Multiscale Filtering and Neural Network Classification for ...€¦ · Multiscale Filtering and...

Article ID: WMC003588 ISSN 2046-1690

Multiscale Filtering and Neural NetworkClassification for Segmentation and Analysis ofRetinal Vessels.Corresponding Author:Prof. Maurizio Baroni,assistant professor, Department of Electronics and Telecommunications, via S.Marta 3, 50139 - Italy

Submitting Author:Prof. Maurizio Baroni,assistant professor, Department of Electronics and Telecommunications, via S.Marta 3, 50139 - Italy

Article ID: WMC003588

Article Type: Original Articles

Submitted on:17-Jul-2012, 12:36:37 PM GMT Published on: 17-Jul-2012, 04:55:01 PM GMT

Article URL: http://www.webmedcentral.com/article_view/3588

Subject Categories:BIOMEDICAL ENGINEERING

Keywords:Vascular disease, Retinopathy of prematurity, Retina, Image Processing, vessel segmentation,artificial neural network, tracking

How to cite the article:Baroni M, Fortunato P, Pollazzi L, La Torre A. Multiscale Filtering and Neural NetworkClassification for Segmentation and Analysis of Retinal Vessels. . WebmedCentral BIOMEDICAL ENGINEERING2012;3(7):WMC003588

Copyright: This is an open-access article distributed under the terms of the Creative Commons AttributionLicense(CC-BY), which permits unrestricted use, distribution, and reproduction in any medium, provided theoriginal author and source are credited.

Source(s) of Funding:

This work was supported by a research fund of Florence University.

Competing Interests:

The authors have no financial or personal relationships with other people or organizations that couldinappropriately influence this work. No Ethical Approval is required.

Webmedcentral > Original Articles of 22

http://www.webmedcentral.com/article_view/3588

http://creativecommons.org/licenses/by/3.0/

http://creativecommons.org/licenses/by/3.0/

WMC003588 Downloaded from http://www.webmedcentral.com on 25-Jul-2012, 08:00:01 PM

Multiscale Filtering and Neural NetworkClassification for Segmentation and Analysis ofRetinal Vessels.Author(s): Baroni M, Fortunato P, Pollazzi L, La Torre A

Abstract

The blood vessel segmentation method described inthis paper aims at minimizing the false positive rate,while maintaining high accuracy. Though developedfor color retinal images, it can be applied to othertree-like vascular images as well. A computer visionapproach is devised that mimics the image reading byexpert people, based on a two stage process,perception and interpretation. The first stage adoptsmultiscale filtering to detect objects of different sizes: atwo scale Laplacian of Gaussian scheme is used withthe related sigma values chosen according to thesmallest and greatest vessel widths. An approximatesegmentation is achieved simply by means of theLaplacian sign. The interpretation stage isapplication-specific and accomplishes classificationand quantitative analysis. The skeleton of the binarystructures is subdivided in vessel segments, theirfeatures (position, orientation, length and width) arefed into an art i f ic ial neural network, afterback-propagation training. The segments classified asvessels are assembled into the vascular tree byrule-based tracking. Results are evaluated on STAREand DRIVE data bases. Accuracy is 95% and the falsepositive rate is decreased to about 1%. Theapplication on fundus images of infants withretinopathy of prematurity is also described.

Introduction

As average life increases, pathologies like age-relatedmacular degeneration are becoming more frequent; onthe other hand, the progressive improvement inneonatal care leads to an increasing of prematureinfants with lower gestational age and weight [1]. Manyother pathologies may affect blood vessels in all thebody districts, which can be investigated by a widerange of angiographic modalities. In order to supportmedical research and routine examinations, usuallycarried on through visual inspection of angiograms,many computer methods have been developed (recentreviews cite over 200 papers [2,3]): they aim to earlydiagnosis, accurate quantification, treatment and

surgery planning. Moreover, vessel analysis can beuseful for image registration, as needed in multi-modalimaging and patient follow-up. The first step of thesemethods is segmentation: vessels are detected amongother structures and background. Then, vesselfeatures are measured to quantify and follow uppathological alterations. Typical vessel features arewidth, length, branching angles and tortuosity.Unfortunately, manual segmentation of vesselnetworks is too time-consuming and their evaluationby visual inspection is qualitative and subjective; onthe other hand, the numerous computer methodsdeveloped are not always fully automatic nor they warrant optimal performances. Therefore, vesselsegmentation and analysis are still open areas formore research.

Our specific interest regards retinal vessels that maybe affected in the same way of the body vasculatureand can be observed directly and non-invasively bycolor fundus photography. To process the retinalimages obtained at our institute, a few publishedmethods have been taken into consideration [4-6]. Inthis activity some difficulties have been met. First,implementation time is long, particularly whenimplementation details are not all known. Second,also processing time is long for some methods. Last,accuracy depends on image quality: acquisition andpathological artifacts can not be always distinguishedby blood vessels, so the analysis results are biased bythe false positive rate.

To overcome these difficulties and incorporatetogether both segmentation and analysis capabilities,a computer vision approach was devised in order tomimic the image reading by expert people, based onboth perceptual and cognitive stages. It is well knownthat low level vision (perception) adopts multiscalefiltering [7], which enhances and detects objects withdifferent sizes over background. Both vascular andnonvascular structures are detected at this level,unless some vessel model is used. Consideringvascular trees, i.e. vessels pruning from a knownstarting point or region, a tracking process can inprinciple avoid over-segmentation and false positives.Hence, low level segmented structures are



skelotonized and partitioned into single segments, byusing terminal, bifurcation and crossing points; vesselsegments are measured and their features are usedfor classification (vessel/non-vessel) and for trackingthe vessel tree from the starting location.

To this aims, a supervised artificial neural network(ANN) mimics the cognitive image interpretation stepand a rule-based algorithm is used for tracking.Relevant knowledge has been previously incorporatedin the ANN system by means of a training procedure.In this way, a data-driven step extracts vessel features,then a knowledge-based step accomplishes bothvessel segmentation and analysis.

Methods

2.1 Previous work:

Segmentation methods can be classified by means ofseveral criteria [2,3]. Specifically, vessel segmentationmethods can detect whole lumen or its centerline.However, it is difficult to assign a unique classification,because most methods use multiple techniques. Ingeneral, they rely on the elongated and branchedgeometry of blood vessels as well as on their intensity.These structures are enhanced by morphologicoperations or differential filtering and then they aredetected by thresholding, region-growing, iterativetracking, active contours, pattern recognition orartificial intelligence approaches. In [8] a rotating 2Dmatched filtering is used, where vessels are modeledas short, linear, fixed-width segments with a Gaussianprofile: among all the rotation angles, the highestresponse is retained at each pixel and segmentation isachieved by local thresholding. In [4] the matchedfiltered image is thresholded to extract seed points thatare grown according to heuristic criteria. In [9] anadaptive local thresholding of the Gaussian matchedfiltered images is proposed, combined with theresponse to the first difference of Gaussian filter, inorder to reduce the false positives. In [5] vessels areenhanced by multiscale filtering and then aresegmented by region growing, using the histogramsof the local maxima of Hessian features. Gaussianfiltering features are also classified in vessel andnon-vessel according to a ridge-based k-NN classifier[10] or according to a RAdius based ClusteringALgorithm [11]. Also the Gabor wavelet transform canextract vessel features to be classified [6].Mathematical morphology exploits features of thevasculature shape that are known prior, such as itbeing piecewise linear and connected [12,13]. Rather

than applying models or filters to the whole image,vessel tracking approaches start from an initial pointand detect vessel centerlines or boundaries byanalyzing the pixels orthogonal to the trackingdirection. Most of these implementations require userintervention for selecting starting and ending points,e.g. [14]. Also robust automatic approaches have beendeveloped, e.g. [15] detects vessel centerline pixelswith tramline filter and then grows vessel segmentswith an active contour model. However, in fundusimages, the optic nerve head (optic disk) is a naturalchoice as the initial point, since retinal vessels departfrom it. Among the various approaches for optic diskdetection [16], only those which detect its location anddo not segment its boundary are considered here. In[17] vessels are segmented through a complex set of rules and optic disk is detected by means of Houghtransform. The OD has been localized according tobrightness, shape and size, e.g. as the area with thehighest variation in intensity: unfortunately, theseapproaches fail in presence of pathological artifacts.Better approaches are based on the fact that theretinal vasculature originates from the OD following asimilar directional pattern: the vessel tree is firstdetected, then the OD is located by using the conceptof fuzzy convergence [18], the intersection of the twomain vascular groups or direction matched filtering[16]. Once vessels are segmented, their length andwidth as well as other relevant features must bemeasured. Various analysis papers take their inputfrom a pre-segmented binary vascular image andquantify the parameters of the entire vascular networkor those of selected vessels. Other papers report bothsegmentation and analysis methods [14,16]. Amongthose dedicated to retinal vessels, in last years therewas an increasing interest in developing softwarepackages for the analysis of retinopathy of prematurity(ROP), such as the ROPtools program, based on [14],the Retinal Image multiScale Analysis (RISA) software,based on [19] and Computer-Aided Image Analysis ofthe Retina (CAIAR) program, based on [17].

2.2 Retinal images:

Many databases of retinal images are now availableon-line for validation and comparison of segmentationand analysis methods. Since 2000, the STARE(STructural Analysis Retina) project [4] has published20 color fundus images (with 700x605 pixels and 8bits for each RGB channel) related to 10 normal eyesand 10 pathological ones. The DRIVE (Digital RetinalImages for Vessel Extraction) project has published itssoftware as well as 40 color fundus images (with565x584 resolution) divided into training and test sets



[10], which contain only seven eyes showing mildretinopathy. A more complete pathological database isDIARETDB1 ( Diabetic Retinopathy DataBase): thisproject is concentrated on pathological signs withoutconsidering retinal vessels. Another database offundus images is available together with retinal vesselprofiles for vessel width measurement [20]. In thepresent work, both STARE and DRIVE databaseshave been used for comparison purposes, as theyprovide also the vessel segmentation of two experts.Besides, a consecutive case series of eleven eyes ofseven premature infants have been also examined;digital 640x480 images were acquired with 130 degreeRetCam II Imaging System (Clarity Medical Systems,Pleasanton,CA) according to standard ROP protocol.

2.3 Present work:2.3.1 Image pre-processing and segmentation ofblood vessels:

Even if routine fundus images are color photographs,appearing reddish, the green channel of theRGB-representation (illustration 1a) is given as inputto the processing steps, as it has the best contrast: infact the green light is highly absorbed by blood, whilethe blue light is absorbed and dispersed equally in theeye and the red light has the least absorption and isreflected mainly by the choroid [21]. In this work,retinal images have been enhanced by Laplacian ofGaussian (LoG) filtering, implemented in its fast,separable form, as mono-dimensional convolutions,instead of 2D convolution, according to Marr & Hildreth.A resulting image is shown in illustration 1d, togetherwith the results achieved by rotating 2D Gaussianmatched filter [8] (illustration 1b) and multiple orientedGabor wavelets [6] (illustration 1c) for a comparisonpurpose. Gaussian filters can be tuned by varying thesigma parameter and they respond to vessels as wellas to spurious objects, so that the extracted featuresshould be examined by a subsequent classifier. Onthe other hand, vessel width varies from tens of pixelsto sub-pixel level. It is well known how such filterspreserve the objects of size tuned to their scale,whereas other objects are smoothed and smeared, tillto cancel the finest ones. Therefore, instead ofchoosing a unique scale, multiscale filtering [7] hasbeen adopted to simulate the visual perception ofhuman observers: edges are detected at a coarsescale and then are localized at the finest scale. Thegreatest sigma value has been chosen equal to 3 forpublic database and 2 for our ROP image set; oursmallest sigma value is 1. The window size, w, is thesuperior odd number computed with this standard

formula: w = 2?2·3? .The outputs of LoG filters with these scale values arevery similar to the response of a Gaussian matchedfilter, because they extract the vessels with their wholewidth. Using smaller sigma values or applying thesevalues to higher resolution images (or to larger vessels)results in extracting the edges of vessels rather thanwhole vessels.Unlike other enhancement methods, LoG filtering isfast and does not need further processing for vesselsegmentation, such as region growing, thresholding orclassification algorithms. In fact, it allows segmentingthe vessels simply by the output sign. Twodimensional Laplacian changes sign acrossboundaries and therefore segmentation is achieved bya unique threshold: pixels with negative Laplacian aremarked as vessels (white), and the other pixels asbackground (black). From the multiscale outputs,usually, the local maxima over scale are kept pixel bypixel. In our experiments a novel scheme is used tocombine the two scale outputs (illustration 2): abit-wise AND is operated at pixel level, between thethresholded coarse and fine scale Laplacian images.In this way, the small vessels that are detected also atthe coarse scale, preserve their original width,whereas the noisy structures detected only at the finescale are canceled. As a possible drawback of thisscheme, larger vessels may exhibit spurious holes,where their width is over the filter scale. Moreover, allfilters produce spurious vessel segments: specificallythe second derivative calculation results in jaggedvessel borders. To avoid artifacts in the subsequentskeletonizing step, simple median filtering is used, asa cleaning procedure on the segmented images,according to the following sequence:

Ir = [(Is1‡ m5) AND Is2] ‡ m3

where Ir is the resulting segmented image (illustration2c), I?1 and I?2 are the binary fine and coarse scaleLaplacian images, which are combined by means of apixel-by-pixel AND operator; finally, ‡ indicates 2Dconvolution and m5 and m3 indicate median filters,with 5x5 and 3x3 windows, respectively. Then, aclassical thinning algorithm [22] is used to extract theskeleton of the low level vascular tree (illustration 2d).Thinning deletes border points without destroyingconnectivity and its result is a connected, topologicalpreserving central axis of the vascular tree. Threetypes of significant points are detected in the skeleton[19]: terminal points, bifurcation points and crossingpoints. Based on these points, skeletons arepartitioned into single segments. Two neighboringbifurcation points are merged into one crossing points,



according to [19]. A tracking algorithm is needed tonavigate from a branch to another; beforehand,however, every skeleton branching should beclassified to avoid false positives.

2.3.2 Vessel classification and tracking:

Rather than classi fy ing pixel features asvessel/nonvessel, in this work candidate vessels arefirst extracted by the LoG sign and then classifiedaccording to a symbolic representation of vascular tree,as described by human experts: retinal vessels areelongated, connected, dark structures; they originatefrom the optic disc; they branch repeatedly, withtapering widths; they may cross or touch each other,without looping paths. Hence, unlike other proposedmethods, the optic disc (OD) is detected prior tovessel segmentation. To this aim, the brightest andgreatest blob has been detected by smoothing thegreen image and by matching a 2D Gaussian to yieldan approximate location of the optic nerve head. Ahuman operator looks at this result and when thebrightest blob is produced by pathological structures orwhen optic disc is not in the field of view, the correctinitial location of vessel tree is indicated manually.From experts’ descriptions the most relevant featuresfor classification can be derived. Specifically, eachskeleton vessel segment is tracked from an outermostpoint to the other one, by using a chain code [23], andthe following measurements are made: length, width,inner and outer intensity, position with respect to theoptic disc. This analysis module gives a length value,L (along the skeleton segment) and mean and s.d.values of width, Wm and Wsd (evaluated on the localwidths estimated at every three skeleton points as theshortest path across the binary vessel). Moreover, anaverage value of inner intensity, I, is computedthrough region growing on the green image withineach single binary region. To try a distinction betweentrue vessels and false positives, also an outer intensityindex, O, is evaluated as follows. The average graylevel, O1 and O2, and their s.d. are computed for boththe regions besides each candidate binary region, withequal length and half width. If ?=|O1-O2|> s.d., theouter intensity index, O, is set to -?, indicating apossible false positive (e.g. due to OD border,between the bright disc and the dark retinalbackground); otherwise O is set to (O1+O2)/2. Thisvalue indicates a possible true vessel or an artifactbetween two pathological structures (e.g. exudates)which can be discriminated when the latter are muchbrighter than normal retinal background. Finally, byusing the two outermost points of each segment, twodistances from OD, d1 and d2, are computed for each

candidate vessel segment. Thus for each candidatevessel we have seven features that are fed into anartif icial neural network (ANN) to make theclassification. Among the various ANN classifiers, athree layers feed-forward architecture has beenchosen and trained through the back-propagationalgorithm based on batch gradient descent withmomentum. The input layer has 7 neurons, accordingto the aforementioned feature vectors, and the outputlayer has one neuron. Both input and output havelinear transfer functions. The neurons of the hiddenlayer have log-sigmoid transfer functions and theirnumber, n, has been empirically determined, in orderto maximize the mean square error during training andtesting as well as to improve generalization and avoidover-fitting. To the latter aim, the total number of thevessel segments in the training set is taken muchgreater than the number of parameters in the network.The training set includes one half (10) of the STAREimages and one half (20) of the DRIVE images. Theother halves of both databases and our ROP casesare included in the test set. A human expert manuallylabeled the candidate vessel/nonvessel segments(over 100 for each image); their features werecomputed, normalized in the range [0,1], and fed intothe ANN. The maximal number of epochs was 3000and the goal mean square error was 0.005. Severalexperiments were made, by varying the number n ofhidden neurons, the momentum value, the learningrate and the heuristics for faster convergence (variablelearning rate and resilient back-propagation). Once thebest network was trained, it was simulated with thetest set for performance evaluation. The binarysegments classified as vessels are then assembledinto their vascular tree by a tracking algorithm: to thisaim, they are sorted based on the distance from theoptic disc (OD) and tracked by chain code [23],starting from their terminal point nearer to OD.According to this distance sorting, the vesselsegments are labeled as roots (parents) of partialvessel trees; for each parent, its branches (daughters)are then tracked from the current bifurcation point till toanother bifurcation point or till to their terminal point.The daughters of every parent vessel, vk, arenumbered as 2vk and 2vk +1 in order to track the treein either direction. Tracking continues iteratively untilno other daughters are detected and no other vesselsegment has remained unlabeled and identified as anew parent. The final step of tracking examines againthe labeled vessels, trying to connect them if their firstterminal point (the terminal point nearer to OD) is within a short distance (r) from the second terminalpoint (the farther one from OD) of another labeledsegment. This tracking is similar to [19] but it does not



need user interaction and the r distance allows tofollow interrupted vessels or disconnected branches,which may turn out from image artifacts due toprocessing or pathology.Finally, the vessel trees are filled from the labeledskeletons. From these vessel data, further geometricaland topological analyses may be accomplished.

Results

Several experiments were undertaken to devise theoptimal settings of the proposed method. Ad hocsoftware was written using a C++ programmingenvironment and a neural network toolkit, developedbeforehand at our lab. The STARE and DRIVE publicdatabases provided a most valuable reference forthese activities. Specifically, the proposed method isevaluated by computing sensitivity and specificity atpixel level with respect to manually segmented vesselsavailable in both databases. The first observersegmentation is taken as ground truth. Any pixel whichis classified as vessel in both the ground truth andsegmented image is accumulated among the truepositives; any pixel classified as vessel in thesegmented image but not in ground truth is countedamong the false positives. To make a comparison withother methods, also accuracy is evaluated and thesame regions of interest were used: specifically, thespatial masks available on-line for DRIVE images,whereas the regions of interest for STARE databasewere generated by the code provided by Soares [6].The best performance of Gaussian smoothing for opticdisk matching was obtained with a sigma value of 5. InDRIVE database, OD has not been correctly detectedin 3 out of 40 images (7.5%); in STARE database theundetected ODs have been 5 out of 20 (25%),because of the more severe pathological signs.Segmenting with two scale LoG and partitioning of itsskeletons produced at least 2000 candidate vesselsegments for every image, which usually werereduced to less than 500 after artificial neural networkclassification. In the first network configuration, thenumber n of hidden layer neurons was 15. In thistraining phase, the expert was asked to label the mostrelevant vessels, i.e. the candidate segmentsdeparting from OD and their first order ramifications.Unfortunately, even if the mean square error was verygood (2x10-3), several vessels were not classified,giving a low sensitivity (53%). This problem isworsened in ROP images, where most vessels arevery thin. Therefore, another session of training wasaccomplished in order to label also vessel segments ofsecond and third order ramifications. Finally, two ROP

images were added to the training set. In summary,the last network configuration has a number n=20 ofhidden layer neurons and its training convergence isquickened by using resilient back-propagation. Themean square error of 0.05 was reached within 3000iterations during training, as well as its value in thetesting phase was about 0.012. For a typical fundusimage the processing time was nearly one minute andhalf, included about 40” for tracking and filling, using aPC with 1 GHz clock and 1 GB memory. Examples ofsegmentation results are shown in illustrations 3 and 4.A normal DRIVE image segmented by the firstobserver, by the Primitive-Based-Method [10], and bythe proposed method, after the first and last phase oftraining, is shown in illustration 3: sensitivity is clearlyincreased from illustration 3c to illustration 3d.Segmentation results for a pathological STARE image,shown in illustration 4a, can be seen for the firstobserver (illustration 4b), for the Hoover’s method [4](illustration 4c) and for the proposed method, labeledas LoG-ANN (illustration 4d). Illustration 5 presents theaverage values of specificity, sensitivity and accuracycomputed on the DRIVE test images by differentmethods: the performance measures are reportedfrom their original papers for the PBM method [10], forthe second observer (as a measure of humanvariability), for Mendonça et al. [12], for Soares etal.[6], Al-Diri et al.[15], for Zhang et al. [9] and for theproposed method. The same results of these methodsapplied to STARE database are presented inillustration 6, also given for normal and pathologicalimages. Results of Staal et al.[10] and Al-Diri et al.[15],are reported only in the last part of illustration 6, sincethey do not make distinction between normal andabnormal cases. Instead, the results of Hoover et al.[4]are reported in illustration 6. It is worth noting that thistable is not limited to 10 test images, but it refers to allthe 20 STARE images, for comparison purposes. Totalaccuracy, computed on both databases, was nearly0.95 ( 0.9485 ± 0.003), which is comparable with thebest accuracy values reported in literature. However,the false positive rate was very close to 1%, so thatspecificity is better than 98%, whereas the valuesreported in literature are not better than 96 or 97 %.On the other hand, sensitivity is less than 70% andcannot reach the higher values reported in somepapers (over 72% for [6,9,12], about 77% and 90% forthe second observer in DRIVE and STARE imagesrespectively). Finally, it is worth noting the optimalperformance of the proposed method in discriminatingbetween vessels and other spurious structures, asvisible in pathological eyes (illustration 4), and in lowquality images (illustration 7). With regard to tracking,an exhaustive evaluation has not yet done. Because of



segmentation errors or processing artifacts, somevessel segments are missing and some other vesseldaughters are incorrectly identified as new parents.However, the last tracking step is able to followinterrupted vessels and to connect some isolatedsegments to the nearest vascular tree. To this aim thevalue r = 20 was empirically chosen for the distanceparameter.

Discussion

Computer analysis methods of fundus images aim atreducing the amount of manual interaction andlowering qualitative and subjective evaluation,specifically in the follow-up of patients. As an attemptto face the problems still open to discussion, thispaper presents a method for vessel segmentation andanalysis, based on a computer vision approach thatmimics the image reading by expert people (detectionand interpretation). In the first stage, multiscalefiltering detects candidate vessel segments of differentsizes. Many matched filters can enhance tubularobjects and reduce noise at the same time: e.g. Gaborwavelet can be tuned to specific spatial frequenciesand has directional capabilities. However, Laplacian ofGaussian filtering has been adopted on boththeoretical and practical grounds: it approximates thevisual pre-processing at the retina level and is simpleand fast. It allows multiscale enhancement as well assegmentation by sign thresholding, provided that itspass-band is tuned to the vessel width. For this reason,the sigma parameter must be tailored to the imageresolution and to the expected sizes of vessels. As thesecond order spatial derivatives extract vesseltopology even in presence of intensity variations andlow contrast, no preprocessing was used forbrightness normalization. The interpretation stageaccomplishes quantitative analysis and classificationof candidate vessels. In general, quantitative featurescan be classified by supervised methods, such asBayes rule [6], k-nearest-neighbor classifier [10],clustering algorithm [11] and artificial neutral network[24]. ANN classifiers were also used for detection andevaluation of retinal lesions [25]. They are relativelysimple and very robust, able to incorporate nonlinearmappings between a set of input variables and a set ofoutput variables. The features used in the proposedclassifier are pre-selected, based on domainknowledge; there is no selection stage. Training wasdone in two phases and neural network parameterswere carefully tuned, so it may be difficult to apply thismethod to images with different resolution withoutre-sampling them. The artificial neural network

demands the use of manual label ing andtime-consuming training, but the simulation of networkis accurate and fast. The success achieved in usingsimple features, like length, width, intensity, orientationand distance from OD, might support us in tryinganother classification scheme, based on symbolicrules, even if these features interact perhaps in acomplex manner. The use of OD is of particularinterest both to classify vessel segments and toassemble the vascular tree. A more robust and fullyautomated algorithm for OD location is desired inorder to avoid the user interactivity in marking thatpoint. Differences between our results and manualsegmentation are not too far from the inter-observervariability. Accuracy is better than that of the secondobserver. Specificity is better than the values reportedin literature, on the contrary, sensitivity is worse.However, our true positives refer to retinal vasculararcades and their principal ramifications: to our opinion,evaluation of all the vessels in second and third orderramifications is not necessary for most patients. A lowfalse positive rate in vessel segmentation is veryimportant in applications like ROP, where therapy isdecided according to normal or abnormal width andtortuosity of retinal vessels, as shown in illustration 7.In this application it is essential to make thesemeasurements on the whole length of principalvessels, whereas their thinner branching may giveminor information. Tracking of classified vesselsallows to make tortuosity analysis as well as to refinewidth measurement. In fact, after LOG segmentation,thin vessels may not preserve their natural tapering.Therefore, further analysis can be achieved on theintensity profiles in the gray level image [26]. Overall,the proposed method exhibits a low computationalcomplexity and a good performance. We plan to usethe measurements of vessel width and tortuosity for amore complete validation of the proposed method, aswell as to apply this method to other types of vascularimages. Preliminary results were obtained in retinalfluoroangiograms and coronary angiograms, whereasthe neo-vascularization analysis in cornealphotography needs a modification because there isnot an unique starting point like OD. Also results onROP images are encouraging, as ophthalmologists’evaluation of the width enlargement and the severity oftortuosity is subjective. Therefore, automatic andaccurate methods are very useful for the analysis ofretinal images at regular time intervals, to evaluate theprogression and therapy efficacy.

References



1. Early Treatment for Retinopathy of PrematurityCooperative Group. The Incidence and Course ofRetinopathy of Prematurity: Findings From the EarlyTreatment for Retinopathy of Prematurity Study.Pediatrics 2005; 116:15-23.2. Kirbas C, Quek F. A Review of Vessel ExtractionTechniques and Algorithms. ACM Computing Surveys2004; 36:81–121.3. Lesage D, Angelini ED, Bloch I, Funka-Lea G. Areview of 3D vessel lumen segmentation techniques:Models, features and extraction schemes. MedicalImage Analysis 2009; 13: 819–845.4. Hoover A, Kouznetsova V, Goldbaum M. Locatingblood vessels in retinal images by piecewise thresholdprobing of a matched filter response. IEEE Trans MedImaging 2000; 19:203–210.5. Martinez-Perez ME, Hughes AD, Thom SA, BharathAA, Parker KH. Segmentation of blood vessels fromred-free and fluorescein retinal images. Medical ImageAnalysis 2007; 11: 47–61.6. Soares JVB, Leandro JG, Cesar RM Jr., Jelinek HF,Cree MJ. Retinal Vessel Segmentation Using the 2-DGabor Wavelet and Supervised Classification. IEEETrans Med Imaging 2006; 23:1214-1222.7. Lindeberg T (1994) Scale-space theory in computervision. Kluwer Academic Publisher, The Netherlands.8. Chaudhuri S, Chatterjee S, Katz N, Nelson M,Goldbaum M. Detection of blood vessels in retinalimages using two-dimensional matched filters. IEEETrans. Med. Imag. 1989; 8: 263-269.9. Zhang B, Zhang L, Zhang L, Karray F. Retinal Vessel extraction by matched filter with first-orderderivative of Gaussian. Computers in Biology andMedicine 2010; 40:438–445.10. Staal J, Abramoff MD, Niemeijer M, Viergever MA,van Ginneken B. Ridge-based vessel segmentation incolor images for the retina. IEEE Trans Med Imaging2004; 23:501–509.11. Salem SA, Salem NM, Nandi AK. Segmentation ofretinal blood vessels using a novel clustering algorithm(RACAL) with a partial supervision strategy. Med BioEng Comput 2007; 45:261–273.12. Mendonça AM, Campilho A. Segmentation ofRetinal Blood Vessels by Combining the Detection ofCenterlines and Morphological Reconstruction. IEEETrans Med Imaging 2006; 25:1200–1213.13. Zana F, Klein JC. Segmentation of Vessel-LikePatterns Using Mathematical Morphology andCurvature Evaluation. IEEE Trans Image Processing2001; 10: 1010- 1019.14. Aylward SR, Bullitt E. Initialization, noise,singularities, and scale in height ridge traversal fortubular object centerline extraction. IEEE Trans MedImaging 2002; 21: 61-75.

15. Al-Diri B, Hunter A, Steel D. An Active ContourModel for Segmenting and Measuring Retinal Vessels.IEEE Trans.Medical Imaging 2009; 28:1488-97.16. Youssif AA, Ghalwash AZ, Ghoneim ASA. OpticDisc Detection From Normalized Digital FundusImages by Means of a Vessels’ Direction MatchedFilter. IEEE Trans Med Imaging 2008; 27: 11-18.17. Wilson CM, Cocker KD, Moseley MJ, Paterson C,Clay ST, Schulenburg WE, Mills MD, Ells AL, ParkerKH, Quinn GE, Fielder AR, Jeffrey Ng. ComputerizedAnalysis of Retinal Vessel Width and Tortuosity inPremature Infants. Invest Ophthalmol Vis Sci. 2008;49:3577–3785.18. Hoover A, Goldbaum M (2003) Locating the OpticNerve in a Retinal Image Using the FuzzyConvergence of the Blood Vessels. IEEETrans.Medical Imaging, 22, 951-958.19. Martinez-Perez ME, Hughes AD, Stanton AV,Thom SA, Chapman N, Bharath AA,. Parker KH (2002)Retinal Vascular Tree Morphology: A Semi-AutomaticQuantification. IEEE Trans Biomed Eng 49: 912-917.20. Al-Diri B, Hunter A, Steel D, Habib M, Hudaib T,Berry S. REVIEW - A Reference Data Set for RetinalVessel Profiles. 30th Annual Int. Conf. IEEE Eng. inMedicine and Biology Society, 2008.21. Preece SJ, Claridge E. Monte Carlo modeling ofthe spectral reflectance of the human eye. Physics inMedicine and Biology 2001; 47: 2863–2877.22. Zhang TY, Suen CY. A Fast Parallel Algorithm forThinning Digital Patterns. Communications of the ACM1984; 27: 236-239.23. Freeman H. On the encoding of arbitrarygeometric configurations. IRE Transactions onElectronic Computers 1961, EC- 10: 260-268.24. Nekovei R, Sun Y. Back-propagation network andits configuration for blood vessel detection inangiograms. IEEE Trans. Neural Netw. 1995; 6: 64–72.25. García M, López MI, Álvarez D, Hornero R.Assessment of four neural network based classifiers toautomatically detect red lesions in retinal images.Medical Engineering & Physics 2010; 32:1085–1093.26. Lowell J, Hunter A, Steel D, Basu A, Ryder R,Kennedy RL. Measurement of Retinal Vessel WidthsFrom Fundus Images Based on 2-D Modeling. IEEETrans Med Imaging 2004; 23: 1196-1204.



Illustrations

Illustration 1

Illustration 1a) Green channel of a normal fundus image (10_test) of DRIVE database.b) Gaussian matched filter response.c) Gaborfilter output. d) Laplacian of Gaussian.



Illustration 2

Illustration 2a) LoG at fine scale (sigma=1). b) LoG at coarse scale (sigma=3).c) combined LoG. d) skeleton of candidate vessels:terminal points are shown in red, while branching and crossing points are shown in blue.



Illustration 3

Illustration 3a) manual segmentation of 10_test DRIVE image.b) segmentation by the PBM method.c) segmentation by the firstANN.d) segmentation by the final ANN.



Illustration 4

Illustration 4a) an abnormal fundus image (img002) of STARE database.b) manual segmentation.c) segmentation by theHoover’s method.d) segmentation by the LoG-ANN.



method specificity (%) sensitivity (%) accuracy (%)2nd observer 97.25 77.61 94.73

Staal et al.[10] 97.73 71.94 94.42Mendonça et al.[12] 97.64 73.44 94.52

Soares et al.[6] 97.88 72.83 94.66Al-Diri et al.[15] 95.51 72.82 -Zhang et al.[9] 97.24 71.20 93.82

LoG-ANN 98.61 64.76 94.93

Illustration 5

Performance of vessel segmentation methods (DRIVE database).



method specificity (%) sensitivity (%) accuracy (%)normal cases

2nd observer 92.36 96.46 92.83Hoover et al.[4] 96.62 67.66 93.24

Mendonça et al.[12] 97.91 72.58 94.92Soares et al.[6] 98.12 75.54 95.42Zhang et al.[9] 97.79 75.26 95.10

LoG-ANN 98.82 69.37 94.95abnormal cases

2nd observer 95.44 82.52 94.25Hoover et al. [4] 94.72 67.36 92.11

Mendonça et al.[12] 96.69 67.33 93.88Soares et al.[6] 96.82 68.69 94.16Zhang et al.[9] 96.73 71.66 94.39

LoG-ANN 98.62 56.75 94.61all cases

2nd observer 93.90 89.49 93.54Hoover et al. [4] 95.67 67.51 92.67Staal et al.[10] 98.10 69.70 95.16

Mendonça et al.[12] 97.30 69.96 94.40Soares et al.[6] 97.48 71.65 94.80

Al-Diri et al.[15] 96.81 75.21 -Zhang et al.[9] 97.53 71.77 94.84

LoG-ANN 98.72 63.06 94.78

Illustration 6

Performance of vessel segmentation methods (STARE database).



Illustration 7

Illustration 7Follow-up of a premature infant with retinopathy (ROP) and width measurements of the segmented vessels.



Reviews

Review 1

Review Title: Invited review of this paperPosted by Faculty Dr. Qiang (shawn) Cheng on 23 Jul 2012 07:18:32 PM GMT

What are the main claims of the paper and how important are they?: This paper describes a blood vessel segmentation method, which uses a two-state approach: perception andinterpretation. The perception stage uses two-scale Laplacian of Gaussian, which is inspired by a low-level visionmechanism, to detect (both vascular and non-vascular) objects of different sizes. The interpretation stageincludes partitioning the binary structures into vessel segments and extracting a number of their features, andthen classifying these features with artificial neural networks (ANN). Finally a vascular tree is assembled withrule-based tracking. The proposed method is validated on STARE and DRIVE databases.

Many methods for retinal vessel segmentatino have been proposed in the literature; nonetheless, thecombindation of a two-scale filtering with vescular tree analysis might be considered somewhat novel.

Yes. The relevant work is presented properly.

Yes. The proposed method is validated using the STARE and DRIVE databases, and it is compared with otherstate-of-the-art methods. The experimental results appear to be comparable to (and occasionally better than)several state-of-the-art methods.

If a protocol is provided, for example for a randomized controlled trial, are there any important deviationsfrom it? If so, have the authors explained adequately why the deviations occurred? N/A.

Yes, it appears so. The paper provides enough details and the results appear to be reproducible.

A successful, real-world application might be an ultimate validation of the usefulness of the method.

A successful, real-world application might be an ultimate validation of the usefulness of the method.

Rating: 7

Comment: No.

Competing interests: No

Invited by the author to make a review on this article? : No

Have you previously published on this or a similar topic?: No

References:

Experience and credentials in the specific area of science: Good.

How to cite: Anonymous.Invited review of this paper[Review of the article 'Multiscale Filtering and NeuralNetwork Classification for Segmentation and Analysis of Retinal Vessels. ' by ].WebmedCentral1970;3(7):WMCRW002116


http://http://www.webmedcentral.com/Article_Review_View/2116

http://www.webmedcentral.com/faculty/profile/4738


Review 2

Review Title: Multiscale Filtering and Neural Network Classification forSegmentation and Analysis of Retinal VesselsPosted by Prof. Sergio Fonda on 23 Jul 2012 01:49:42 PM GMT

What are the main claims of the paper and how important are they?: NA

NA

NA

NA

If a protocol is provided, for example for a randomized controlled trial, are there any important deviationsfrom it? If so, have the authors explained adequately why the deviations occurred? NA

NA

NA

NA

Rating: 7

Comment: An accurate segmentation of retinal blood vessels has to be done in order to reach a useful diagnosticconclusion about their morphological changes in important diseases like diabetes and hypertension. All thisprocess, if manually done, is largely time consuming and presents difficulties in patient’s results comparisons.The main paper’s claim is an automatic procedure including both vessel tree segmentation (detection of objectsagainst the retina image background) and its consequent diagnostic analysis performed by an artificial neuralnetwork operating according to a “human vision inspired” approach. Owing to the still difficult automatic achievingof reliable diagnostic assessment, it may be promising.

Many algorithms and methods of segmentation, included the use of artificial neural networks, have been alreadyintroduced; however the automatic integration of conventional multiscale segmentation and “human vision based”vessel tree analysis, presented in this paper, may be considered a novel approach.

The paper could refer also to some recent review on the topic, as for example:MM Fraz et al. Blood vessel segmentation methodologies in retinal images – A survey.Computer Methods and Programs in B iomedic ine, Avai lab le on l ine 21 Apr i l 2012.http://dx.doi.org/10.1016/j.cmpb.2012.03.009 or some old basic paper published for the ophthalmologicalcommunity: Chanjira Sinthanayothin et al. Automated localisation of the optic disc, fovea, and retinal bloodvessels from digital colour fundus images. Br J Ophthalmol 1999;83:902–910

The results support the claims but it needs a better way of detecting OD in pathological retinas.

The authors use STARE and DRIVE databases offering sensitivity, specificity and accuracy evaluation of the test;this presents a solid reproducibility.

Despite being the exploitation of STARE and DRIVE very useful to test algorithms, a clinical trial on patientsshould be introduced as soon as possible to assess drawbacks, limits and advantages of this method.




http://www.webmedcentral.com/user/5562


The paper is a good contribution to the field of automatic diagnosis of retinal pathologies but is not outstanding.However it should be included in review lectures about this topic.

Some difficulty arises to interpret correctly the analytic notations reported at page 4 and 5 of pdf file, due tounusual characters/symbols used (typo or font errors? Both in the online writeup and in pdf text a question markis typed, e.g. w = 2?2·3? or I?1, I?2 )




References: NA

Experience and credentials in the specific area of science: I was involved in Ophthalmological research years ago

How to cite: Fonda S.Multiscale Filtering and Neural Network Classification for Segmentation and Analysis ofRetinal Vessels[Review of the article 'Multiscale Filtering and Neural Network Classification for Segmentation andAnalysis of Retinal Vessels. ' by ].WebmedCentral 1970;3(7):WMCRW002113



Review 3

Review Title: Multiscale Filtering and Neural Network Classification forSegmentation and Analysis of Retinal VesselsPosted by Prof. Sergio Fonda on 22 Jul 2012 04:17:30 PM GMT

What are the main claims of the paper and how important are they?: An accurate segmentation of retinal blood vessels has to be done in order to reach a useful diagnosticconclusion about their morphological changes in important diseases like diabetes and hypertension. All thisprocess, if manually done, is largely time consuming and presents difficulties in patient’s results comparisons.The main paper’s claim is an automatic procedure including both vessel tree segmentation (detection of objectsagainst the retina image background) and its consequent diagnostic analysis performed by an artificial neuralnetwork operating according to a “human vision inspired” approach. Owing to the still difficult automatic achievingof reliable diagnostic assessment, it may be promising.

Many algorithms and methods of segmentation, included the use of artificial neural networks, have been alreadyintroduced; however the automatic integration of conventional multiscale segmentation and “human vision based”vessel tree analysis, presented in this paper, may be considered a novel approach.

Yes. However the paper could refer also to some recent review on the topic, as for example: MM Fraz et al. Blood vessel segmentation methodologies in retinal images – A survey.Computer Methods and Programs in B iomedic ine, Avai lab le on l ine 21 Apr i l 2012.http://dx.doi.org/10.1016/j.cmpb.2012.03.009

or some old basic paper published for the ophthalmological community:Chanjira Sinthanayothin et al. Automated localisation of the optic disc, fovea, and retinal blood vessels fromdigital colour fundus images. Br J Ophthalmol 1999;83:902–910

Yes, but it needs a better way of detecting optic disc in pathological retinas

If a protocol is provided, for example for a randomized controlled trial, are there any important deviationsfrom it? If so, have the authors explained adequately why the deviations occurred? NA

The authors use STARE and DRIVE databases, offering sensitivity, specificity and accuracy evaluation of thetest; this presents a solid reproducibility

Despite being the exploitation of STARE and DRIVE very useful to test algorithms, a clinical trial on patientsshould be introduced as soon as possible to assess drawbacks, limits and advantages of this method

Despite being the exploitation of STARE and DRIVE very useful to test algorithms, a clinical trial on patientsshould be introduced as soon as possible to assess drawbacks, limits and advantages of this method

Rating: 7

Comment: Difficulty to interpret correctly the analytic notations reported at page 4 and 5, due to unusual characters/symbolsused (typo or font errors? Both online and in pdf text a question mark is typed, e.g. w = 2?2·3? or I?1, I?2 )


Invited by the author to make a review on this article? : Yes




http://www.webmedcentral.com/review_user/9529/Reviewer


Have you previously published on this or a similar topic?: Yes

References: I was involved in ophthalmological research years ago. Sergio Fonda, MS; Bruno Bagolini, MD, RelativePhotometric Measurements of Retinal Circulation (Dromofluorograms)_A Television Technique. Arch Ophthalmol.1977;95(2):302-307. doi:10.1001/archopht.1977.04450020103017

How to cite: Fonda S.Multiscale Filtering and Neural Network Classification for Segmentation and Analysis ofRetinal Vessels[Review of the article 'Multiscale Filtering and Neural Network Classification for Segmentation andAnalysis of Retinal Vessels. ' by ].WebmedCentral 1970;3(7):WMCRW002111



Review 4

Review Title: Multiscale Filtering and Neural Network Classification forSegmentation and Analysis of Retinal Vessels.Posted by Faculty Dr. M. A. Balafar on 18 Jul 2012 07:45:03 AM GMT

What are the main claims of the paper and how important are they?: A computer vision approach which mimics the image reading by expert people.

so so

Yes

Yes

If a protocol is provided, for example for a randomized controlled trial, are there any important deviationsfrom it? If so, have the authors explained adequately why the deviations occurred? It is not the case

Yes

No

No

Rating: 7

Comment: It would better to utilize diagrams for methodology explaination to increase readibility of the paper




References:

Experience and credentials in the specific area of science: Good

How to cite: Balafar M.Multiscale Filtering and Neural Network Classification for Segmentation and Analysis ofRetinal Vessels.[Review of the article 'Multiscale Filtering and Neural Network Classification for Segmentationand Analysis of Retinal Vessels. ' by ].WebmedCentral 1970;3(7):WMCRW002098




http://www.webmedcentral.com/faculty/profile/3227


DisclaimerThis article has been downloaded from WebmedCentral. With our unique author driven post publication peerreview, contents posted on this web portal do not undergo any prepublication peer or editorial review. It iscompletely the responsibility of the authors to ensure not only scientific and ethical standards of the manuscriptbut also its grammatical accuracy. Authors must ensure that they obtain all the necessary permissions beforesubmitting any information that requires obtaining a consent or approval from a third party. Authors should alsoensure not to submit any information which they do not have the copyright of or of which they have transferredthe copyrights to a third party.

Contents on WebmedCentral are purely for biomedical researchers and scientists. They are not meant to cater tothe needs of an individual patient. The web portal or any content(s) therein is neither designed to support, norreplace, the relationship that exists between a patient/site visitor and his/her physician. Your use of theWebmedCentral site and its contents is entirely at your own risk. We do not take any responsibility for any harmthat you may suffer or inflict on a third person by following the contents of this website.


Multiscale Filtering and Neural Network Classification for ...€¦ · Multiscale Filtering and...

Documents

Transcript of Multiscale Filtering and Neural Network Classification for ...€¦ · Multiscale Filtering and...