Table 2 Next based on Fisherrsquos linear discriminant the mostsignificant 25were selected on a teaching phase and then usedin the classification phases (see Table 2)
Finally the Support Vector Machine (SVM) with Gaus-sian kernel function [41ndash46] was applied as a classifier torecognize the hemorrhage areas and to eliminate them fromthe specimen map
The third step of the algorithmwas an estimate of the localdensity of immunopositive cells using the reduced resolutionWSI The local maxima of the immunopositive cell densitiesare hot spots To select these the mathematical morphologyand proportion of the color components were used It wasfound that u of the CIE Luv representation of colors is strictlyassociated with the red color and can be used to differentiatethe immunopositive cells from the remainder of the imageThe extended regional minima transformation is applied toevaluate the spatial relation of the stained brown objects totheir neighboring environmentThe densitymapwas createdbased on the isolatedmarks representing the immunoreactivetumor cells
The fourth and final step of the proposedmethod focusedon the fields of quantification selection based on an artificialmodel of field spatial dispersion To prevent all fields ofquantification being chosen the penalty functionwas definedfrom one large dominant hot spot with a high Ki-67 index bythe following formula
penalty = 1minus120588sum
119894
1
(
radic
(119909 minus 119909
119894)
2+ (119910 minus 119910
119894)
2)
05 (4)
Analytical Cellular Pathology 5
Table2Sign
ificant
texturefeatures(features
in8colorc
hann
elso
rsum
ofchannels(a)colorc
hann
elso
rsum
ofchannelsin
features
(b))
(a)
Color
compo
nent
R(1ndash
8)G
(9ndash16)
B(17ndash24)
u(lu
v)+C
(CMYK
)(25ndash32)
C(33ndash40
)M
(41ndash48)
Y(49ndash
56)
K(57ndash64
)
Num
bero
ffeatures
35
1113
mdash252627293032
333435373840
4142434546
48
4954
63
Featuren
ame
(i)En
ergy
(ii)C
ontrast
(i)En
ergy
(ii)C
ontrast
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)
Hom
ogeneity
Clustershade
(b)
12
34
56
78
PSM
Varia
nce
Energy
Correlatio
nCon
trast
Hom
ogeneity
Clustershade
Clusterp
rominence
(u+C)
CM
Y(u
+C)
CM
RG(u+C)
CM
mdashR
G(u+C)
CM
(u+C)
CM
YK
(u+C)
CM
6 Analytical Cellular Pathology
WSI image (8-fold reduction) Channel R
Channel B
Difference map
Binary map
Otsu thresholding
Specimen map
Morphological
filterin
g
RGB CMYK Luv
Texture Texture Texture
Feature selection
SVM classifier
Hemorrhage and
Map without hemorrhages Immunopositive cells inside Hot-spot selection
Cell recognition for one hot-
WSI file with full resolutionreading the selected hot-spot
field of view
Select
ed field
locat
ion
(RGB)
(RGB)
the map spot
background map
Division of f by 120601(f)
Division of f by 120601(f)
Figure 2 The schema of the algorithm for hot-spot localization
which was based on information about the distance betweenthe designated areas and the position of another candidate forhot spot
An increase in the 120588 value shows an increase in the scat-tering of the areas of quantification selectionThe 120588 value hadbeen chosen experimentally (see Section 3) The proposedfunction combined selection of fields of quantification fromdifferent localization in the specimen according to a gradualreduction of the concentration of immunopositive cellsHowever when hot-spot areas other than the dominant oneshow a significantly lower density of immunopositive nucleithe candidates from dominant region will still be selectedfirstThe final analysis of theKi-67 index in all chosen areas ofquantification was performed on full resolution images withthe method published earlier and described in [45]
26 Evaluation of the Concordance of Selected Hot-Spot FieldsTo evaluate the concordance of hot-spot field localizationbetween the expertsrsquo and automatic results the localizationconcordance measure (LCM) was proposed This measureassumed that (1) those fields at a shorter distance shouldhave a reduced impact on the LCM and (2) the significanceof fields should relate to their Ki-67 index The localization
concordancemeasure was calculated according to the follow-ing formula
LCM
= sum
119894
(119908
119894lowast sigm(min
119895
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
dist [(119909119895 119910
119895) (119909
119894 119910
119894)]
(4FOVsize)
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
))
where 119908119894=
119871
119864119894
119871
119864
(5)
in which 119871
119864is the level of the Ki-67 index for the expert
and FOVsize is a one field of view size A low value ofthe LCM shows a similarity in the areas of quantificationselection by algorithm and expert This means that theexpertrsquos selected fields of quantification are represented ornear the fields selected by the proposed method for examplethey represent the same tumor area If expert and algorithmselect fields of view from different virtual slide areas thelocalization measure LCM is high The proposed measureallows both the evaluation of the similarity of choice of hot-spot fields and the identification of the best penalty factorThe proposedmeasure can be used in cases of both inter- andintraobservation variability
Analytical Cellular Pathology 7
27 Study Design Both the proposed automated methodof area of quantification selection and the two pathologistswere used to review all the samples (A and B sets) usingdigital representation of glass slides with an 8x reducedWSI resolution while expert quantification was done infull resolution and for 10 areas of unknown location Theoutcomes were then averaged to give the final result
Each of the pathologists chose 20 fields of quantificationfor eachWSI One pathologist had two additional sessions fortheWSI from set A to estimate interobserver variability Eachadditional session was performed with at least a one-monthdelay between sessionsThe order of samples was randomizedfor each session
Then automated hot-spot selection software was used toselect 20 fields of quantification as an area which fulfill twocriteria
(i) Biggest number of immunopositive nuclei in compar-ison to the others
(ii) That distances between new and previously foundareas are large enough tomeet the requirements of theabove defined penalty function that is (4)
The scores of the Ki-67 index from the areas of quantifi-cation chosen by the pathologists and the automatic methodwere produced using software which segments nuclei insubimages from WSI in full resolution and then classifiesthem into immunopositive and immunonegative classes andestimates the Ki-67 index This software was published in2009 [45] by the principal investigator of the project who iscoauthor of this paperThe ratio of immunopositive nuclei tothe number of all cells in each area of quantification and themean of these ratios for each WSI were sent for computer-enabled statistical analysis
28 Statistical Analysis The scores from the expert micro-scopic examination and all automatic scores from areaschosen by both pathologists and scores from the proposedautomatic method of area of quantification selection wereanalyzed using agreement analysis since IHC interpretationis a subjective process of evaluation For this process a truescore is not available Besides agreement between digital andoptical scores was not the primary objective of the study butrather this is considered the reference value while agreementbetween the automatic and pathologistsrsquo hot spots and area ofquantification selection in IHC assessment was the main aimof the study
The primary objective of the investigationwas to find pat-terns of agreement between manual human and automatedselection of the area of quantification inWSIThe commonlyused concordance measure Kendallrsquos tau-b was used as inGavrielides and coworkers [7 8] The test was calculatedseparately for setsA andB in pooled and categorizedgroupedversions in both pairwise and cumulated versions
Kendallrsquos tau-b is a rank-based correlation metric whichcalculates the difference between the rate of concordanceand discordance [46ndash48] The range of Kendallrsquos tau-b is minus1to 1 where 1 indicates perfect agreement minus1 indicates dataare inverted (perfect agreement inversion) and 0 indicates
no relationship Kendallrsquos tau-b was computed according toBalboaca and Jantschi [48] using dedicated software preparedin MATLAB
Kendallrsquos tau-b values were utilized to quantify interob-server and intraobserver agreements The interobserver agree-ment was estimated between all pairs (1) between patholo-gists themselves ((2) and (3)) between each pathologist andclassical expert microscopic reviewing ((4) and (5)) betweeneach pathologist and the proposed automated method andadditionally in a grouped version between (6) the mean ofpathologist and classical expert microscopic reviewing and(7) the mean of the pathologist and proposed automatedmethod applied to WSI and (8) between classical expertmicroscopic reviewing and proposed automated methodapplied to WSI The intraobserver agreement (agreementbetween the scores of the same observer in various sessionsof area selection) was estimated between all pairs of threeindependent scorings from one pathologist (1) the first andthe second scores (2) the second and the third scores and(3) the first and the third scores both in pooled data and incategorized data Because of the small number of WSIs frompatients in grade III the results relate to only two categoriesgrade I and grade II in diagnosed meningioma patientsConfidence intervals for the overall agreement measureswere calculated applying bootstrap analysis using a proceduredescribed in detail in the study by Gavrielides et al [7]
Software for bootstrap was implemented using MATLAB(MathWorks Natick MA USA) functions
3 Results
First the influence of the 120588 value on the penalty factorwas examined A subset of twelve WSIs from set A waschosen for this analysis The hot-spot localization and areasof quantification selections performed by pathologists andthe automated proposed method were compared using theLCM measure The results of Ki-67 index estimations andLCM measures for 120588 from 01 to 05 with an increment of005 are presented in Figure 3The best concordance betweenthe automatically selected areas of quantification and thoseselected by pathologists is for a 120588 value equal to 02 For this 120588value LCM is minimal for a relatively high value of the Ki-67index
The dispersion of areas of quantification chosen bythe proposed automatic method and pathologists can beobserved in Figures 4 and 5
Figure 4 presents the distribution of the areas of quan-tification selected in hot spots found for two WSIs bytwo pathologists (red and yellow rectangles) and by theproposed automatic method (black rectangles) In the topline (Figures 4(a) and 4(b)) it can be observed that thereis no agreement between both pathologists and that thedistribution of regions is different and inhomogeneous so themeasure of concordance LCM is 9 The distance betweenthe proposed automatic method and the mean measurefor both pathologists is 86 The bottom line (Figures 4(c)and 4(d)) shows good agreement in areas of quantificationdistribution Their measure of concordance LCM is 29
8 Analytical Cellular Pathology
4
5
6
7
8
9
10
202224262830323436
01 015 02 025 03 04 05
LCM
Ki-6
7 in
dex
()
Mean for AUMean for humanLCM
120588
Figure 3 Examination of Ki-67 index and LCM in respect to the 120588
factor of the penalty function
between pathologists while between the proposed automaticmethod and the mean measure for both pathologists it is 34
Figure 5 presents results of three repetitions of the selec-tions of the areas of quantification from one pathologist(Figures 5(a) 5(b) and 5(c) blind trial) and from automatedmethod (Figure 5(d)) using one of the WSIs from set A It isvisible that this one person chose a region of quantificationin various parts of specimen The third trial is significantlydifferent from the previous two rather similar trials but theKi-67 indexes for each of them are similar (107 106 and119 for pathologist and 132 for automated method)
As the selected fields of quantification were previouslyquantified by the software the Ki-67 index for each area andfor each specimen becomes the data for statistical analysis
In Figure 6 where all results for Ki-67 quantificationusing all early interdicted methods of its estimation (man-ually experts by means of two pathologistsrsquo semiauto-matic approach and fully automatic approach) are shownthe general tendency for a relationship between them canbe observed Quantification with the manual microscopicmethod produces the lowest Ki-67 proliferation index valueswhile the automatic methods produce the highest values ofthis index This pattern is biased by one specimen from setA and for 3 specimens from set B In the first case thelower result for the pathologists is caused by an undervaluedscore from one pathologist In the other cases from set Bthe pattern is reversed and the highest values for this indexappear for the manual microscopic method Expose controlof WSI shows that there are very small hot spots in each ofthese three specimens It seems that when hot spots do notcover the whole area of quantification (although they fulfillthe criterion that about 80 of the area of quantificationshould be covered by a hot spot) it causes various resultsfrom the pathologists and automatic method In such casesan expert performing microscopic scoring used to deal withpart of field of quantification restricted to hot spots while theautomatic method diluted the score by counting the numberof cells from the whole area of the rectangle
The results of inter- and intraobserver variability mea-surements are presented in Tables 3 4 and 5
Table 3 presents the results of pairwise agreement usingKendallrsquos tau-b analysis for interobserver variability as acoefficient of concordance along with confidence intervals(95 confidence level) constructed using bootstrap analysisof the samples (100 order changes) It can be observed that allconcordance between all pairs calculated for set A is biggerthan the analogous values for set B This can be explained bythe fact that WSIs in set A were prepared using autostainerand the same set of new chemicals while WSIs in set B wereregular glass slides prepared earlier some with and somewithout autostainer and using the chemicals available at thetime Visual examination shows that glasses in set A are ofreally good quality and homogenous in performance whilethe glasses from set B are not This inhomogeneity amongglasses from set B led to differing interpretations by the twopathologists which is seen as a decrease in agreement betweenthem (from 092 to 086) and between each of them andthe automated method (from 082 and 081 to 078 and 076resp)
Both parts of Table 3 A and B show an overall tendencyfor the highest correlation to be between both pathologistsand the lowest agreement to be between a classical micro-scopic expert scoring and the proposed automated methodemployed on WSI The concordance between both pathol-ogists and the other two methods of scoring are betweenthese two extremes but the concordance is greater betweenthe pathologist and the proposed automated method thanbetween the pathologists and the classical microscopic score
Table 4 presents the results of pooled expert agreement intwo categories grade I and grade II (WHO categorization ofmeningioma) usingKendallrsquos tau-b analysis for interobservervariability The coefficient of concordance is presented withconfidence intervals (95 confidence level) calculated usingbootstrap analysis of the samples (100 order changes) It canbe observed that concordance is higher in category grade IIthan in grade I in both subsets A and B This fact can beexplained as the reason that grade I patientsrsquo scores are usuallylower (up to 8 see Figure 4) than grade II patientsrsquo scores(up to 20 see Figure 4) This means that the hot spots aremore intensive and visible in comparison to the surroundingspace in those specimens from patients diagnosed as gradeII This visibility is more important for the pathologiststhan for the automated method So coincidence betweenpathologistsrsquo digital scoring and the proposed automatedmethod is very high (086 for set A and 08 for set B) ifgrade II patientsrsquo sections are analyzed while for grade Ipatientsrsquo sections the coincidence is lower (079 and 078)but it still shows coincidence For the coincidence betweenexpertsrsquo digital scoring and the manual microscopic expertscoring the results show similar patterns but the numbers arelower (087 043 and 06 and 03 resp)
The coincidence between the results of the proposedautomatic method and manual microscopic expert scoring isambiguousThis coincidence is rather low except for grade IIof set A
Table 5 presents the results of pairwise agreement analysisfor uncategorized and categorized data using Kendallrsquos tau-bfor intraobserver variability One of the pathologists repeatedthe scoring procedure 3 times with a delay long enough
Analytical Cellular Pathology 9
966 1088
372 165
176 147
(a)
990
310
179
(b)
1185 1212
743 924
655 578
(c)
1308
860
577
(d)
Figure 4 TwoWSIs with an area of quantification inside a hot spot marked by two experts ((a) and (c) red and yellow rectangles) and chosenby the proposed automatic method ((b) and (d) black rectangles) with examples of 3 areas of quantification for each method and for bothWSIs with large and small variation in Ki-67 index value calculated as percentage of immunopositive nuclei to the whole number of nucleiin presented area
10 Analytical Cellular Pathology
1420
1253
703
(a)
1573
1274
655
(b)
1522
1518
703
(c)
1891
1500
880
(d)
Figure 5The results of hot spot and area of quantification selections by one pathologist ((a) (b) and (c)) and by proposed automatedmethod(d) with examples of 3 areas of quantification chosen approximately in the same region (hot spot) for each selection The Ki-67 index valuecalculated as percentage of immunopositive nuclei to the whole number of nuclei in presented area differs slightly
Table 3The results of the pairwise agreement analysis of data without taking into account categorical information produced using Kendallrsquostau-b analysis for interobserver variability on the WSI from sets A and B separately TM means classical expert microscopic review P1 andP2 mean pathologists and AU means proposed automatic method
TM P1 P2 AUA
TM 085281 (078355 087879) 079221 (071429 082684) 068831 (060173 068831)P1 085281 (078355 087619) 092208 (088745 093074) 081818 (076623 081818)P2 079221 (071429 082684) 092208 (088745 093074) 080952 (075758 080952)AU 068831 (060173 068831) 081818 (076623 081818) 080952 (075758 080952)
BTM 067687 (065 070017) 058503 (055629 060697) 054762 (051888 056871)P1 067687 (064966 070034) 086735 (085782 087398) 078231 (077262 078912)P2 058503 (055629 060714) 086735 (085748 087381) 07551 (075 076531)AU 054762 (051871 056871) 078231 (077228 078912) 07551 (075 076531)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 3: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/3.jpg)
Analytical Cellular Pathology 3
offering hardware and software support but did not make anysuggestions as to how to gather information about hot spotsor how to choose areas for quantification
24 Textural FeaturesApplied in the ProposedMethod Tofindhot-spot localizations a texture analysis was performed onWSI The normalized probabilities
119875
119904(119894) and
119875
119889(119894) of the 119894th
intensity on the basis of histograms of the sum and differenceimages [37] were used These images were formed from theoriginal image by applying the relative translation (1198891 1198892) Let119891
119896119897mean the intensity of a pixel at (119896 119897)th position in the gray
scale (each of RGB channels) and the image was translated bya fixed displacement (1198891 1198892)
119904
119896119897= 119891
119896119897+119891
119896+1198891 119897+1198892
119889
119896119897= 119891
119896119897minus119891
119896+1198891 119897+1198892
(1)
where 119904 and 119889 represent the sum and difference images Thenormalized sum and difference probabilities were estimatedby
119875
119904(119894) =
ℎ
119904(119894)
119873
119875
119889(119894) =
ℎ
119889(119894)
119873
(2)
where 119873 is the total number of pixels in the image Weused the modified formulas of Unser features [37] which arepresented in Table 1They were applied over a given regionΩ
associated with each pixel of the image In our notation 119873Ω
represents the total number of pixels in Ω region and 119904(x)and 119889(x) represent the pixel values of the sum and differenceimages
The determination of the image resolution and Ω radiuswhich allow the best characterization of the local structuresin images was achieved
The texture analysis was performed in the following stepsThe sum and difference images on the basis of the originalimage and the original image translated by 3 pixels werecalculated for each of the RGB channelsThen the diskmaskswith a radius of 10 pixels selected the set of the neighborhoodregion masks for each pixel location For a neighborhoodof size of 5 8 10 12 15 and 20 pixels the radius size of 10pixels appears to be the best and this was used in furtherexperiments
For the texture features defined in Table 1 the com-putation complexity problems were obvious These wereassociated with the traveling location of the central pixel andits neighboring region Ω This was solved by applying thearray operations The process of adding the pixel values insum and difference images was realized quickly by applyingthe average filtering of the image (embedded imfilter functionin MATLAB) Thereby the mean mask for the whole imagecould be calculated in only one analysis The (119896 119897)th coor-dinate of this mask represented the region center located inthis point To efficiently implement this method of featurecalculation the array form of operations was applied Forexample the variance feature (the second row in Table 1)
Table 1 Modified definitions of Unser features
Name Modified computational formula
Mean 1198911 =
sumxisinΩ 119904 (x)2119873Ω
= 120583
Ω
Variance 1198912 =
12(
sumxisinΩ (119904 (x) minus 2120583Ω)
2+ sumxisinΩ 119889 (x)2
119873
Ω
)
Energy 1198913 =
sumxisinΩ 119904 (x)2sdot sumxisinΩ 119889(x)
2
119873
2Ω
Correlation 1198914 =
12(
sumxisinΩ (119904 (x) minus 2 sdot 120583
Ω)
2minus sumxisinΩ 119889 (x)2
119873
Ω
)
Contrast 1198915 =
sumxisinΩ 119889(x)2
119873
Ω
Homogeneity 1198916 =
sumxisinΩ (11 + 119889(x)2)119873
Ω
Cluster shade 1198917 =
sumxisinΩ (119904 (x) minus 2120583Ω)
3
119873
Ω
Clusterprominence 1198918 =
sumxisinΩ (119904 (x) minus 2120583Ω)
4
119873
Ω
could be computed according to the following (modified)expression
1198912
=
sumxisinΩ 119904 (x)2minus 4120583ΩsumxisinΩ 119904 (x) + 4120583
Ω
2119873
Ω+ sumxisinΩ 119889 (x)2
2119873Ω
(3)
The first term of this relation was calculated by applyingthe filtering of the array-squared sum image (the Hadamardproduct) and the second by array-fashion multiplication ofthemean of the image and the filtered sum image In the sameway the other terms were calculated Thereby the texturefeature computation time has been significantly decreased
25 Automatic Hot Spot and Area of Quantification Selec-tion The proposed method for the hot-spot localizationand area of quantification selection based on mathematicalmorphology texture classification and controlled dispersionwas described in this section
An analysis of the information contained in WSI after aresolution decrease on various scales showed that the texturein the original image is redundant and the resolution can bedecreased To localize hot spots information about the ratioof brown (red) to blue pixels as a basic feature and some otherfeatures described below were needed All features were alsovisible in images with the resolution decreased by up to 8xwhile at a 16x decrease they were not visibleThis is presentedin Figure 1
It appears that an eightfold reduction of the resolutiondoes not disturb the required further textural features (sizeof object-cell nuclei is decreased from 128 plusmn 51 for brownand 102 plusmn 73 for blue in original image to 18 plusmn 9 and 10 plusmn
6 for the selected 8x decreased resolution resp) and enablesthe evaluation to be performed by a computer and by apathologist with a direct visual examination
4 Analytical Cellular Pathology
1x 2x 4x 8x 16x
Figure 1 Fragments of WSI in original and decreased resolution 2x 4x 8x and 16x
The proposed method of analysis of WSI in decreasedresolutions uses the following steps (1) the specimen map isestablished (2) the texture quantification and classificationare done to eliminate hemorrhage areas from the specimenmap (3) the hot spots are detected and finally (4) basedon the proposed penalty function selection of the area ofquantification inside selected hot spots is performed Thegeneral schema of the algorithm for steps 1ndash3 is presented inFigure 2
In the first step a map of the specimen was created usingthe thresholding procedure and morphological filtering [39ndash42] To do this a whole slide image was used to produce asupported image by the morphological operation of openingand brightness equalization This was performed using astructuring element shaped like a disk with a large radius(100 pixels) The operation of the division of each RGB colorcomponent of the image by its version after morphologicalopening was performed independently for each channelAfterwards components from channels B and R were pro-cessed with the Otsu thresholdingmethod [43] Additionallymorphological operations such as erosion dilatation andhole filling were performed to filter the specimen map
The next step which eliminates hemorrhage areas fromthe specimen map was performed by differentiating thetumor area from hemorrhage areas using texture analysis andclassification The local textural descriptors came from theUnser features [33 40] and were applied independently forRGB and CMYK color channels and also for the combined u(from CIE Luv) and C (form CMYK) representation A setof 64 textures was created as 8 features defined in Table 1by 8 color channels or sums of channels as presented in
Table 2 Next based on Fisherrsquos linear discriminant the mostsignificant 25were selected on a teaching phase and then usedin the classification phases (see Table 2)
Finally the Support Vector Machine (SVM) with Gaus-sian kernel function [41ndash46] was applied as a classifier torecognize the hemorrhage areas and to eliminate them fromthe specimen map
The third step of the algorithmwas an estimate of the localdensity of immunopositive cells using the reduced resolutionWSI The local maxima of the immunopositive cell densitiesare hot spots To select these the mathematical morphologyand proportion of the color components were used It wasfound that u of the CIE Luv representation of colors is strictlyassociated with the red color and can be used to differentiatethe immunopositive cells from the remainder of the imageThe extended regional minima transformation is applied toevaluate the spatial relation of the stained brown objects totheir neighboring environmentThe densitymapwas createdbased on the isolatedmarks representing the immunoreactivetumor cells
The fourth and final step of the proposedmethod focusedon the fields of quantification selection based on an artificialmodel of field spatial dispersion To prevent all fields ofquantification being chosen the penalty functionwas definedfrom one large dominant hot spot with a high Ki-67 index bythe following formula
penalty = 1minus120588sum
119894
1
(
radic
(119909 minus 119909
119894)
2+ (119910 minus 119910
119894)
2)
05 (4)
Analytical Cellular Pathology 5
Table2Sign
ificant
texturefeatures(features
in8colorc
hann
elso
rsum
ofchannels(a)colorc
hann
elso
rsum
ofchannelsin
features
(b))
(a)
Color
compo
nent
R(1ndash
8)G
(9ndash16)
B(17ndash24)
u(lu
v)+C
(CMYK
)(25ndash32)
C(33ndash40
)M
(41ndash48)
Y(49ndash
56)
K(57ndash64
)
Num
bero
ffeatures
35
1113
mdash252627293032
333435373840
4142434546
48
4954
63
Featuren
ame
(i)En
ergy
(ii)C
ontrast
(i)En
ergy
(ii)C
ontrast
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)
Hom
ogeneity
Clustershade
(b)
12
34
56
78
PSM
Varia
nce
Energy
Correlatio
nCon
trast
Hom
ogeneity
Clustershade
Clusterp
rominence
(u+C)
CM
Y(u
+C)
CM
RG(u+C)
CM
mdashR
G(u+C)
CM
(u+C)
CM
YK
(u+C)
CM
6 Analytical Cellular Pathology
WSI image (8-fold reduction) Channel R
Channel B
Difference map
Binary map
Otsu thresholding
Specimen map
Morphological
filterin
g
RGB CMYK Luv
Texture Texture Texture
Feature selection
SVM classifier
Hemorrhage and
Map without hemorrhages Immunopositive cells inside Hot-spot selection
Cell recognition for one hot-
WSI file with full resolutionreading the selected hot-spot
field of view
Select
ed field
locat
ion
(RGB)
(RGB)
the map spot
background map
Division of f by 120601(f)
Division of f by 120601(f)
Figure 2 The schema of the algorithm for hot-spot localization
which was based on information about the distance betweenthe designated areas and the position of another candidate forhot spot
An increase in the 120588 value shows an increase in the scat-tering of the areas of quantification selectionThe 120588 value hadbeen chosen experimentally (see Section 3) The proposedfunction combined selection of fields of quantification fromdifferent localization in the specimen according to a gradualreduction of the concentration of immunopositive cellsHowever when hot-spot areas other than the dominant oneshow a significantly lower density of immunopositive nucleithe candidates from dominant region will still be selectedfirstThe final analysis of theKi-67 index in all chosen areas ofquantification was performed on full resolution images withthe method published earlier and described in [45]
26 Evaluation of the Concordance of Selected Hot-Spot FieldsTo evaluate the concordance of hot-spot field localizationbetween the expertsrsquo and automatic results the localizationconcordance measure (LCM) was proposed This measureassumed that (1) those fields at a shorter distance shouldhave a reduced impact on the LCM and (2) the significanceof fields should relate to their Ki-67 index The localization
concordancemeasure was calculated according to the follow-ing formula
LCM
= sum
119894
(119908
119894lowast sigm(min
119895
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
dist [(119909119895 119910
119895) (119909
119894 119910
119894)]
(4FOVsize)
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
))
where 119908119894=
119871
119864119894
119871
119864
(5)
in which 119871
119864is the level of the Ki-67 index for the expert
and FOVsize is a one field of view size A low value ofthe LCM shows a similarity in the areas of quantificationselection by algorithm and expert This means that theexpertrsquos selected fields of quantification are represented ornear the fields selected by the proposed method for examplethey represent the same tumor area If expert and algorithmselect fields of view from different virtual slide areas thelocalization measure LCM is high The proposed measureallows both the evaluation of the similarity of choice of hot-spot fields and the identification of the best penalty factorThe proposedmeasure can be used in cases of both inter- andintraobservation variability
Analytical Cellular Pathology 7
27 Study Design Both the proposed automated methodof area of quantification selection and the two pathologistswere used to review all the samples (A and B sets) usingdigital representation of glass slides with an 8x reducedWSI resolution while expert quantification was done infull resolution and for 10 areas of unknown location Theoutcomes were then averaged to give the final result
Each of the pathologists chose 20 fields of quantificationfor eachWSI One pathologist had two additional sessions fortheWSI from set A to estimate interobserver variability Eachadditional session was performed with at least a one-monthdelay between sessionsThe order of samples was randomizedfor each session
Then automated hot-spot selection software was used toselect 20 fields of quantification as an area which fulfill twocriteria
(i) Biggest number of immunopositive nuclei in compar-ison to the others
(ii) That distances between new and previously foundareas are large enough tomeet the requirements of theabove defined penalty function that is (4)
The scores of the Ki-67 index from the areas of quantifi-cation chosen by the pathologists and the automatic methodwere produced using software which segments nuclei insubimages from WSI in full resolution and then classifiesthem into immunopositive and immunonegative classes andestimates the Ki-67 index This software was published in2009 [45] by the principal investigator of the project who iscoauthor of this paperThe ratio of immunopositive nuclei tothe number of all cells in each area of quantification and themean of these ratios for each WSI were sent for computer-enabled statistical analysis
28 Statistical Analysis The scores from the expert micro-scopic examination and all automatic scores from areaschosen by both pathologists and scores from the proposedautomatic method of area of quantification selection wereanalyzed using agreement analysis since IHC interpretationis a subjective process of evaluation For this process a truescore is not available Besides agreement between digital andoptical scores was not the primary objective of the study butrather this is considered the reference value while agreementbetween the automatic and pathologistsrsquo hot spots and area ofquantification selection in IHC assessment was the main aimof the study
The primary objective of the investigationwas to find pat-terns of agreement between manual human and automatedselection of the area of quantification inWSIThe commonlyused concordance measure Kendallrsquos tau-b was used as inGavrielides and coworkers [7 8] The test was calculatedseparately for setsA andB in pooled and categorizedgroupedversions in both pairwise and cumulated versions
Kendallrsquos tau-b is a rank-based correlation metric whichcalculates the difference between the rate of concordanceand discordance [46ndash48] The range of Kendallrsquos tau-b is minus1to 1 where 1 indicates perfect agreement minus1 indicates dataare inverted (perfect agreement inversion) and 0 indicates
no relationship Kendallrsquos tau-b was computed according toBalboaca and Jantschi [48] using dedicated software preparedin MATLAB
Kendallrsquos tau-b values were utilized to quantify interob-server and intraobserver agreements The interobserver agree-ment was estimated between all pairs (1) between patholo-gists themselves ((2) and (3)) between each pathologist andclassical expert microscopic reviewing ((4) and (5)) betweeneach pathologist and the proposed automated method andadditionally in a grouped version between (6) the mean ofpathologist and classical expert microscopic reviewing and(7) the mean of the pathologist and proposed automatedmethod applied to WSI and (8) between classical expertmicroscopic reviewing and proposed automated methodapplied to WSI The intraobserver agreement (agreementbetween the scores of the same observer in various sessionsof area selection) was estimated between all pairs of threeindependent scorings from one pathologist (1) the first andthe second scores (2) the second and the third scores and(3) the first and the third scores both in pooled data and incategorized data Because of the small number of WSIs frompatients in grade III the results relate to only two categoriesgrade I and grade II in diagnosed meningioma patientsConfidence intervals for the overall agreement measureswere calculated applying bootstrap analysis using a proceduredescribed in detail in the study by Gavrielides et al [7]
Software for bootstrap was implemented using MATLAB(MathWorks Natick MA USA) functions
3 Results
First the influence of the 120588 value on the penalty factorwas examined A subset of twelve WSIs from set A waschosen for this analysis The hot-spot localization and areasof quantification selections performed by pathologists andthe automated proposed method were compared using theLCM measure The results of Ki-67 index estimations andLCM measures for 120588 from 01 to 05 with an increment of005 are presented in Figure 3The best concordance betweenthe automatically selected areas of quantification and thoseselected by pathologists is for a 120588 value equal to 02 For this 120588value LCM is minimal for a relatively high value of the Ki-67index
The dispersion of areas of quantification chosen bythe proposed automatic method and pathologists can beobserved in Figures 4 and 5
Figure 4 presents the distribution of the areas of quan-tification selected in hot spots found for two WSIs bytwo pathologists (red and yellow rectangles) and by theproposed automatic method (black rectangles) In the topline (Figures 4(a) and 4(b)) it can be observed that thereis no agreement between both pathologists and that thedistribution of regions is different and inhomogeneous so themeasure of concordance LCM is 9 The distance betweenthe proposed automatic method and the mean measurefor both pathologists is 86 The bottom line (Figures 4(c)and 4(d)) shows good agreement in areas of quantificationdistribution Their measure of concordance LCM is 29
8 Analytical Cellular Pathology
4
5
6
7
8
9
10
202224262830323436
01 015 02 025 03 04 05
LCM
Ki-6
7 in
dex
()
Mean for AUMean for humanLCM
120588
Figure 3 Examination of Ki-67 index and LCM in respect to the 120588
factor of the penalty function
between pathologists while between the proposed automaticmethod and the mean measure for both pathologists it is 34
Figure 5 presents results of three repetitions of the selec-tions of the areas of quantification from one pathologist(Figures 5(a) 5(b) and 5(c) blind trial) and from automatedmethod (Figure 5(d)) using one of the WSIs from set A It isvisible that this one person chose a region of quantificationin various parts of specimen The third trial is significantlydifferent from the previous two rather similar trials but theKi-67 indexes for each of them are similar (107 106 and119 for pathologist and 132 for automated method)
As the selected fields of quantification were previouslyquantified by the software the Ki-67 index for each area andfor each specimen becomes the data for statistical analysis
In Figure 6 where all results for Ki-67 quantificationusing all early interdicted methods of its estimation (man-ually experts by means of two pathologistsrsquo semiauto-matic approach and fully automatic approach) are shownthe general tendency for a relationship between them canbe observed Quantification with the manual microscopicmethod produces the lowest Ki-67 proliferation index valueswhile the automatic methods produce the highest values ofthis index This pattern is biased by one specimen from setA and for 3 specimens from set B In the first case thelower result for the pathologists is caused by an undervaluedscore from one pathologist In the other cases from set Bthe pattern is reversed and the highest values for this indexappear for the manual microscopic method Expose controlof WSI shows that there are very small hot spots in each ofthese three specimens It seems that when hot spots do notcover the whole area of quantification (although they fulfillthe criterion that about 80 of the area of quantificationshould be covered by a hot spot) it causes various resultsfrom the pathologists and automatic method In such casesan expert performing microscopic scoring used to deal withpart of field of quantification restricted to hot spots while theautomatic method diluted the score by counting the numberof cells from the whole area of the rectangle
The results of inter- and intraobserver variability mea-surements are presented in Tables 3 4 and 5
Table 3 presents the results of pairwise agreement usingKendallrsquos tau-b analysis for interobserver variability as acoefficient of concordance along with confidence intervals(95 confidence level) constructed using bootstrap analysisof the samples (100 order changes) It can be observed that allconcordance between all pairs calculated for set A is biggerthan the analogous values for set B This can be explained bythe fact that WSIs in set A were prepared using autostainerand the same set of new chemicals while WSIs in set B wereregular glass slides prepared earlier some with and somewithout autostainer and using the chemicals available at thetime Visual examination shows that glasses in set A are ofreally good quality and homogenous in performance whilethe glasses from set B are not This inhomogeneity amongglasses from set B led to differing interpretations by the twopathologists which is seen as a decrease in agreement betweenthem (from 092 to 086) and between each of them andthe automated method (from 082 and 081 to 078 and 076resp)
Both parts of Table 3 A and B show an overall tendencyfor the highest correlation to be between both pathologistsand the lowest agreement to be between a classical micro-scopic expert scoring and the proposed automated methodemployed on WSI The concordance between both pathol-ogists and the other two methods of scoring are betweenthese two extremes but the concordance is greater betweenthe pathologist and the proposed automated method thanbetween the pathologists and the classical microscopic score
Table 4 presents the results of pooled expert agreement intwo categories grade I and grade II (WHO categorization ofmeningioma) usingKendallrsquos tau-b analysis for interobservervariability The coefficient of concordance is presented withconfidence intervals (95 confidence level) calculated usingbootstrap analysis of the samples (100 order changes) It canbe observed that concordance is higher in category grade IIthan in grade I in both subsets A and B This fact can beexplained as the reason that grade I patientsrsquo scores are usuallylower (up to 8 see Figure 4) than grade II patientsrsquo scores(up to 20 see Figure 4) This means that the hot spots aremore intensive and visible in comparison to the surroundingspace in those specimens from patients diagnosed as gradeII This visibility is more important for the pathologiststhan for the automated method So coincidence betweenpathologistsrsquo digital scoring and the proposed automatedmethod is very high (086 for set A and 08 for set B) ifgrade II patientsrsquo sections are analyzed while for grade Ipatientsrsquo sections the coincidence is lower (079 and 078)but it still shows coincidence For the coincidence betweenexpertsrsquo digital scoring and the manual microscopic expertscoring the results show similar patterns but the numbers arelower (087 043 and 06 and 03 resp)
The coincidence between the results of the proposedautomatic method and manual microscopic expert scoring isambiguousThis coincidence is rather low except for grade IIof set A
Table 5 presents the results of pairwise agreement analysisfor uncategorized and categorized data using Kendallrsquos tau-bfor intraobserver variability One of the pathologists repeatedthe scoring procedure 3 times with a delay long enough
Analytical Cellular Pathology 9
966 1088
372 165
176 147
(a)
990
310
179
(b)
1185 1212
743 924
655 578
(c)
1308
860
577
(d)
Figure 4 TwoWSIs with an area of quantification inside a hot spot marked by two experts ((a) and (c) red and yellow rectangles) and chosenby the proposed automatic method ((b) and (d) black rectangles) with examples of 3 areas of quantification for each method and for bothWSIs with large and small variation in Ki-67 index value calculated as percentage of immunopositive nuclei to the whole number of nucleiin presented area
10 Analytical Cellular Pathology
1420
1253
703
(a)
1573
1274
655
(b)
1522
1518
703
(c)
1891
1500
880
(d)
Figure 5The results of hot spot and area of quantification selections by one pathologist ((a) (b) and (c)) and by proposed automatedmethod(d) with examples of 3 areas of quantification chosen approximately in the same region (hot spot) for each selection The Ki-67 index valuecalculated as percentage of immunopositive nuclei to the whole number of nuclei in presented area differs slightly
Table 3The results of the pairwise agreement analysis of data without taking into account categorical information produced using Kendallrsquostau-b analysis for interobserver variability on the WSI from sets A and B separately TM means classical expert microscopic review P1 andP2 mean pathologists and AU means proposed automatic method
TM P1 P2 AUA
TM 085281 (078355 087879) 079221 (071429 082684) 068831 (060173 068831)P1 085281 (078355 087619) 092208 (088745 093074) 081818 (076623 081818)P2 079221 (071429 082684) 092208 (088745 093074) 080952 (075758 080952)AU 068831 (060173 068831) 081818 (076623 081818) 080952 (075758 080952)
BTM 067687 (065 070017) 058503 (055629 060697) 054762 (051888 056871)P1 067687 (064966 070034) 086735 (085782 087398) 078231 (077262 078912)P2 058503 (055629 060714) 086735 (085748 087381) 07551 (075 076531)AU 054762 (051871 056871) 078231 (077228 078912) 07551 (075 076531)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 4: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/4.jpg)
4 Analytical Cellular Pathology
1x 2x 4x 8x 16x
Figure 1 Fragments of WSI in original and decreased resolution 2x 4x 8x and 16x
The proposed method of analysis of WSI in decreasedresolutions uses the following steps (1) the specimen map isestablished (2) the texture quantification and classificationare done to eliminate hemorrhage areas from the specimenmap (3) the hot spots are detected and finally (4) basedon the proposed penalty function selection of the area ofquantification inside selected hot spots is performed Thegeneral schema of the algorithm for steps 1ndash3 is presented inFigure 2
In the first step a map of the specimen was created usingthe thresholding procedure and morphological filtering [39ndash42] To do this a whole slide image was used to produce asupported image by the morphological operation of openingand brightness equalization This was performed using astructuring element shaped like a disk with a large radius(100 pixels) The operation of the division of each RGB colorcomponent of the image by its version after morphologicalopening was performed independently for each channelAfterwards components from channels B and R were pro-cessed with the Otsu thresholdingmethod [43] Additionallymorphological operations such as erosion dilatation andhole filling were performed to filter the specimen map
The next step which eliminates hemorrhage areas fromthe specimen map was performed by differentiating thetumor area from hemorrhage areas using texture analysis andclassification The local textural descriptors came from theUnser features [33 40] and were applied independently forRGB and CMYK color channels and also for the combined u(from CIE Luv) and C (form CMYK) representation A setof 64 textures was created as 8 features defined in Table 1by 8 color channels or sums of channels as presented in
Table 2 Next based on Fisherrsquos linear discriminant the mostsignificant 25were selected on a teaching phase and then usedin the classification phases (see Table 2)
Finally the Support Vector Machine (SVM) with Gaus-sian kernel function [41ndash46] was applied as a classifier torecognize the hemorrhage areas and to eliminate them fromthe specimen map
The third step of the algorithmwas an estimate of the localdensity of immunopositive cells using the reduced resolutionWSI The local maxima of the immunopositive cell densitiesare hot spots To select these the mathematical morphologyand proportion of the color components were used It wasfound that u of the CIE Luv representation of colors is strictlyassociated with the red color and can be used to differentiatethe immunopositive cells from the remainder of the imageThe extended regional minima transformation is applied toevaluate the spatial relation of the stained brown objects totheir neighboring environmentThe densitymapwas createdbased on the isolatedmarks representing the immunoreactivetumor cells
The fourth and final step of the proposedmethod focusedon the fields of quantification selection based on an artificialmodel of field spatial dispersion To prevent all fields ofquantification being chosen the penalty functionwas definedfrom one large dominant hot spot with a high Ki-67 index bythe following formula
penalty = 1minus120588sum
119894
1
(
radic
(119909 minus 119909
119894)
2+ (119910 minus 119910
119894)
2)
05 (4)
Analytical Cellular Pathology 5
Table2Sign
ificant
texturefeatures(features
in8colorc
hann
elso
rsum
ofchannels(a)colorc
hann
elso
rsum
ofchannelsin
features
(b))
(a)
Color
compo
nent
R(1ndash
8)G
(9ndash16)
B(17ndash24)
u(lu
v)+C
(CMYK
)(25ndash32)
C(33ndash40
)M
(41ndash48)
Y(49ndash
56)
K(57ndash64
)
Num
bero
ffeatures
35
1113
mdash252627293032
333435373840
4142434546
48
4954
63
Featuren
ame
(i)En
ergy
(ii)C
ontrast
(i)En
ergy
(ii)C
ontrast
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)
Hom
ogeneity
Clustershade
(b)
12
34
56
78
PSM
Varia
nce
Energy
Correlatio
nCon
trast
Hom
ogeneity
Clustershade
Clusterp
rominence
(u+C)
CM
Y(u
+C)
CM
RG(u+C)
CM
mdashR
G(u+C)
CM
(u+C)
CM
YK
(u+C)
CM
6 Analytical Cellular Pathology
WSI image (8-fold reduction) Channel R
Channel B
Difference map
Binary map
Otsu thresholding
Specimen map
Morphological
filterin
g
RGB CMYK Luv
Texture Texture Texture
Feature selection
SVM classifier
Hemorrhage and
Map without hemorrhages Immunopositive cells inside Hot-spot selection
Cell recognition for one hot-
WSI file with full resolutionreading the selected hot-spot
field of view
Select
ed field
locat
ion
(RGB)
(RGB)
the map spot
background map
Division of f by 120601(f)
Division of f by 120601(f)
Figure 2 The schema of the algorithm for hot-spot localization
which was based on information about the distance betweenthe designated areas and the position of another candidate forhot spot
An increase in the 120588 value shows an increase in the scat-tering of the areas of quantification selectionThe 120588 value hadbeen chosen experimentally (see Section 3) The proposedfunction combined selection of fields of quantification fromdifferent localization in the specimen according to a gradualreduction of the concentration of immunopositive cellsHowever when hot-spot areas other than the dominant oneshow a significantly lower density of immunopositive nucleithe candidates from dominant region will still be selectedfirstThe final analysis of theKi-67 index in all chosen areas ofquantification was performed on full resolution images withthe method published earlier and described in [45]
26 Evaluation of the Concordance of Selected Hot-Spot FieldsTo evaluate the concordance of hot-spot field localizationbetween the expertsrsquo and automatic results the localizationconcordance measure (LCM) was proposed This measureassumed that (1) those fields at a shorter distance shouldhave a reduced impact on the LCM and (2) the significanceof fields should relate to their Ki-67 index The localization
concordancemeasure was calculated according to the follow-ing formula
LCM
= sum
119894
(119908
119894lowast sigm(min
119895
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
dist [(119909119895 119910
119895) (119909
119894 119910
119894)]
(4FOVsize)
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
))
where 119908119894=
119871
119864119894
119871
119864
(5)
in which 119871
119864is the level of the Ki-67 index for the expert
and FOVsize is a one field of view size A low value ofthe LCM shows a similarity in the areas of quantificationselection by algorithm and expert This means that theexpertrsquos selected fields of quantification are represented ornear the fields selected by the proposed method for examplethey represent the same tumor area If expert and algorithmselect fields of view from different virtual slide areas thelocalization measure LCM is high The proposed measureallows both the evaluation of the similarity of choice of hot-spot fields and the identification of the best penalty factorThe proposedmeasure can be used in cases of both inter- andintraobservation variability
Analytical Cellular Pathology 7
27 Study Design Both the proposed automated methodof area of quantification selection and the two pathologistswere used to review all the samples (A and B sets) usingdigital representation of glass slides with an 8x reducedWSI resolution while expert quantification was done infull resolution and for 10 areas of unknown location Theoutcomes were then averaged to give the final result
Each of the pathologists chose 20 fields of quantificationfor eachWSI One pathologist had two additional sessions fortheWSI from set A to estimate interobserver variability Eachadditional session was performed with at least a one-monthdelay between sessionsThe order of samples was randomizedfor each session
Then automated hot-spot selection software was used toselect 20 fields of quantification as an area which fulfill twocriteria
(i) Biggest number of immunopositive nuclei in compar-ison to the others
(ii) That distances between new and previously foundareas are large enough tomeet the requirements of theabove defined penalty function that is (4)
The scores of the Ki-67 index from the areas of quantifi-cation chosen by the pathologists and the automatic methodwere produced using software which segments nuclei insubimages from WSI in full resolution and then classifiesthem into immunopositive and immunonegative classes andestimates the Ki-67 index This software was published in2009 [45] by the principal investigator of the project who iscoauthor of this paperThe ratio of immunopositive nuclei tothe number of all cells in each area of quantification and themean of these ratios for each WSI were sent for computer-enabled statistical analysis
28 Statistical Analysis The scores from the expert micro-scopic examination and all automatic scores from areaschosen by both pathologists and scores from the proposedautomatic method of area of quantification selection wereanalyzed using agreement analysis since IHC interpretationis a subjective process of evaluation For this process a truescore is not available Besides agreement between digital andoptical scores was not the primary objective of the study butrather this is considered the reference value while agreementbetween the automatic and pathologistsrsquo hot spots and area ofquantification selection in IHC assessment was the main aimof the study
The primary objective of the investigationwas to find pat-terns of agreement between manual human and automatedselection of the area of quantification inWSIThe commonlyused concordance measure Kendallrsquos tau-b was used as inGavrielides and coworkers [7 8] The test was calculatedseparately for setsA andB in pooled and categorizedgroupedversions in both pairwise and cumulated versions
Kendallrsquos tau-b is a rank-based correlation metric whichcalculates the difference between the rate of concordanceand discordance [46ndash48] The range of Kendallrsquos tau-b is minus1to 1 where 1 indicates perfect agreement minus1 indicates dataare inverted (perfect agreement inversion) and 0 indicates
no relationship Kendallrsquos tau-b was computed according toBalboaca and Jantschi [48] using dedicated software preparedin MATLAB
Kendallrsquos tau-b values were utilized to quantify interob-server and intraobserver agreements The interobserver agree-ment was estimated between all pairs (1) between patholo-gists themselves ((2) and (3)) between each pathologist andclassical expert microscopic reviewing ((4) and (5)) betweeneach pathologist and the proposed automated method andadditionally in a grouped version between (6) the mean ofpathologist and classical expert microscopic reviewing and(7) the mean of the pathologist and proposed automatedmethod applied to WSI and (8) between classical expertmicroscopic reviewing and proposed automated methodapplied to WSI The intraobserver agreement (agreementbetween the scores of the same observer in various sessionsof area selection) was estimated between all pairs of threeindependent scorings from one pathologist (1) the first andthe second scores (2) the second and the third scores and(3) the first and the third scores both in pooled data and incategorized data Because of the small number of WSIs frompatients in grade III the results relate to only two categoriesgrade I and grade II in diagnosed meningioma patientsConfidence intervals for the overall agreement measureswere calculated applying bootstrap analysis using a proceduredescribed in detail in the study by Gavrielides et al [7]
Software for bootstrap was implemented using MATLAB(MathWorks Natick MA USA) functions
3 Results
First the influence of the 120588 value on the penalty factorwas examined A subset of twelve WSIs from set A waschosen for this analysis The hot-spot localization and areasof quantification selections performed by pathologists andthe automated proposed method were compared using theLCM measure The results of Ki-67 index estimations andLCM measures for 120588 from 01 to 05 with an increment of005 are presented in Figure 3The best concordance betweenthe automatically selected areas of quantification and thoseselected by pathologists is for a 120588 value equal to 02 For this 120588value LCM is minimal for a relatively high value of the Ki-67index
The dispersion of areas of quantification chosen bythe proposed automatic method and pathologists can beobserved in Figures 4 and 5
Figure 4 presents the distribution of the areas of quan-tification selected in hot spots found for two WSIs bytwo pathologists (red and yellow rectangles) and by theproposed automatic method (black rectangles) In the topline (Figures 4(a) and 4(b)) it can be observed that thereis no agreement between both pathologists and that thedistribution of regions is different and inhomogeneous so themeasure of concordance LCM is 9 The distance betweenthe proposed automatic method and the mean measurefor both pathologists is 86 The bottom line (Figures 4(c)and 4(d)) shows good agreement in areas of quantificationdistribution Their measure of concordance LCM is 29
8 Analytical Cellular Pathology
4
5
6
7
8
9
10
202224262830323436
01 015 02 025 03 04 05
LCM
Ki-6
7 in
dex
()
Mean for AUMean for humanLCM
120588
Figure 3 Examination of Ki-67 index and LCM in respect to the 120588
factor of the penalty function
between pathologists while between the proposed automaticmethod and the mean measure for both pathologists it is 34
Figure 5 presents results of three repetitions of the selec-tions of the areas of quantification from one pathologist(Figures 5(a) 5(b) and 5(c) blind trial) and from automatedmethod (Figure 5(d)) using one of the WSIs from set A It isvisible that this one person chose a region of quantificationin various parts of specimen The third trial is significantlydifferent from the previous two rather similar trials but theKi-67 indexes for each of them are similar (107 106 and119 for pathologist and 132 for automated method)
As the selected fields of quantification were previouslyquantified by the software the Ki-67 index for each area andfor each specimen becomes the data for statistical analysis
In Figure 6 where all results for Ki-67 quantificationusing all early interdicted methods of its estimation (man-ually experts by means of two pathologistsrsquo semiauto-matic approach and fully automatic approach) are shownthe general tendency for a relationship between them canbe observed Quantification with the manual microscopicmethod produces the lowest Ki-67 proliferation index valueswhile the automatic methods produce the highest values ofthis index This pattern is biased by one specimen from setA and for 3 specimens from set B In the first case thelower result for the pathologists is caused by an undervaluedscore from one pathologist In the other cases from set Bthe pattern is reversed and the highest values for this indexappear for the manual microscopic method Expose controlof WSI shows that there are very small hot spots in each ofthese three specimens It seems that when hot spots do notcover the whole area of quantification (although they fulfillthe criterion that about 80 of the area of quantificationshould be covered by a hot spot) it causes various resultsfrom the pathologists and automatic method In such casesan expert performing microscopic scoring used to deal withpart of field of quantification restricted to hot spots while theautomatic method diluted the score by counting the numberof cells from the whole area of the rectangle
The results of inter- and intraobserver variability mea-surements are presented in Tables 3 4 and 5
Table 3 presents the results of pairwise agreement usingKendallrsquos tau-b analysis for interobserver variability as acoefficient of concordance along with confidence intervals(95 confidence level) constructed using bootstrap analysisof the samples (100 order changes) It can be observed that allconcordance between all pairs calculated for set A is biggerthan the analogous values for set B This can be explained bythe fact that WSIs in set A were prepared using autostainerand the same set of new chemicals while WSIs in set B wereregular glass slides prepared earlier some with and somewithout autostainer and using the chemicals available at thetime Visual examination shows that glasses in set A are ofreally good quality and homogenous in performance whilethe glasses from set B are not This inhomogeneity amongglasses from set B led to differing interpretations by the twopathologists which is seen as a decrease in agreement betweenthem (from 092 to 086) and between each of them andthe automated method (from 082 and 081 to 078 and 076resp)
Both parts of Table 3 A and B show an overall tendencyfor the highest correlation to be between both pathologistsand the lowest agreement to be between a classical micro-scopic expert scoring and the proposed automated methodemployed on WSI The concordance between both pathol-ogists and the other two methods of scoring are betweenthese two extremes but the concordance is greater betweenthe pathologist and the proposed automated method thanbetween the pathologists and the classical microscopic score
Table 4 presents the results of pooled expert agreement intwo categories grade I and grade II (WHO categorization ofmeningioma) usingKendallrsquos tau-b analysis for interobservervariability The coefficient of concordance is presented withconfidence intervals (95 confidence level) calculated usingbootstrap analysis of the samples (100 order changes) It canbe observed that concordance is higher in category grade IIthan in grade I in both subsets A and B This fact can beexplained as the reason that grade I patientsrsquo scores are usuallylower (up to 8 see Figure 4) than grade II patientsrsquo scores(up to 20 see Figure 4) This means that the hot spots aremore intensive and visible in comparison to the surroundingspace in those specimens from patients diagnosed as gradeII This visibility is more important for the pathologiststhan for the automated method So coincidence betweenpathologistsrsquo digital scoring and the proposed automatedmethod is very high (086 for set A and 08 for set B) ifgrade II patientsrsquo sections are analyzed while for grade Ipatientsrsquo sections the coincidence is lower (079 and 078)but it still shows coincidence For the coincidence betweenexpertsrsquo digital scoring and the manual microscopic expertscoring the results show similar patterns but the numbers arelower (087 043 and 06 and 03 resp)
The coincidence between the results of the proposedautomatic method and manual microscopic expert scoring isambiguousThis coincidence is rather low except for grade IIof set A
Table 5 presents the results of pairwise agreement analysisfor uncategorized and categorized data using Kendallrsquos tau-bfor intraobserver variability One of the pathologists repeatedthe scoring procedure 3 times with a delay long enough
Analytical Cellular Pathology 9
966 1088
372 165
176 147
(a)
990
310
179
(b)
1185 1212
743 924
655 578
(c)
1308
860
577
(d)
Figure 4 TwoWSIs with an area of quantification inside a hot spot marked by two experts ((a) and (c) red and yellow rectangles) and chosenby the proposed automatic method ((b) and (d) black rectangles) with examples of 3 areas of quantification for each method and for bothWSIs with large and small variation in Ki-67 index value calculated as percentage of immunopositive nuclei to the whole number of nucleiin presented area
10 Analytical Cellular Pathology
1420
1253
703
(a)
1573
1274
655
(b)
1522
1518
703
(c)
1891
1500
880
(d)
Figure 5The results of hot spot and area of quantification selections by one pathologist ((a) (b) and (c)) and by proposed automatedmethod(d) with examples of 3 areas of quantification chosen approximately in the same region (hot spot) for each selection The Ki-67 index valuecalculated as percentage of immunopositive nuclei to the whole number of nuclei in presented area differs slightly
Table 3The results of the pairwise agreement analysis of data without taking into account categorical information produced using Kendallrsquostau-b analysis for interobserver variability on the WSI from sets A and B separately TM means classical expert microscopic review P1 andP2 mean pathologists and AU means proposed automatic method
TM P1 P2 AUA
TM 085281 (078355 087879) 079221 (071429 082684) 068831 (060173 068831)P1 085281 (078355 087619) 092208 (088745 093074) 081818 (076623 081818)P2 079221 (071429 082684) 092208 (088745 093074) 080952 (075758 080952)AU 068831 (060173 068831) 081818 (076623 081818) 080952 (075758 080952)
BTM 067687 (065 070017) 058503 (055629 060697) 054762 (051888 056871)P1 067687 (064966 070034) 086735 (085782 087398) 078231 (077262 078912)P2 058503 (055629 060714) 086735 (085748 087381) 07551 (075 076531)AU 054762 (051871 056871) 078231 (077228 078912) 07551 (075 076531)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 5: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/5.jpg)
Analytical Cellular Pathology 5
Table2Sign
ificant
texturefeatures(features
in8colorc
hann
elso
rsum
ofchannels(a)colorc
hann
elso
rsum
ofchannelsin
features
(b))
(a)
Color
compo
nent
R(1ndash
8)G
(9ndash16)
B(17ndash24)
u(lu
v)+C
(CMYK
)(25ndash32)
C(33ndash40
)M
(41ndash48)
Y(49ndash
56)
K(57ndash64
)
Num
bero
ffeatures
35
1113
mdash252627293032
333435373840
4142434546
48
4954
63
Featuren
ame
(i)En
ergy
(ii)C
ontrast
(i)En
ergy
(ii)C
ontrast
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)V
ariance
(iii)En
ergy
(iv)C
ontrast
(v)H
omogeneity
(vi)Cluster
prom
inence
(i)PS
M(ii)
Hom
ogeneity
Clustershade
(b)
12
34
56
78
PSM
Varia
nce
Energy
Correlatio
nCon
trast
Hom
ogeneity
Clustershade
Clusterp
rominence
(u+C)
CM
Y(u
+C)
CM
RG(u+C)
CM
mdashR
G(u+C)
CM
(u+C)
CM
YK
(u+C)
CM
6 Analytical Cellular Pathology
WSI image (8-fold reduction) Channel R
Channel B
Difference map
Binary map
Otsu thresholding
Specimen map
Morphological
filterin
g
RGB CMYK Luv
Texture Texture Texture
Feature selection
SVM classifier
Hemorrhage and
Map without hemorrhages Immunopositive cells inside Hot-spot selection
Cell recognition for one hot-
WSI file with full resolutionreading the selected hot-spot
field of view
Select
ed field
locat
ion
(RGB)
(RGB)
the map spot
background map
Division of f by 120601(f)
Division of f by 120601(f)
Figure 2 The schema of the algorithm for hot-spot localization
which was based on information about the distance betweenthe designated areas and the position of another candidate forhot spot
An increase in the 120588 value shows an increase in the scat-tering of the areas of quantification selectionThe 120588 value hadbeen chosen experimentally (see Section 3) The proposedfunction combined selection of fields of quantification fromdifferent localization in the specimen according to a gradualreduction of the concentration of immunopositive cellsHowever when hot-spot areas other than the dominant oneshow a significantly lower density of immunopositive nucleithe candidates from dominant region will still be selectedfirstThe final analysis of theKi-67 index in all chosen areas ofquantification was performed on full resolution images withthe method published earlier and described in [45]
26 Evaluation of the Concordance of Selected Hot-Spot FieldsTo evaluate the concordance of hot-spot field localizationbetween the expertsrsquo and automatic results the localizationconcordance measure (LCM) was proposed This measureassumed that (1) those fields at a shorter distance shouldhave a reduced impact on the LCM and (2) the significanceof fields should relate to their Ki-67 index The localization
concordancemeasure was calculated according to the follow-ing formula
LCM
= sum
119894
(119908
119894lowast sigm(min
119895
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
dist [(119909119895 119910
119895) (119909
119894 119910
119894)]
(4FOVsize)
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
))
where 119908119894=
119871
119864119894
119871
119864
(5)
in which 119871
119864is the level of the Ki-67 index for the expert
and FOVsize is a one field of view size A low value ofthe LCM shows a similarity in the areas of quantificationselection by algorithm and expert This means that theexpertrsquos selected fields of quantification are represented ornear the fields selected by the proposed method for examplethey represent the same tumor area If expert and algorithmselect fields of view from different virtual slide areas thelocalization measure LCM is high The proposed measureallows both the evaluation of the similarity of choice of hot-spot fields and the identification of the best penalty factorThe proposedmeasure can be used in cases of both inter- andintraobservation variability
Analytical Cellular Pathology 7
27 Study Design Both the proposed automated methodof area of quantification selection and the two pathologistswere used to review all the samples (A and B sets) usingdigital representation of glass slides with an 8x reducedWSI resolution while expert quantification was done infull resolution and for 10 areas of unknown location Theoutcomes were then averaged to give the final result
Each of the pathologists chose 20 fields of quantificationfor eachWSI One pathologist had two additional sessions fortheWSI from set A to estimate interobserver variability Eachadditional session was performed with at least a one-monthdelay between sessionsThe order of samples was randomizedfor each session
Then automated hot-spot selection software was used toselect 20 fields of quantification as an area which fulfill twocriteria
(i) Biggest number of immunopositive nuclei in compar-ison to the others
(ii) That distances between new and previously foundareas are large enough tomeet the requirements of theabove defined penalty function that is (4)
The scores of the Ki-67 index from the areas of quantifi-cation chosen by the pathologists and the automatic methodwere produced using software which segments nuclei insubimages from WSI in full resolution and then classifiesthem into immunopositive and immunonegative classes andestimates the Ki-67 index This software was published in2009 [45] by the principal investigator of the project who iscoauthor of this paperThe ratio of immunopositive nuclei tothe number of all cells in each area of quantification and themean of these ratios for each WSI were sent for computer-enabled statistical analysis
28 Statistical Analysis The scores from the expert micro-scopic examination and all automatic scores from areaschosen by both pathologists and scores from the proposedautomatic method of area of quantification selection wereanalyzed using agreement analysis since IHC interpretationis a subjective process of evaluation For this process a truescore is not available Besides agreement between digital andoptical scores was not the primary objective of the study butrather this is considered the reference value while agreementbetween the automatic and pathologistsrsquo hot spots and area ofquantification selection in IHC assessment was the main aimof the study
The primary objective of the investigationwas to find pat-terns of agreement between manual human and automatedselection of the area of quantification inWSIThe commonlyused concordance measure Kendallrsquos tau-b was used as inGavrielides and coworkers [7 8] The test was calculatedseparately for setsA andB in pooled and categorizedgroupedversions in both pairwise and cumulated versions
Kendallrsquos tau-b is a rank-based correlation metric whichcalculates the difference between the rate of concordanceand discordance [46ndash48] The range of Kendallrsquos tau-b is minus1to 1 where 1 indicates perfect agreement minus1 indicates dataare inverted (perfect agreement inversion) and 0 indicates
no relationship Kendallrsquos tau-b was computed according toBalboaca and Jantschi [48] using dedicated software preparedin MATLAB
Kendallrsquos tau-b values were utilized to quantify interob-server and intraobserver agreements The interobserver agree-ment was estimated between all pairs (1) between patholo-gists themselves ((2) and (3)) between each pathologist andclassical expert microscopic reviewing ((4) and (5)) betweeneach pathologist and the proposed automated method andadditionally in a grouped version between (6) the mean ofpathologist and classical expert microscopic reviewing and(7) the mean of the pathologist and proposed automatedmethod applied to WSI and (8) between classical expertmicroscopic reviewing and proposed automated methodapplied to WSI The intraobserver agreement (agreementbetween the scores of the same observer in various sessionsof area selection) was estimated between all pairs of threeindependent scorings from one pathologist (1) the first andthe second scores (2) the second and the third scores and(3) the first and the third scores both in pooled data and incategorized data Because of the small number of WSIs frompatients in grade III the results relate to only two categoriesgrade I and grade II in diagnosed meningioma patientsConfidence intervals for the overall agreement measureswere calculated applying bootstrap analysis using a proceduredescribed in detail in the study by Gavrielides et al [7]
Software for bootstrap was implemented using MATLAB(MathWorks Natick MA USA) functions
3 Results
First the influence of the 120588 value on the penalty factorwas examined A subset of twelve WSIs from set A waschosen for this analysis The hot-spot localization and areasof quantification selections performed by pathologists andthe automated proposed method were compared using theLCM measure The results of Ki-67 index estimations andLCM measures for 120588 from 01 to 05 with an increment of005 are presented in Figure 3The best concordance betweenthe automatically selected areas of quantification and thoseselected by pathologists is for a 120588 value equal to 02 For this 120588value LCM is minimal for a relatively high value of the Ki-67index
The dispersion of areas of quantification chosen bythe proposed automatic method and pathologists can beobserved in Figures 4 and 5
Figure 4 presents the distribution of the areas of quan-tification selected in hot spots found for two WSIs bytwo pathologists (red and yellow rectangles) and by theproposed automatic method (black rectangles) In the topline (Figures 4(a) and 4(b)) it can be observed that thereis no agreement between both pathologists and that thedistribution of regions is different and inhomogeneous so themeasure of concordance LCM is 9 The distance betweenthe proposed automatic method and the mean measurefor both pathologists is 86 The bottom line (Figures 4(c)and 4(d)) shows good agreement in areas of quantificationdistribution Their measure of concordance LCM is 29
8 Analytical Cellular Pathology
4
5
6
7
8
9
10
202224262830323436
01 015 02 025 03 04 05
LCM
Ki-6
7 in
dex
()
Mean for AUMean for humanLCM
120588
Figure 3 Examination of Ki-67 index and LCM in respect to the 120588
factor of the penalty function
between pathologists while between the proposed automaticmethod and the mean measure for both pathologists it is 34
Figure 5 presents results of three repetitions of the selec-tions of the areas of quantification from one pathologist(Figures 5(a) 5(b) and 5(c) blind trial) and from automatedmethod (Figure 5(d)) using one of the WSIs from set A It isvisible that this one person chose a region of quantificationin various parts of specimen The third trial is significantlydifferent from the previous two rather similar trials but theKi-67 indexes for each of them are similar (107 106 and119 for pathologist and 132 for automated method)
As the selected fields of quantification were previouslyquantified by the software the Ki-67 index for each area andfor each specimen becomes the data for statistical analysis
In Figure 6 where all results for Ki-67 quantificationusing all early interdicted methods of its estimation (man-ually experts by means of two pathologistsrsquo semiauto-matic approach and fully automatic approach) are shownthe general tendency for a relationship between them canbe observed Quantification with the manual microscopicmethod produces the lowest Ki-67 proliferation index valueswhile the automatic methods produce the highest values ofthis index This pattern is biased by one specimen from setA and for 3 specimens from set B In the first case thelower result for the pathologists is caused by an undervaluedscore from one pathologist In the other cases from set Bthe pattern is reversed and the highest values for this indexappear for the manual microscopic method Expose controlof WSI shows that there are very small hot spots in each ofthese three specimens It seems that when hot spots do notcover the whole area of quantification (although they fulfillthe criterion that about 80 of the area of quantificationshould be covered by a hot spot) it causes various resultsfrom the pathologists and automatic method In such casesan expert performing microscopic scoring used to deal withpart of field of quantification restricted to hot spots while theautomatic method diluted the score by counting the numberof cells from the whole area of the rectangle
The results of inter- and intraobserver variability mea-surements are presented in Tables 3 4 and 5
Table 3 presents the results of pairwise agreement usingKendallrsquos tau-b analysis for interobserver variability as acoefficient of concordance along with confidence intervals(95 confidence level) constructed using bootstrap analysisof the samples (100 order changes) It can be observed that allconcordance between all pairs calculated for set A is biggerthan the analogous values for set B This can be explained bythe fact that WSIs in set A were prepared using autostainerand the same set of new chemicals while WSIs in set B wereregular glass slides prepared earlier some with and somewithout autostainer and using the chemicals available at thetime Visual examination shows that glasses in set A are ofreally good quality and homogenous in performance whilethe glasses from set B are not This inhomogeneity amongglasses from set B led to differing interpretations by the twopathologists which is seen as a decrease in agreement betweenthem (from 092 to 086) and between each of them andthe automated method (from 082 and 081 to 078 and 076resp)
Both parts of Table 3 A and B show an overall tendencyfor the highest correlation to be between both pathologistsand the lowest agreement to be between a classical micro-scopic expert scoring and the proposed automated methodemployed on WSI The concordance between both pathol-ogists and the other two methods of scoring are betweenthese two extremes but the concordance is greater betweenthe pathologist and the proposed automated method thanbetween the pathologists and the classical microscopic score
Table 4 presents the results of pooled expert agreement intwo categories grade I and grade II (WHO categorization ofmeningioma) usingKendallrsquos tau-b analysis for interobservervariability The coefficient of concordance is presented withconfidence intervals (95 confidence level) calculated usingbootstrap analysis of the samples (100 order changes) It canbe observed that concordance is higher in category grade IIthan in grade I in both subsets A and B This fact can beexplained as the reason that grade I patientsrsquo scores are usuallylower (up to 8 see Figure 4) than grade II patientsrsquo scores(up to 20 see Figure 4) This means that the hot spots aremore intensive and visible in comparison to the surroundingspace in those specimens from patients diagnosed as gradeII This visibility is more important for the pathologiststhan for the automated method So coincidence betweenpathologistsrsquo digital scoring and the proposed automatedmethod is very high (086 for set A and 08 for set B) ifgrade II patientsrsquo sections are analyzed while for grade Ipatientsrsquo sections the coincidence is lower (079 and 078)but it still shows coincidence For the coincidence betweenexpertsrsquo digital scoring and the manual microscopic expertscoring the results show similar patterns but the numbers arelower (087 043 and 06 and 03 resp)
The coincidence between the results of the proposedautomatic method and manual microscopic expert scoring isambiguousThis coincidence is rather low except for grade IIof set A
Table 5 presents the results of pairwise agreement analysisfor uncategorized and categorized data using Kendallrsquos tau-bfor intraobserver variability One of the pathologists repeatedthe scoring procedure 3 times with a delay long enough
Analytical Cellular Pathology 9
966 1088
372 165
176 147
(a)
990
310
179
(b)
1185 1212
743 924
655 578
(c)
1308
860
577
(d)
Figure 4 TwoWSIs with an area of quantification inside a hot spot marked by two experts ((a) and (c) red and yellow rectangles) and chosenby the proposed automatic method ((b) and (d) black rectangles) with examples of 3 areas of quantification for each method and for bothWSIs with large and small variation in Ki-67 index value calculated as percentage of immunopositive nuclei to the whole number of nucleiin presented area
10 Analytical Cellular Pathology
1420
1253
703
(a)
1573
1274
655
(b)
1522
1518
703
(c)
1891
1500
880
(d)
Figure 5The results of hot spot and area of quantification selections by one pathologist ((a) (b) and (c)) and by proposed automatedmethod(d) with examples of 3 areas of quantification chosen approximately in the same region (hot spot) for each selection The Ki-67 index valuecalculated as percentage of immunopositive nuclei to the whole number of nuclei in presented area differs slightly
Table 3The results of the pairwise agreement analysis of data without taking into account categorical information produced using Kendallrsquostau-b analysis for interobserver variability on the WSI from sets A and B separately TM means classical expert microscopic review P1 andP2 mean pathologists and AU means proposed automatic method
TM P1 P2 AUA
TM 085281 (078355 087879) 079221 (071429 082684) 068831 (060173 068831)P1 085281 (078355 087619) 092208 (088745 093074) 081818 (076623 081818)P2 079221 (071429 082684) 092208 (088745 093074) 080952 (075758 080952)AU 068831 (060173 068831) 081818 (076623 081818) 080952 (075758 080952)
BTM 067687 (065 070017) 058503 (055629 060697) 054762 (051888 056871)P1 067687 (064966 070034) 086735 (085782 087398) 078231 (077262 078912)P2 058503 (055629 060714) 086735 (085748 087381) 07551 (075 076531)AU 054762 (051871 056871) 078231 (077228 078912) 07551 (075 076531)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 6: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/6.jpg)
6 Analytical Cellular Pathology
WSI image (8-fold reduction) Channel R
Channel B
Difference map
Binary map
Otsu thresholding
Specimen map
Morphological
filterin
g
RGB CMYK Luv
Texture Texture Texture
Feature selection
SVM classifier
Hemorrhage and
Map without hemorrhages Immunopositive cells inside Hot-spot selection
Cell recognition for one hot-
WSI file with full resolutionreading the selected hot-spot
field of view
Select
ed field
locat
ion
(RGB)
(RGB)
the map spot
background map
Division of f by 120601(f)
Division of f by 120601(f)
Figure 2 The schema of the algorithm for hot-spot localization
which was based on information about the distance betweenthe designated areas and the position of another candidate forhot spot
An increase in the 120588 value shows an increase in the scat-tering of the areas of quantification selectionThe 120588 value hadbeen chosen experimentally (see Section 3) The proposedfunction combined selection of fields of quantification fromdifferent localization in the specimen according to a gradualreduction of the concentration of immunopositive cellsHowever when hot-spot areas other than the dominant oneshow a significantly lower density of immunopositive nucleithe candidates from dominant region will still be selectedfirstThe final analysis of theKi-67 index in all chosen areas ofquantification was performed on full resolution images withthe method published earlier and described in [45]
26 Evaluation of the Concordance of Selected Hot-Spot FieldsTo evaluate the concordance of hot-spot field localizationbetween the expertsrsquo and automatic results the localizationconcordance measure (LCM) was proposed This measureassumed that (1) those fields at a shorter distance shouldhave a reduced impact on the LCM and (2) the significanceof fields should relate to their Ki-67 index The localization
concordancemeasure was calculated according to the follow-ing formula
LCM
= sum
119894
(119908
119894lowast sigm(min
119895
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
dist [(119909119895 119910
119895) (119909
119894 119910
119894)]
(4FOVsize)
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
1003816
))
where 119908119894=
119871
119864119894
119871
119864
(5)
in which 119871
119864is the level of the Ki-67 index for the expert
and FOVsize is a one field of view size A low value ofthe LCM shows a similarity in the areas of quantificationselection by algorithm and expert This means that theexpertrsquos selected fields of quantification are represented ornear the fields selected by the proposed method for examplethey represent the same tumor area If expert and algorithmselect fields of view from different virtual slide areas thelocalization measure LCM is high The proposed measureallows both the evaluation of the similarity of choice of hot-spot fields and the identification of the best penalty factorThe proposedmeasure can be used in cases of both inter- andintraobservation variability
Analytical Cellular Pathology 7
27 Study Design Both the proposed automated methodof area of quantification selection and the two pathologistswere used to review all the samples (A and B sets) usingdigital representation of glass slides with an 8x reducedWSI resolution while expert quantification was done infull resolution and for 10 areas of unknown location Theoutcomes were then averaged to give the final result
Each of the pathologists chose 20 fields of quantificationfor eachWSI One pathologist had two additional sessions fortheWSI from set A to estimate interobserver variability Eachadditional session was performed with at least a one-monthdelay between sessionsThe order of samples was randomizedfor each session
Then automated hot-spot selection software was used toselect 20 fields of quantification as an area which fulfill twocriteria
(i) Biggest number of immunopositive nuclei in compar-ison to the others
(ii) That distances between new and previously foundareas are large enough tomeet the requirements of theabove defined penalty function that is (4)
The scores of the Ki-67 index from the areas of quantifi-cation chosen by the pathologists and the automatic methodwere produced using software which segments nuclei insubimages from WSI in full resolution and then classifiesthem into immunopositive and immunonegative classes andestimates the Ki-67 index This software was published in2009 [45] by the principal investigator of the project who iscoauthor of this paperThe ratio of immunopositive nuclei tothe number of all cells in each area of quantification and themean of these ratios for each WSI were sent for computer-enabled statistical analysis
28 Statistical Analysis The scores from the expert micro-scopic examination and all automatic scores from areaschosen by both pathologists and scores from the proposedautomatic method of area of quantification selection wereanalyzed using agreement analysis since IHC interpretationis a subjective process of evaluation For this process a truescore is not available Besides agreement between digital andoptical scores was not the primary objective of the study butrather this is considered the reference value while agreementbetween the automatic and pathologistsrsquo hot spots and area ofquantification selection in IHC assessment was the main aimof the study
The primary objective of the investigationwas to find pat-terns of agreement between manual human and automatedselection of the area of quantification inWSIThe commonlyused concordance measure Kendallrsquos tau-b was used as inGavrielides and coworkers [7 8] The test was calculatedseparately for setsA andB in pooled and categorizedgroupedversions in both pairwise and cumulated versions
Kendallrsquos tau-b is a rank-based correlation metric whichcalculates the difference between the rate of concordanceand discordance [46ndash48] The range of Kendallrsquos tau-b is minus1to 1 where 1 indicates perfect agreement minus1 indicates dataare inverted (perfect agreement inversion) and 0 indicates
no relationship Kendallrsquos tau-b was computed according toBalboaca and Jantschi [48] using dedicated software preparedin MATLAB
Kendallrsquos tau-b values were utilized to quantify interob-server and intraobserver agreements The interobserver agree-ment was estimated between all pairs (1) between patholo-gists themselves ((2) and (3)) between each pathologist andclassical expert microscopic reviewing ((4) and (5)) betweeneach pathologist and the proposed automated method andadditionally in a grouped version between (6) the mean ofpathologist and classical expert microscopic reviewing and(7) the mean of the pathologist and proposed automatedmethod applied to WSI and (8) between classical expertmicroscopic reviewing and proposed automated methodapplied to WSI The intraobserver agreement (agreementbetween the scores of the same observer in various sessionsof area selection) was estimated between all pairs of threeindependent scorings from one pathologist (1) the first andthe second scores (2) the second and the third scores and(3) the first and the third scores both in pooled data and incategorized data Because of the small number of WSIs frompatients in grade III the results relate to only two categoriesgrade I and grade II in diagnosed meningioma patientsConfidence intervals for the overall agreement measureswere calculated applying bootstrap analysis using a proceduredescribed in detail in the study by Gavrielides et al [7]
Software for bootstrap was implemented using MATLAB(MathWorks Natick MA USA) functions
3 Results
First the influence of the 120588 value on the penalty factorwas examined A subset of twelve WSIs from set A waschosen for this analysis The hot-spot localization and areasof quantification selections performed by pathologists andthe automated proposed method were compared using theLCM measure The results of Ki-67 index estimations andLCM measures for 120588 from 01 to 05 with an increment of005 are presented in Figure 3The best concordance betweenthe automatically selected areas of quantification and thoseselected by pathologists is for a 120588 value equal to 02 For this 120588value LCM is minimal for a relatively high value of the Ki-67index
The dispersion of areas of quantification chosen bythe proposed automatic method and pathologists can beobserved in Figures 4 and 5
Figure 4 presents the distribution of the areas of quan-tification selected in hot spots found for two WSIs bytwo pathologists (red and yellow rectangles) and by theproposed automatic method (black rectangles) In the topline (Figures 4(a) and 4(b)) it can be observed that thereis no agreement between both pathologists and that thedistribution of regions is different and inhomogeneous so themeasure of concordance LCM is 9 The distance betweenthe proposed automatic method and the mean measurefor both pathologists is 86 The bottom line (Figures 4(c)and 4(d)) shows good agreement in areas of quantificationdistribution Their measure of concordance LCM is 29
8 Analytical Cellular Pathology
4
5
6
7
8
9
10
202224262830323436
01 015 02 025 03 04 05
LCM
Ki-6
7 in
dex
()
Mean for AUMean for humanLCM
120588
Figure 3 Examination of Ki-67 index and LCM in respect to the 120588
factor of the penalty function
between pathologists while between the proposed automaticmethod and the mean measure for both pathologists it is 34
Figure 5 presents results of three repetitions of the selec-tions of the areas of quantification from one pathologist(Figures 5(a) 5(b) and 5(c) blind trial) and from automatedmethod (Figure 5(d)) using one of the WSIs from set A It isvisible that this one person chose a region of quantificationin various parts of specimen The third trial is significantlydifferent from the previous two rather similar trials but theKi-67 indexes for each of them are similar (107 106 and119 for pathologist and 132 for automated method)
As the selected fields of quantification were previouslyquantified by the software the Ki-67 index for each area andfor each specimen becomes the data for statistical analysis
In Figure 6 where all results for Ki-67 quantificationusing all early interdicted methods of its estimation (man-ually experts by means of two pathologistsrsquo semiauto-matic approach and fully automatic approach) are shownthe general tendency for a relationship between them canbe observed Quantification with the manual microscopicmethod produces the lowest Ki-67 proliferation index valueswhile the automatic methods produce the highest values ofthis index This pattern is biased by one specimen from setA and for 3 specimens from set B In the first case thelower result for the pathologists is caused by an undervaluedscore from one pathologist In the other cases from set Bthe pattern is reversed and the highest values for this indexappear for the manual microscopic method Expose controlof WSI shows that there are very small hot spots in each ofthese three specimens It seems that when hot spots do notcover the whole area of quantification (although they fulfillthe criterion that about 80 of the area of quantificationshould be covered by a hot spot) it causes various resultsfrom the pathologists and automatic method In such casesan expert performing microscopic scoring used to deal withpart of field of quantification restricted to hot spots while theautomatic method diluted the score by counting the numberof cells from the whole area of the rectangle
The results of inter- and intraobserver variability mea-surements are presented in Tables 3 4 and 5
Table 3 presents the results of pairwise agreement usingKendallrsquos tau-b analysis for interobserver variability as acoefficient of concordance along with confidence intervals(95 confidence level) constructed using bootstrap analysisof the samples (100 order changes) It can be observed that allconcordance between all pairs calculated for set A is biggerthan the analogous values for set B This can be explained bythe fact that WSIs in set A were prepared using autostainerand the same set of new chemicals while WSIs in set B wereregular glass slides prepared earlier some with and somewithout autostainer and using the chemicals available at thetime Visual examination shows that glasses in set A are ofreally good quality and homogenous in performance whilethe glasses from set B are not This inhomogeneity amongglasses from set B led to differing interpretations by the twopathologists which is seen as a decrease in agreement betweenthem (from 092 to 086) and between each of them andthe automated method (from 082 and 081 to 078 and 076resp)
Both parts of Table 3 A and B show an overall tendencyfor the highest correlation to be between both pathologistsand the lowest agreement to be between a classical micro-scopic expert scoring and the proposed automated methodemployed on WSI The concordance between both pathol-ogists and the other two methods of scoring are betweenthese two extremes but the concordance is greater betweenthe pathologist and the proposed automated method thanbetween the pathologists and the classical microscopic score
Table 4 presents the results of pooled expert agreement intwo categories grade I and grade II (WHO categorization ofmeningioma) usingKendallrsquos tau-b analysis for interobservervariability The coefficient of concordance is presented withconfidence intervals (95 confidence level) calculated usingbootstrap analysis of the samples (100 order changes) It canbe observed that concordance is higher in category grade IIthan in grade I in both subsets A and B This fact can beexplained as the reason that grade I patientsrsquo scores are usuallylower (up to 8 see Figure 4) than grade II patientsrsquo scores(up to 20 see Figure 4) This means that the hot spots aremore intensive and visible in comparison to the surroundingspace in those specimens from patients diagnosed as gradeII This visibility is more important for the pathologiststhan for the automated method So coincidence betweenpathologistsrsquo digital scoring and the proposed automatedmethod is very high (086 for set A and 08 for set B) ifgrade II patientsrsquo sections are analyzed while for grade Ipatientsrsquo sections the coincidence is lower (079 and 078)but it still shows coincidence For the coincidence betweenexpertsrsquo digital scoring and the manual microscopic expertscoring the results show similar patterns but the numbers arelower (087 043 and 06 and 03 resp)
The coincidence between the results of the proposedautomatic method and manual microscopic expert scoring isambiguousThis coincidence is rather low except for grade IIof set A
Table 5 presents the results of pairwise agreement analysisfor uncategorized and categorized data using Kendallrsquos tau-bfor intraobserver variability One of the pathologists repeatedthe scoring procedure 3 times with a delay long enough
Analytical Cellular Pathology 9
966 1088
372 165
176 147
(a)
990
310
179
(b)
1185 1212
743 924
655 578
(c)
1308
860
577
(d)
Figure 4 TwoWSIs with an area of quantification inside a hot spot marked by two experts ((a) and (c) red and yellow rectangles) and chosenby the proposed automatic method ((b) and (d) black rectangles) with examples of 3 areas of quantification for each method and for bothWSIs with large and small variation in Ki-67 index value calculated as percentage of immunopositive nuclei to the whole number of nucleiin presented area
10 Analytical Cellular Pathology
1420
1253
703
(a)
1573
1274
655
(b)
1522
1518
703
(c)
1891
1500
880
(d)
Figure 5The results of hot spot and area of quantification selections by one pathologist ((a) (b) and (c)) and by proposed automatedmethod(d) with examples of 3 areas of quantification chosen approximately in the same region (hot spot) for each selection The Ki-67 index valuecalculated as percentage of immunopositive nuclei to the whole number of nuclei in presented area differs slightly
Table 3The results of the pairwise agreement analysis of data without taking into account categorical information produced using Kendallrsquostau-b analysis for interobserver variability on the WSI from sets A and B separately TM means classical expert microscopic review P1 andP2 mean pathologists and AU means proposed automatic method
TM P1 P2 AUA
TM 085281 (078355 087879) 079221 (071429 082684) 068831 (060173 068831)P1 085281 (078355 087619) 092208 (088745 093074) 081818 (076623 081818)P2 079221 (071429 082684) 092208 (088745 093074) 080952 (075758 080952)AU 068831 (060173 068831) 081818 (076623 081818) 080952 (075758 080952)
BTM 067687 (065 070017) 058503 (055629 060697) 054762 (051888 056871)P1 067687 (064966 070034) 086735 (085782 087398) 078231 (077262 078912)P2 058503 (055629 060714) 086735 (085748 087381) 07551 (075 076531)AU 054762 (051871 056871) 078231 (077228 078912) 07551 (075 076531)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 7: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/7.jpg)
Analytical Cellular Pathology 7
27 Study Design Both the proposed automated methodof area of quantification selection and the two pathologistswere used to review all the samples (A and B sets) usingdigital representation of glass slides with an 8x reducedWSI resolution while expert quantification was done infull resolution and for 10 areas of unknown location Theoutcomes were then averaged to give the final result
Each of the pathologists chose 20 fields of quantificationfor eachWSI One pathologist had two additional sessions fortheWSI from set A to estimate interobserver variability Eachadditional session was performed with at least a one-monthdelay between sessionsThe order of samples was randomizedfor each session
Then automated hot-spot selection software was used toselect 20 fields of quantification as an area which fulfill twocriteria
(i) Biggest number of immunopositive nuclei in compar-ison to the others
(ii) That distances between new and previously foundareas are large enough tomeet the requirements of theabove defined penalty function that is (4)
The scores of the Ki-67 index from the areas of quantifi-cation chosen by the pathologists and the automatic methodwere produced using software which segments nuclei insubimages from WSI in full resolution and then classifiesthem into immunopositive and immunonegative classes andestimates the Ki-67 index This software was published in2009 [45] by the principal investigator of the project who iscoauthor of this paperThe ratio of immunopositive nuclei tothe number of all cells in each area of quantification and themean of these ratios for each WSI were sent for computer-enabled statistical analysis
28 Statistical Analysis The scores from the expert micro-scopic examination and all automatic scores from areaschosen by both pathologists and scores from the proposedautomatic method of area of quantification selection wereanalyzed using agreement analysis since IHC interpretationis a subjective process of evaluation For this process a truescore is not available Besides agreement between digital andoptical scores was not the primary objective of the study butrather this is considered the reference value while agreementbetween the automatic and pathologistsrsquo hot spots and area ofquantification selection in IHC assessment was the main aimof the study
The primary objective of the investigationwas to find pat-terns of agreement between manual human and automatedselection of the area of quantification inWSIThe commonlyused concordance measure Kendallrsquos tau-b was used as inGavrielides and coworkers [7 8] The test was calculatedseparately for setsA andB in pooled and categorizedgroupedversions in both pairwise and cumulated versions
Kendallrsquos tau-b is a rank-based correlation metric whichcalculates the difference between the rate of concordanceand discordance [46ndash48] The range of Kendallrsquos tau-b is minus1to 1 where 1 indicates perfect agreement minus1 indicates dataare inverted (perfect agreement inversion) and 0 indicates
no relationship Kendallrsquos tau-b was computed according toBalboaca and Jantschi [48] using dedicated software preparedin MATLAB
Kendallrsquos tau-b values were utilized to quantify interob-server and intraobserver agreements The interobserver agree-ment was estimated between all pairs (1) between patholo-gists themselves ((2) and (3)) between each pathologist andclassical expert microscopic reviewing ((4) and (5)) betweeneach pathologist and the proposed automated method andadditionally in a grouped version between (6) the mean ofpathologist and classical expert microscopic reviewing and(7) the mean of the pathologist and proposed automatedmethod applied to WSI and (8) between classical expertmicroscopic reviewing and proposed automated methodapplied to WSI The intraobserver agreement (agreementbetween the scores of the same observer in various sessionsof area selection) was estimated between all pairs of threeindependent scorings from one pathologist (1) the first andthe second scores (2) the second and the third scores and(3) the first and the third scores both in pooled data and incategorized data Because of the small number of WSIs frompatients in grade III the results relate to only two categoriesgrade I and grade II in diagnosed meningioma patientsConfidence intervals for the overall agreement measureswere calculated applying bootstrap analysis using a proceduredescribed in detail in the study by Gavrielides et al [7]
Software for bootstrap was implemented using MATLAB(MathWorks Natick MA USA) functions
3 Results
First the influence of the 120588 value on the penalty factorwas examined A subset of twelve WSIs from set A waschosen for this analysis The hot-spot localization and areasof quantification selections performed by pathologists andthe automated proposed method were compared using theLCM measure The results of Ki-67 index estimations andLCM measures for 120588 from 01 to 05 with an increment of005 are presented in Figure 3The best concordance betweenthe automatically selected areas of quantification and thoseselected by pathologists is for a 120588 value equal to 02 For this 120588value LCM is minimal for a relatively high value of the Ki-67index
The dispersion of areas of quantification chosen bythe proposed automatic method and pathologists can beobserved in Figures 4 and 5
Figure 4 presents the distribution of the areas of quan-tification selected in hot spots found for two WSIs bytwo pathologists (red and yellow rectangles) and by theproposed automatic method (black rectangles) In the topline (Figures 4(a) and 4(b)) it can be observed that thereis no agreement between both pathologists and that thedistribution of regions is different and inhomogeneous so themeasure of concordance LCM is 9 The distance betweenthe proposed automatic method and the mean measurefor both pathologists is 86 The bottom line (Figures 4(c)and 4(d)) shows good agreement in areas of quantificationdistribution Their measure of concordance LCM is 29
8 Analytical Cellular Pathology
4
5
6
7
8
9
10
202224262830323436
01 015 02 025 03 04 05
LCM
Ki-6
7 in
dex
()
Mean for AUMean for humanLCM
120588
Figure 3 Examination of Ki-67 index and LCM in respect to the 120588
factor of the penalty function
between pathologists while between the proposed automaticmethod and the mean measure for both pathologists it is 34
Figure 5 presents results of three repetitions of the selec-tions of the areas of quantification from one pathologist(Figures 5(a) 5(b) and 5(c) blind trial) and from automatedmethod (Figure 5(d)) using one of the WSIs from set A It isvisible that this one person chose a region of quantificationin various parts of specimen The third trial is significantlydifferent from the previous two rather similar trials but theKi-67 indexes for each of them are similar (107 106 and119 for pathologist and 132 for automated method)
As the selected fields of quantification were previouslyquantified by the software the Ki-67 index for each area andfor each specimen becomes the data for statistical analysis
In Figure 6 where all results for Ki-67 quantificationusing all early interdicted methods of its estimation (man-ually experts by means of two pathologistsrsquo semiauto-matic approach and fully automatic approach) are shownthe general tendency for a relationship between them canbe observed Quantification with the manual microscopicmethod produces the lowest Ki-67 proliferation index valueswhile the automatic methods produce the highest values ofthis index This pattern is biased by one specimen from setA and for 3 specimens from set B In the first case thelower result for the pathologists is caused by an undervaluedscore from one pathologist In the other cases from set Bthe pattern is reversed and the highest values for this indexappear for the manual microscopic method Expose controlof WSI shows that there are very small hot spots in each ofthese three specimens It seems that when hot spots do notcover the whole area of quantification (although they fulfillthe criterion that about 80 of the area of quantificationshould be covered by a hot spot) it causes various resultsfrom the pathologists and automatic method In such casesan expert performing microscopic scoring used to deal withpart of field of quantification restricted to hot spots while theautomatic method diluted the score by counting the numberof cells from the whole area of the rectangle
The results of inter- and intraobserver variability mea-surements are presented in Tables 3 4 and 5
Table 3 presents the results of pairwise agreement usingKendallrsquos tau-b analysis for interobserver variability as acoefficient of concordance along with confidence intervals(95 confidence level) constructed using bootstrap analysisof the samples (100 order changes) It can be observed that allconcordance between all pairs calculated for set A is biggerthan the analogous values for set B This can be explained bythe fact that WSIs in set A were prepared using autostainerand the same set of new chemicals while WSIs in set B wereregular glass slides prepared earlier some with and somewithout autostainer and using the chemicals available at thetime Visual examination shows that glasses in set A are ofreally good quality and homogenous in performance whilethe glasses from set B are not This inhomogeneity amongglasses from set B led to differing interpretations by the twopathologists which is seen as a decrease in agreement betweenthem (from 092 to 086) and between each of them andthe automated method (from 082 and 081 to 078 and 076resp)
Both parts of Table 3 A and B show an overall tendencyfor the highest correlation to be between both pathologistsand the lowest agreement to be between a classical micro-scopic expert scoring and the proposed automated methodemployed on WSI The concordance between both pathol-ogists and the other two methods of scoring are betweenthese two extremes but the concordance is greater betweenthe pathologist and the proposed automated method thanbetween the pathologists and the classical microscopic score
Table 4 presents the results of pooled expert agreement intwo categories grade I and grade II (WHO categorization ofmeningioma) usingKendallrsquos tau-b analysis for interobservervariability The coefficient of concordance is presented withconfidence intervals (95 confidence level) calculated usingbootstrap analysis of the samples (100 order changes) It canbe observed that concordance is higher in category grade IIthan in grade I in both subsets A and B This fact can beexplained as the reason that grade I patientsrsquo scores are usuallylower (up to 8 see Figure 4) than grade II patientsrsquo scores(up to 20 see Figure 4) This means that the hot spots aremore intensive and visible in comparison to the surroundingspace in those specimens from patients diagnosed as gradeII This visibility is more important for the pathologiststhan for the automated method So coincidence betweenpathologistsrsquo digital scoring and the proposed automatedmethod is very high (086 for set A and 08 for set B) ifgrade II patientsrsquo sections are analyzed while for grade Ipatientsrsquo sections the coincidence is lower (079 and 078)but it still shows coincidence For the coincidence betweenexpertsrsquo digital scoring and the manual microscopic expertscoring the results show similar patterns but the numbers arelower (087 043 and 06 and 03 resp)
The coincidence between the results of the proposedautomatic method and manual microscopic expert scoring isambiguousThis coincidence is rather low except for grade IIof set A
Table 5 presents the results of pairwise agreement analysisfor uncategorized and categorized data using Kendallrsquos tau-bfor intraobserver variability One of the pathologists repeatedthe scoring procedure 3 times with a delay long enough
Analytical Cellular Pathology 9
966 1088
372 165
176 147
(a)
990
310
179
(b)
1185 1212
743 924
655 578
(c)
1308
860
577
(d)
Figure 4 TwoWSIs with an area of quantification inside a hot spot marked by two experts ((a) and (c) red and yellow rectangles) and chosenby the proposed automatic method ((b) and (d) black rectangles) with examples of 3 areas of quantification for each method and for bothWSIs with large and small variation in Ki-67 index value calculated as percentage of immunopositive nuclei to the whole number of nucleiin presented area
10 Analytical Cellular Pathology
1420
1253
703
(a)
1573
1274
655
(b)
1522
1518
703
(c)
1891
1500
880
(d)
Figure 5The results of hot spot and area of quantification selections by one pathologist ((a) (b) and (c)) and by proposed automatedmethod(d) with examples of 3 areas of quantification chosen approximately in the same region (hot spot) for each selection The Ki-67 index valuecalculated as percentage of immunopositive nuclei to the whole number of nuclei in presented area differs slightly
Table 3The results of the pairwise agreement analysis of data without taking into account categorical information produced using Kendallrsquostau-b analysis for interobserver variability on the WSI from sets A and B separately TM means classical expert microscopic review P1 andP2 mean pathologists and AU means proposed automatic method
TM P1 P2 AUA
TM 085281 (078355 087879) 079221 (071429 082684) 068831 (060173 068831)P1 085281 (078355 087619) 092208 (088745 093074) 081818 (076623 081818)P2 079221 (071429 082684) 092208 (088745 093074) 080952 (075758 080952)AU 068831 (060173 068831) 081818 (076623 081818) 080952 (075758 080952)
BTM 067687 (065 070017) 058503 (055629 060697) 054762 (051888 056871)P1 067687 (064966 070034) 086735 (085782 087398) 078231 (077262 078912)P2 058503 (055629 060714) 086735 (085748 087381) 07551 (075 076531)AU 054762 (051871 056871) 078231 (077228 078912) 07551 (075 076531)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 8: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/8.jpg)
8 Analytical Cellular Pathology
4
5
6
7
8
9
10
202224262830323436
01 015 02 025 03 04 05
LCM
Ki-6
7 in
dex
()
Mean for AUMean for humanLCM
120588
Figure 3 Examination of Ki-67 index and LCM in respect to the 120588
factor of the penalty function
between pathologists while between the proposed automaticmethod and the mean measure for both pathologists it is 34
Figure 5 presents results of three repetitions of the selec-tions of the areas of quantification from one pathologist(Figures 5(a) 5(b) and 5(c) blind trial) and from automatedmethod (Figure 5(d)) using one of the WSIs from set A It isvisible that this one person chose a region of quantificationin various parts of specimen The third trial is significantlydifferent from the previous two rather similar trials but theKi-67 indexes for each of them are similar (107 106 and119 for pathologist and 132 for automated method)
As the selected fields of quantification were previouslyquantified by the software the Ki-67 index for each area andfor each specimen becomes the data for statistical analysis
In Figure 6 where all results for Ki-67 quantificationusing all early interdicted methods of its estimation (man-ually experts by means of two pathologistsrsquo semiauto-matic approach and fully automatic approach) are shownthe general tendency for a relationship between them canbe observed Quantification with the manual microscopicmethod produces the lowest Ki-67 proliferation index valueswhile the automatic methods produce the highest values ofthis index This pattern is biased by one specimen from setA and for 3 specimens from set B In the first case thelower result for the pathologists is caused by an undervaluedscore from one pathologist In the other cases from set Bthe pattern is reversed and the highest values for this indexappear for the manual microscopic method Expose controlof WSI shows that there are very small hot spots in each ofthese three specimens It seems that when hot spots do notcover the whole area of quantification (although they fulfillthe criterion that about 80 of the area of quantificationshould be covered by a hot spot) it causes various resultsfrom the pathologists and automatic method In such casesan expert performing microscopic scoring used to deal withpart of field of quantification restricted to hot spots while theautomatic method diluted the score by counting the numberof cells from the whole area of the rectangle
The results of inter- and intraobserver variability mea-surements are presented in Tables 3 4 and 5
Table 3 presents the results of pairwise agreement usingKendallrsquos tau-b analysis for interobserver variability as acoefficient of concordance along with confidence intervals(95 confidence level) constructed using bootstrap analysisof the samples (100 order changes) It can be observed that allconcordance between all pairs calculated for set A is biggerthan the analogous values for set B This can be explained bythe fact that WSIs in set A were prepared using autostainerand the same set of new chemicals while WSIs in set B wereregular glass slides prepared earlier some with and somewithout autostainer and using the chemicals available at thetime Visual examination shows that glasses in set A are ofreally good quality and homogenous in performance whilethe glasses from set B are not This inhomogeneity amongglasses from set B led to differing interpretations by the twopathologists which is seen as a decrease in agreement betweenthem (from 092 to 086) and between each of them andthe automated method (from 082 and 081 to 078 and 076resp)
Both parts of Table 3 A and B show an overall tendencyfor the highest correlation to be between both pathologistsand the lowest agreement to be between a classical micro-scopic expert scoring and the proposed automated methodemployed on WSI The concordance between both pathol-ogists and the other two methods of scoring are betweenthese two extremes but the concordance is greater betweenthe pathologist and the proposed automated method thanbetween the pathologists and the classical microscopic score
Table 4 presents the results of pooled expert agreement intwo categories grade I and grade II (WHO categorization ofmeningioma) usingKendallrsquos tau-b analysis for interobservervariability The coefficient of concordance is presented withconfidence intervals (95 confidence level) calculated usingbootstrap analysis of the samples (100 order changes) It canbe observed that concordance is higher in category grade IIthan in grade I in both subsets A and B This fact can beexplained as the reason that grade I patientsrsquo scores are usuallylower (up to 8 see Figure 4) than grade II patientsrsquo scores(up to 20 see Figure 4) This means that the hot spots aremore intensive and visible in comparison to the surroundingspace in those specimens from patients diagnosed as gradeII This visibility is more important for the pathologiststhan for the automated method So coincidence betweenpathologistsrsquo digital scoring and the proposed automatedmethod is very high (086 for set A and 08 for set B) ifgrade II patientsrsquo sections are analyzed while for grade Ipatientsrsquo sections the coincidence is lower (079 and 078)but it still shows coincidence For the coincidence betweenexpertsrsquo digital scoring and the manual microscopic expertscoring the results show similar patterns but the numbers arelower (087 043 and 06 and 03 resp)
The coincidence between the results of the proposedautomatic method and manual microscopic expert scoring isambiguousThis coincidence is rather low except for grade IIof set A
Table 5 presents the results of pairwise agreement analysisfor uncategorized and categorized data using Kendallrsquos tau-bfor intraobserver variability One of the pathologists repeatedthe scoring procedure 3 times with a delay long enough
Analytical Cellular Pathology 9
966 1088
372 165
176 147
(a)
990
310
179
(b)
1185 1212
743 924
655 578
(c)
1308
860
577
(d)
Figure 4 TwoWSIs with an area of quantification inside a hot spot marked by two experts ((a) and (c) red and yellow rectangles) and chosenby the proposed automatic method ((b) and (d) black rectangles) with examples of 3 areas of quantification for each method and for bothWSIs with large and small variation in Ki-67 index value calculated as percentage of immunopositive nuclei to the whole number of nucleiin presented area
10 Analytical Cellular Pathology
1420
1253
703
(a)
1573
1274
655
(b)
1522
1518
703
(c)
1891
1500
880
(d)
Figure 5The results of hot spot and area of quantification selections by one pathologist ((a) (b) and (c)) and by proposed automatedmethod(d) with examples of 3 areas of quantification chosen approximately in the same region (hot spot) for each selection The Ki-67 index valuecalculated as percentage of immunopositive nuclei to the whole number of nuclei in presented area differs slightly
Table 3The results of the pairwise agreement analysis of data without taking into account categorical information produced using Kendallrsquostau-b analysis for interobserver variability on the WSI from sets A and B separately TM means classical expert microscopic review P1 andP2 mean pathologists and AU means proposed automatic method
TM P1 P2 AUA
TM 085281 (078355 087879) 079221 (071429 082684) 068831 (060173 068831)P1 085281 (078355 087619) 092208 (088745 093074) 081818 (076623 081818)P2 079221 (071429 082684) 092208 (088745 093074) 080952 (075758 080952)AU 068831 (060173 068831) 081818 (076623 081818) 080952 (075758 080952)
BTM 067687 (065 070017) 058503 (055629 060697) 054762 (051888 056871)P1 067687 (064966 070034) 086735 (085782 087398) 078231 (077262 078912)P2 058503 (055629 060714) 086735 (085748 087381) 07551 (075 076531)AU 054762 (051871 056871) 078231 (077228 078912) 07551 (075 076531)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 9: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/9.jpg)
Analytical Cellular Pathology 9
966 1088
372 165
176 147
(a)
990
310
179
(b)
1185 1212
743 924
655 578
(c)
1308
860
577
(d)
Figure 4 TwoWSIs with an area of quantification inside a hot spot marked by two experts ((a) and (c) red and yellow rectangles) and chosenby the proposed automatic method ((b) and (d) black rectangles) with examples of 3 areas of quantification for each method and for bothWSIs with large and small variation in Ki-67 index value calculated as percentage of immunopositive nuclei to the whole number of nucleiin presented area
10 Analytical Cellular Pathology
1420
1253
703
(a)
1573
1274
655
(b)
1522
1518
703
(c)
1891
1500
880
(d)
Figure 5The results of hot spot and area of quantification selections by one pathologist ((a) (b) and (c)) and by proposed automatedmethod(d) with examples of 3 areas of quantification chosen approximately in the same region (hot spot) for each selection The Ki-67 index valuecalculated as percentage of immunopositive nuclei to the whole number of nuclei in presented area differs slightly
Table 3The results of the pairwise agreement analysis of data without taking into account categorical information produced using Kendallrsquostau-b analysis for interobserver variability on the WSI from sets A and B separately TM means classical expert microscopic review P1 andP2 mean pathologists and AU means proposed automatic method
TM P1 P2 AUA
TM 085281 (078355 087879) 079221 (071429 082684) 068831 (060173 068831)P1 085281 (078355 087619) 092208 (088745 093074) 081818 (076623 081818)P2 079221 (071429 082684) 092208 (088745 093074) 080952 (075758 080952)AU 068831 (060173 068831) 081818 (076623 081818) 080952 (075758 080952)
BTM 067687 (065 070017) 058503 (055629 060697) 054762 (051888 056871)P1 067687 (064966 070034) 086735 (085782 087398) 078231 (077262 078912)P2 058503 (055629 060714) 086735 (085748 087381) 07551 (075 076531)AU 054762 (051871 056871) 078231 (077228 078912) 07551 (075 076531)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 10: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/10.jpg)
10 Analytical Cellular Pathology
1420
1253
703
(a)
1573
1274
655
(b)
1522
1518
703
(c)
1891
1500
880
(d)
Figure 5The results of hot spot and area of quantification selections by one pathologist ((a) (b) and (c)) and by proposed automatedmethod(d) with examples of 3 areas of quantification chosen approximately in the same region (hot spot) for each selection The Ki-67 index valuecalculated as percentage of immunopositive nuclei to the whole number of nuclei in presented area differs slightly
Table 3The results of the pairwise agreement analysis of data without taking into account categorical information produced using Kendallrsquostau-b analysis for interobserver variability on the WSI from sets A and B separately TM means classical expert microscopic review P1 andP2 mean pathologists and AU means proposed automatic method
TM P1 P2 AUA
TM 085281 (078355 087879) 079221 (071429 082684) 068831 (060173 068831)P1 085281 (078355 087619) 092208 (088745 093074) 081818 (076623 081818)P2 079221 (071429 082684) 092208 (088745 093074) 080952 (075758 080952)AU 068831 (060173 068831) 081818 (076623 081818) 080952 (075758 080952)
BTM 067687 (065 070017) 058503 (055629 060697) 054762 (051888 056871)P1 067687 (064966 070034) 086735 (085782 087398) 078231 (077262 078912)P2 058503 (055629 060714) 086735 (085748 087381) 07551 (075 076531)AU 054762 (051871 056871) 078231 (077228 078912) 07551 (075 076531)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 11: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/11.jpg)
Analytical Cellular Pathology 11
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 212223
()
()
Number of cases
Series A
MicroscopeManual averageAutomatic
0
10
20
30
40
50
60
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Series B
Grade I Grade II Grade III
Figure 6Themean and standard deviation of the scoringKi-67 index for all 50 examined glass slides and theirWSI derived frommeningiomapatients Data are grouped according to the types of measurements grade of malignance (grade I grade II and grade III) and type of samplepreparation method (sets A and B) In each of the three bars the first blue bar shows the result for the manual microscopic expert score andthe second red bar shows the result for the mean of the pathologistsrsquo quantifications usingWSI while the third green bar presents results forthe proposed automated method
Table 4The results of pooled (for all experts) agreement analysis of categorized data usingKendallrsquos tau-b analysis for interobserver variabilityon the WSI from sets A and B separately The data are grouped in categories grade I and grade II while grade III was excluded from theanalysis because of the small number of WSIs (3) in this category TMmeans classical expert microscopic review Human means the mean ofthe pathologists and AU means the proposed automatic method
Human TM AUAmdashgrade I
Human 060606 (054545 069697) 078788 (066667 078788)TM 060606 (054545 069697) 039394 (027273 042424)AU 078788 (066667 078788) 039394 (027273 042424)
Amdashgrade IIHuman 086667 (046667 086667) 086667 (046667 086667)TM 086667 (046667 086667) 1 (046667 1)AU 086667 (046667 086667) 1 (046667 1)
Bmdashgrade IHuman 033333 (028105 039608) 077778 (075163 080392)TM 033333 (028105 039869) 032026 (026797 035948)AU 077778 (075163 080392) 032026 (026928 035948)
Bmdashgrade IIHuman 042857 (02381 052381) 080952 (071429 090476)TM 042857 (02381 052381) 02381 (0047619 033333)AU 080952 (071429 090476) 02381 (0047619 033333)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 12: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/12.jpg)
12 Analytical Cellular Pathology
Ki-67
Mon
ths
0
20
40
60
80
0 6 12 18 24 30
Regression 95 confid
Ki-67 versus monthsMonths = 58024 minus 2061 lowast Ki-67
Correlation r = minus06592
(a)
Mon
ths
0
20
40
60
80
Ki-670 4 8 12 16 20 24
Regression 95 confid
Ki-67 versus months
Correlation r =Months = 51848 minus 2588 lowast Ki-67
minus06233
(b)
Regression 95 confid
Mon
ths
0
20
40
60
80
Ki-670 6 12 18 24
Ki-67 versus months
Correlation r =Months = 57186 minus 2388 lowast Ki-67
minus07014
(c)
Figure 7The regression function estimated for the number ofmonths between recurrences ofmeningioma related to the value of Ki-67 indexcalculated based on (i) fully automatic method (a) (ii) traditional microscopic assessment (b) and (iii) semiautomatic human-computerhybrid approach (c)
to forget the samples The results presented coefficients ofconcordance supportedwith confidence intervals (95 confi-dence level) calculated using bootstrap analysis of the samples(100 order changes) This shows very good agreement forall combinations of three scores performed by one observer(all coefficients are between 085 and 09 for data withoutcategorization and between 07 and 085 for those categorizedin grade I and between 073 and 1 for those categorized ingrade II) Intraobserver variability is significantly smallerthan analogous interobserver analysis results presented inTable 4
4 Discussion and Conclusions
The Ki-67 index obtained for each patient WSI deter-mines the downstream clinical decision which concernspatientsrsquo treatment and in consequence patientsrsquo recoveryrecurrences of the disease or patient death To compare
the Ki-67 index obtained using three methods that istraditional microscopic human-computer hybrid methodand the fully automatic method proposed in this paper inthe context of the final results of the therapy for the patientsthere is a need to know full patientsrsquo case histories which arenot available in the Polish Healthcare System What appearsto be available after expose documentation review is data onthe recurrence of meningioma in those patients who havebeen rehospitalized in the same hospital Among 50 patientswhose samples or glasses were used in these investigationsonly 10 patients have currently returned to the same hospitalwith a recurrence of meningioma So the prediction of theprobability of themeningioma recurrence based on the Ki-67index for all three methods of estimation has been estimatedThe regression functions calculated for the number ofmonthsbetween the cancer surgical treatment and its recurrence inrelation to the value of the Ki-67 index calculated based on 10patientsrsquo information are presented in Figure 7
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 13: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/13.jpg)
Analytical Cellular Pathology 13
Table 5The results of intraobserver agreement examinations using Kendallrsquos tau-b analysis calculated based on pairwise analysis performedin uncategorized (top part) and categorized schemas The results show coincidence between Ki-67 indexes based on 3 selections of fields ofquantification by the same pathologist
P1 (1) P1 (2) P1 (3)All
P1 (1) 085004 (08355 088658) 09032 (088398 091775)P1 (2) 085948 (08329 088745) 088216 (085541 090649)P1 (3) 086165 (083203 08961) 089835 (087965 092554)
Grade IP1 (1) 069697 (066667 072727) 084848 (072727 087879)P1 (2) 069697 (066667 072727) 078788 (075758 084848)P1 (3) 084848 (072727 087879) 078788 (075758 084848)
Grade IIP1 (1) 073333 (073333 073333) 073333 (073333 073333)P1 (2) 073333 (073333 073333) 1 (086667 1)P1 (3) 073333 (073333 073333) 1 (086667 1)
Comparing all three regression function parameters (119886119909+119887) and the value of the correlation there is no significantdifference between them
In summary the results of both the above analysis andthe analysis described in the previous section show that thereis no evidence that either hybrid human-computer aidedor fully automatic selection of the area of quantification issuperior for quantifying the Ki-67 index in meningiomapatient samplesThe results of the study showclose agreementin terms of their correlations with tumor recurrences anda relatively high overall agreement for quantification usingboth methods presented in the paper while the results foreach of the methods and traditional macroscopic estimationby an expert are not so high
In this study the time constraints were not examined butwithout any doubt the automatic area selection followed byautomatic analyses would lead to time saving for pathologists
The agreement observed for the three scoring methodsthat is traditional optical microscope and the method basedon digital modalities used by pathologists to select the regionof quantification together with a fully automatic computeraided version of this selection shows that automation of areaselection inWSI is an effective tool in helping physicians andin increasing the reliability of diagnosis based on immunohis-tochemically stained tissue sections Furthermore discussionof the standardization of meningioma Ki-67 quantification iswelcomed
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This studywas supported by theNational Centre for Researchand Development Poland (Grant PBS2A9212013)
References
[1] S H Swerdlow International Agency for Research on Can-cer and World Health Organization WHO Classification ofTumours of Haematopoietic and Lymphoid Tissue World HealthOrganization Classification of Tumors International Agencyfor Research on Cancer 2008
[2] D L Commins R D Atkinson and M E Burnett ldquoReview ofmeningioma histopathologyrdquo Neurosurgical Focus vol 23 no4 p E3 2007
[3] H Colman C Giannini L Huang et al ldquoAssessment and prog-nostic significance of mitotic index using the mitosis markerphospho-histone H3 in low and intermediate-grade infiltratingastrocytomasrdquo American Journal of Surgical Pathology vol 30no 5 pp 657ndash664 2006
[4] A Terzi E A Saglam A Barak and F Soylemezoglu ldquoThesignificance of immunohistochemical expression of Ki-67 p53p21 and p16 in meningiomas tissue arraysrdquo Pathology Researchand Practice vol 204 no 5 pp 305ndash314 2008
[5] A Cruz-Roa F Gonzalez J Galaro et al ldquoA visual latentsemantic approach for automatic analysis and interpretationof anaplastic medulloblastoma virtual slidesrdquo Medical ImageComputing and Computer-Assisted Intervention vol 15 part 1pp 157ndash164 2012
[6] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology amp LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
[7] M A Gavrielides C Conway N OrsquoFlaherty B D Gallasand S M Hewitt ldquoObserver performance in the use of dgitaland optical microscopy for the interpretation of tissue-basedbiomarkersrdquo Analytical Cellular Pathology vol 2014 Article ID157308 10 pages 2014
[8] M A Gavrielides B D Gallas P Lenz A Badano and S MHewitt ldquoObserver variability in the interpretation of HER2neuimmunohistochemical expression with unaided and computer-aided digital microscopyrdquo Archives of Pathology and LaboratoryMedicine vol 135 no 2 pp 233ndash242 2011
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 14: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/14.jpg)
14 Analytical Cellular Pathology
[9] T Seidal A J Balaton and H Battifora ldquoInterpretation andquantification of immunostainsrdquo The American Journal of Sur-gical Pathology vol 25 no 9 pp 1204ndash1207 2001
[10] G Puppa C Senore K Sheahan et al ldquoDiagnostic repro-ducibility of tumour budding in colorectal cancer a multicen-tre multinational study using virtual microscopyrdquoHistopathol-ogy vol 61 no 4 pp 562ndash575 2012
[11] D S A Sanders HGrabsch RHarrison et al ldquoComparing vir-tual with conventional microscopy for the consensus diagnosisof Barrettrsquos neoplasia in the AspECT Barrettrsquos chemopreventiontrial pathology auditrdquoHistopathology vol 61 no 5 pp 795ndash8002012
[12] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2009
[13] K E Brick J C Sluzevich M A Cappel D J DicaudoN I Comfere and C N Wieland ldquoComparison of virtualmicroscopy and glass slide microscopy among dermatologyresidents during a simulated in-training examinationrdquo Journalof Cutaneous Pathology vol 40 no 9 pp 807ndash811 2013
[14] L Pantanowitz J H Sinard W H Henricks et al ldquoValidatingwhole slide imaging for diagnostic purposes in pathologyguideline fromthe College of American Pathologists Pathologyand Laboratory Quality Centerrdquo Archives of Pathology amp Labo-ratory Medicine vol 137 no 12 pp 1710ndash1722 2013
[15] M G Rojo G Bueno and J Slodkowska ldquoReview of imagingsolutions for integrated quantitative immunohistochemistry inthe Pathology daily practicerdquo Folia Histochemica et Cytobiolog-ica vol 47 no 3 pp 349ndash354 2010
[16] T Markiewicz S Osowski J Patera and W Kozlowski ldquoImageprocessing for accurate cell recognition and count on histologicslidesrdquo Analytical and Quantitative Cytology and Histology vol28 no 5 pp 281ndash291 2006
[17] M Krotkiewicz and K Wojtkiewicz ldquoAn introduction toontology based structured knowledge base system knowledgeacquisition modulerdquo in Intelligent Information and DatabaseSystems 5th Asian Conference ACIIDS 2013 Kuala LumpurMalaysia March 18ndash20 2013 Proceedings Part I vol 7802of Lecture Notes in Computer Science pp 497ndash506 SpringerBerlin Germany 2013
[18] M Bator and L J Chmielewski ldquoFinding regions of interest forcancerous masses enhanced by elimination of linear structuresand considerations on detection correctness measures in mam-mographyrdquo Pattern Analysis and Applications vol 12 no 4 pp377ndash390 2009
[19] C Lopez M Lejeune R Bosch et al ldquoDigital image analysisin breast cancer an example of an automated methodology andthe effects of image compressionrdquo Studies in Health Technologyand Informatics vol 179 pp 155ndash171 2012
[20] C Lopez M Lejeune M T Salvado et al ldquoAutomated quantifi-cation of nuclear immunohistochemical markers with differentcomplexityrdquo Histochemistry and Cell Biology vol 129 no 3 pp379ndash387 2008
[21] U Neuman A Korzynska C Lopez and M Lejeune ldquoSeg-mentation of stained lymphoma tissue section imagesrdquo inInformation Technologies in Biomedicine E Piętka and J KawaEds vol 69 of Advances in Intelligent and Soft Computing pp101ndash113 Springer Berlin Germany 2010
[22] A Korzynska L Roszkowiak C Lopez R Bosch LWitkowskiand M Lejeune ldquoValidation of various adaptive threshold
methods of segmentation applied to follicular lymphoma digi-tal images stained with 331015840-diaminobenzidineamphaematoxylinrdquoDiagnostic Pathology vol 8 no 1 article 48 2013
[23] U Neuman A Korzynska C Lopez M Lejeune ŁRoszkowiak and R Bosch ldquoEqualisation of archival micro-scopic images from immunohistochemically stained tissuesectionsrdquo Biocybernetics and Biomedical Engineering vol 33no 1 pp 63ndash76 2013
[24] T Markiewicz P Wisniewski S Osowski J Patera WKozlowski and R Koktysz ldquoComparative analysis of methodsfor accurate recognition of cells through nuclei staining of Ki-67 in neuroblastoma and estrogenprogesterone status stainingin breast cancerrdquo Analytical and Quantitative Cytology andHistology vol 31 no 1 pp 49ndash62 2009
[25] S di Cataldo E Ficarra andEMacii ldquoAutomated segmentationof tissue images for computerized IHC analysisrdquo ComputerMethods and Programs in Biomedicine vol 100 no 1 pp 1ndash152010
[26] S Di Cataldo E Ficarra A Acquaviva and EMacii ldquoAchievingthe way for automated segmentation of nuclei in cancer tissueimages through morphology-based approach a quantitativeevaluationrdquo Computerized Medical Imaging and Graphics vol34 no 6 pp 453ndash461 2010
[27] M M Fernandez-Carrobles I Tadeo G Bueno et al ldquoTMAvessel segmentation based on color andmorphological featuresapplication to angiogenesis researchrdquoThe Scientific World Jour-nal vol 2013 Article ID 263190 11 pages 2013
[28] G Bueno R Gonzalez O Deniz et al ldquoA parallel solution forhigh resolution histological image analysisrdquo Computer Methodsand Programs in Biomedicine vol 108 no 1 pp 388ndash401 2012
[29] V Roullier O Lezoray V-T Ta and A Elmoataz ldquoMulti-resolution graph-based analysis of histopathological whole slideimages application to mitotic cell extraction and visualizationrdquoComputerizedMedical Imaging andGraphics vol 35 no 7-8 pp603ndash615 2011
[30] S Kothari J H Phan T H Stokes andM DWang ldquoPathologyimaging informatics for quantitative analysis of whole-slideimagesrdquo Journal of the American Medical Informatics Associa-tion vol 20 no 6 pp 1099ndash1108 2013
[31] K Kayser D Radziszowski P Bzdyl R Sommer andG KayserldquoTowards an automated virtual slide screening theoreticalconsiderations and practical experiences of automated tissue-based virtual diagnosis to be implemented in the internetrdquoDiagnostic Pathology vol 1 article 10 2006
[32] J R Gilbertson J Ho L Anthony DM Jukic Y Yagi and A VParwani ldquoPrimary histologic diagnosis using automated wholeslide imaging a validation studyrdquo BMC Clinical Pathology vol6 article 4 2006
[33] B Molnar L Berczi C Diczhazy et al ldquoDigital slide andvirtual microscopy based routine and telepathology evaluationof routine gastrointestinal biopsy specimensrdquo Journal of ClinicalPathology vol 56 no 6 pp 433ndash438 2003
[34] S J Potts D A Eberhard and M E Salama ldquoPracticalapproaches to microvessel analysis hotspots microvessel den-sity and vessel proximityrdquo inMolecular Histopathology and Tis-sue Biomarkers in Drug and Diagnostic Development Methodsin Pharmacology and Toxicology pp 87ndash100 Springer 2014
[35] H Lu T G Papathomas D van Zessen et al ldquoAutomated Selec-tion ofHotspots (ASH) enhanced automated segmentation andadaptive step finding for Ki67 hotspot detection in adrenalcortical cancerrdquo Diagnostic Pathology vol 9 no 1 article 2162014
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 15: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/15.jpg)
Analytical Cellular Pathology 15
[36] S Nakasu D H Li H Okabe M Nakajima and M MatsudaldquoSignificance of MIB-1 staining indices in meningiomas com-parison of two counting methodsrdquo The American Journal ofSurgical Pathology vol 25 no 4 pp 472ndash478 2001
[37] M Unser ldquoSum and difference histograms for texture classi-ficationrdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 8 no 1 pp 118ndash125 1986
[38] httpwwwopenslicecom[39] R Koprowski and Z Wrobel ldquoThe cell structures segmen-
tationrdquo in Computer Recognition Systems M Kurzynski EPuchala M Wozniak and A Zolnierek Eds vol 30 ofAdvances in Soft Computing pp 569ndash576 Springer BerlinGermany 2005
[40] A Korzynska U Neuman C Lopez M Lejeun and R BoschldquoThe method of immunohistochemical images standardiza-tionrdquo in Image Processing andCommunicationsChallenges 2 vol84 of Advances in Intelligent and Soft Computing pp 213ndash221Springer Berlin Germany 2010
[41] T Alvaro-Naranjo M Lejeune M-T Salvado et al ldquoImmuno-histochemical patterns of reactive microenvironment are asso-ciated with clinicobiologic behavior in follicular lymphomapatientsrdquo Journal of Clinical Oncology vol 24 no 34 pp 5350ndash5357 2006
[42] P SoilleMorphological Image AnalysismdashPrinciples and Applica-tions Springer Berlin Germany 2004
[43] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo IEEE SystemsMan and Cybernetics Society vol 9 no1 pp 62ndash66 1979
[44] R O Duda P E Hart and P Stork Pattern Classification andScene Analysis Wiley New York NY USA 2003
[45] B Grala T Markiewicz W Kozłowski S Osowski JSłodkowska and W Papierz ldquoNew automated image analysismethod for the assessment of Ki-67 labeling index inmeningio-masrdquo Folia Histochemica et Cytobiologica vol 47 no 4 pp587ndash592 2009
[46] M Kendall Rank Correlation Methods Charles Griffin and CoLimited London UK 1948
[47] R F Woolson and W R Clarke Statistical Methods for theAnalysis of Biomedical Data JohnWiley amp Sons New York NYUSA 1987
[48] S D Balboaca and L Jantschi ldquoPearson versus SpearmanKendallrsquos tau correlation analysis on structure-activity rela-tionships of biologic active compoundsrdquo Leonardo Journal ofSciences no 9 pp 179ndash200 2006
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom
![Page 16: Research Article Comparison of the Manual, Semiautomatic ...downloads.hindawi.com/journals/acp/2015/498746.pdfscoring with manual and automatic hot- spot selection is high, while agreement](https://reader031.fdocuments.in/reader031/viewer/2022041122/5d0d07be88c9931d148b884a/html5/thumbnails/16.jpg)
Submit your manuscripts athttpwwwhindawicom
Stem CellsInternational
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MEDIATORSINFLAMMATION
of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Behavioural Neurology
EndocrinologyInternational Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Disease Markers
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
BioMed Research International
OncologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Oxidative Medicine and Cellular Longevity
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
PPAR Research
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
ObesityJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational and Mathematical Methods in Medicine
OphthalmologyJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Diabetes ResearchJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Research and TreatmentAIDS
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Gastroenterology Research and Practice
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Parkinsonrsquos Disease
Evidence-Based Complementary and Alternative Medicine
Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom