A Novel Segmentation Approach Combining Region- and Edge...

Research ArticleA Novel Segmentation Approach Combining Region- andEdge-Based Information for Ultrasound Images

Yaozhong Luo1 Longzhong Liu2 Qinghua Huang13 and Xuelong Li4

1School of Electronic and Information Engineering South China University of Technology Guangzhou China2Department of Ultrasound The Cancer Center of Sun Yat-sen University State Key Laboratory of Oncology inSouth China Collaborative Innovation Center for Cancer Medicine Guangdong China3College of Information Engineering Shenzhen University Shenzhen 518060 China4Center for OPTical IMagery Analysis and Learning (OPTIMAL) State Key Laboratory of Transient Optics and PhotonicsXirsquoan Institute of Optics and Precision Mechanics Chinese Academy of Sciences Xirsquoan Shaanxi 710119 China

Correspondence should be addressed to Qinghua Huang qhhuangscuteducn

Received 9 September 2016 Revised 21 January 2017 Accepted 14 March 2017 Published 27 April 2017

Academic Editor Cristiana Corsi

Copyright copy 2017 Yaozhong Luo et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Ultrasound imaging has become one of the most popular medical imaging modalities with numerous diagnostic applicationsHowever ultrasound (US) image segmentation which is the essential process for further analysis is a challenging task due to thepoor image quality In this paper we propose a new segmentation scheme to combine both region- and edge-based information intothe robust graph-based (RGB) segmentation method The only interaction required is to select two diagonal points to determinea region of interest (ROI) on the original image The ROI image is smoothed by a bilateral filter and then contrast-enhanced byhistogram equalizationThen the enhanced image is filtered by pyramidmean shift to improve homogeneityWith the optimizationof particle swarm optimization (PSO) algorithm the RGB segmentation method is performed to segment the filtered image Thesegmentation results of ourmethod have been compared with the corresponding results obtained by three existing approaches andfour metrics have been used to measure the segmentation performance The experimental results show that the method achievesthe best overall performance and gets the lowest ARE (1077) the second highest TPVF (8534) and the second lowest FPVF(448)

1 Introduction

Ultrasound (US) imaging is one of the most popular medicalimaging modalities with numerous diagnostic applicationsdue to the following merits no radiation faster imaginghigher sensitivity and accuracy and lower cost comparedto other imaging modalities such as computed tomography(CT) or magnetic resonance imaging (MRI) [1ndash6] Howeversonography is operator-dependent and reading US imagesrequires well-trained and experienced radiologists To reducethe interobserver variation among different clinicians andhelp them generate more reliable and accurate diagnos-tic conclusions computer-aided diagnosis (CAD) has beenproposed [3 7 8] Generally the CAD system based onthe US image involves the following four steps prepro-cessing segmentation feature extraction and selection and

classification [9 10] Among these four procedures imagesegmentation which separates the lesion region from thebackground is the key to the subsequent processing anddetermines the quality of the final analysis In the pre-vious clinical practice the segmentation task is generallyperformed by manual tracing which is laborious time-consuming and skill- and experience-dependent Conse-quently reliable and automatic segmentation methods arepreferred to segment the ROI from the US image to improvethe automation and robustness of the CAD system Howeveraccurate and automatic US image segmentation remains achallenging task [11ndash13] due to various US artifacts includinghigh speckle noise [14] low signal-to-noise ratio and inten-sity inhomogeneity [15]

In the last decade a large number of segmentationmethods have been developed for US images for example

HindawiBioMed Research InternationalVolume 2017 Article ID 9157341 18 pageshttpsdoiorg10115520179157341

2 BioMed Research International

thresholding-based methods [16ndash18] clustering-based meth-ods [19ndash23] watershed-based methods [24ndash27] graph-based methods [28ndash35] and active contour models [36ndash42]Thresholding is one of the frequently used segmentationtechniques for themonochrome image Yap et al [18] adoptedthe thresholding segmentation to separate the lesion regionfrom the background before detecting the initial boundaryvia edge detection Clustering is a classification technique andhas been successfully applied to image segmentation basedon similarity between image regions or pixels Isa et al [19]used the moving 119896-means clustering to automatically selectthe seed and proposed a modified seed based region growingalgorithm to detect the edge Shan et al [20] used a novel neu-trosophic clustering approach to detect the lesion boundaryMoon et al [22] used the fuzzy C-means (FCM) clusteringto extract the tumor candidates in their CAD system Thewatershed transformation which is frequently used in thesegmentation of grey scale images considers the gradientmagnitude of an image as a topographic surface Chen etal [24] employed the two-pass watershed transformationsto generate the cells and proposed a region-based approachcalled cell-competition algorithm to simultaneously segmentmultiple objects in a sonogram L Zhang and M Zhang [26]used an extended fuzzy watershed method to segment USimages fully automatically The experiments showed that theproposedmethod could get good results on blurryUS images

In the last few years graph-based segmentation hasbecome a research hotspot due to the simple structure andsolid theories In graph-based segmentation the image ismodeled as a weighted undirected graph Zhang et al [28]applied the discriminative graph-cut approach to segment-ing tumors after discrimination between tumors and thebackground via a trained classifier In 2014 Zhou et al [29]proposed a novel US image segmentation method based onmean shift and graph cuts (MSGC) It uses mean shift filterto improve the homogeneity and applies graph-cut methodwhose energy function combines region- and edge-basedinformation to segment US images The result showed thatthe method is rapid and efficient Huang et al [30] designeda novel comparison criterion for pairwise subregions whichtakes local statistics into account to make their method morerobust to noises and hence it was named as robust graph-based (RGB) segmentationmethodThe experimental resultsshowed that accurate segmentation results can be obtained bythis method However two significant parameters determin-ing the segmentation result should be set empirically and fordifferent images they need to be adjusted by repeated teststo obtain good segmentation results In 2013 Huang et al[31] proposed an improvement method for RGB by usingPSO algorithm to optimize the two significant parametersautomatically The between-class variance which denotesthe difference between the reference region and its adjacentregions was introduced as the objective function and themethod was named as parameter-automatically optimizedrobust graph-based (PAORGB) segmentation

The active contour model (ACM) more widely knownas snake is another very popular segmentation method forUS images and has been massively used as an edge-basedsegmentation method This approach attempts to minimize

the energy associated with the initial contour as the sum ofthe internal and external energies During the deformationprocess the force is calculated from the internal energyand external energy The internal energy derived from thecontour model is used to control the shape and regularityof the contour and the external energy derived from theimage feature is used to extract the contour of the desiredobject A 3D snake technique was used by Chang et al [36]to obtain the tumor contour for the pre- and postoperativemalignant tumor excision Jumaat et al [37] applied theBalloon Snake to segment the mass in the US image takenfrom Malaysian population To overcome the curvature andtopology problems in the ACM level set has been employedto improve the US image segmentation Sarti et al [38]used a level set formulation to search the minimal value ofACM and the segmentation results showed that their modelis efficient and flexible Gao et al [40] combined an edgestopping term and an improved gradient vector flow snakein the level set framework to robustly cope with noise and toaccurately extract the low contrast andor concave ultrasonictumor boundaries Liu et al [39] proposed a novel probabilitydensity difference-based active contour method for ultra-sound image segmentation In 2010 Li et al [43] proposed thenew level set evolution model Distance Regularized Level SetEvolution (DRLSE) in which it adds a distance regularizationterm over traditional level set evolution to eliminate theneed for reinitialization in evolution process and improve theefficiency Some researchers combined texture informationwith other methods for US images segmentation [44ndash47]In 2016 Lang et al [44] used a multiscale texture identifierintegrated in a level set framework to capture the spiculatedboundary and showed improved segmentation result

However most of the above methods are purely region-based or edge-based For region-based methods they usehomogeneity statistics and low-level image features likeintensity texture and histogram to assign pixels to objectsTwo pixels would be assigned to the same object if theyare similar in value and connected to each other in somesense The problem of applying these approaches to USimages is that without considering any shape informationthey would classify pixels within the acoustic shadow asbelonging to the tumor while posterior acoustic shadowingis a common artifact in US images [48 49] For edge-based methods (ACM) they are used to handle only theROI not the entire image Although they can obtain theprecise contour of the desired object they are sensitive tonoise and heavily rely on the suitable initial contour whichis very difficult to generate properly Also the deformationprocedure is very time-consuming Therefore segmentationapproaches which integrate region-based techniques andedge-based techniques have been proposed to obtain accuratesegmentation results for US images [50ndash55] Chang et al[50] introduced the concepts of 3D stick 3D morphologicprocess and 3D ACM The 3D stick is used to reducethe speckle noise and enhance the edge information in 3DUS images Then the 3D morphologic process is used toobtain the initial contour of the tumor for the 3D ACMHuang and Chen [51 52] utilized the watershed transformand ACM to overcome the natural properties of US images

BioMed Research International 3

Original USimage

Touching the stop condition

Coping TCI

Bilateral filtering

Histogram equalization

Mean shift filtering

Particles traversal

RGBsegmentation

Calculating thevalue of the

objective function

Update particle

Getting the optimalparameter and

segmentation result

Postprocessing

Final result

Preprocess

YesNo

Figure 1 Flowchart of the proposed approach

(ie speckle noise and tissue-related textures) to segmenttumors precisely In their methods the watershed transformis performed as the automatic initial contouring procedurefor the ACM Then the ACM automatically determines theexquisite contour of the tumor Wang et al [55] presented amultiscale framework for US image segmentation based onspeckle reducing anisotropic diffusion and geodesic activecontour In general the region-based technique is used togenerate the initial contour for the edge-based techniqueThe experimental results of these approaches indicate thataccurate segmentation results can be obtained by combiningregion-based and edge-based information of the US image

In this paper we propose a novel segmentation schemefor US images based on the RGB segmentation method [30]and particle swarm optimization (PSO) algorithm [56 57]In this scheme the PSO is used to optimally set the twosignificant parameters determining the segmentation resultautomatically To combine region-based and edge-basedinformation we consider the optimization as amultiobjectiveproblem comparing with PAORGB We use multiobjectiveoptimization method (maximizing the difference betweentarget and background improving the uniformity within thetarget region and considering the edge gradient) to improvethe segmentation performance We take the uniformity ofthe region and the information of the edge as the objectivecontents in the process of optimization First one rectangleis manually selected to determine the ROI on the originalimage However because of the low contrast and speckle

noises of US images the ROI image is filtered by a bilateralfilter and contrast-enhanced by histogram equalization Nextpyramid mean shift is executed on the enhanced image toimprove homogeneity A novel objective function consistingof three parts corresponding to region-based and edge-basedinformation is designed in the PSOWith the optimization ofPSO the RGB segmentationmethod is performed to segmentthe ROI image Finally the segmented image is processedby morphological opening and closing to refine the tumorcontour

This paper is organized as follows Section 2 introducesthe proposedmethod in detail Next the experimental resultsand comparisons among different methods are presented inSection 3 Finally we provide some discussion and draw theconclusion in Section 4

2 Methods

In this paper our method is called multi-objectivelyoptimized robust graph-based (MOORGB) segmentationmethod which utilizes PSO algorithm to optimize thetwo key parameters of RGB segmentation method Inthe MOORGB a multiobjective optimization functionwhich combines region-based and edge-based information isdesigned in the PSO to optimize the RGB The flowchart ofthe proposed approach is shown in Figure 1 In the rest of thissection we introduce each step in the proposed approach indetail


(a) (b)

Figure 2 Example of extracting a TCI (a) the original image and (b) the TCI image

21 Preprocessing

211 Cropping Tumor Centered ROI According to [11] a goodsegmentation method for clinical US images should havetaken advantage of a priori knowledge to improve the seg-mentation result due to the relatively low quality In additionit is hard to describe the segmentation result quantitativelywithout any a priori knowledge therefore it is difficult todesign objective function(s) without any a priori knowledgeTherefore we employ the a priori knowledge used in [31]namely asking the operator to roughly extract a relativelysmall rectangular ROI (in which the focus of interest is fullycontained and located in the central part) from the US imageIn this way interferences from other unrelated regions can bereduced as much as possible making the segmentation easierand more efficient Besides it gives useful a priori knowledgefor design of objective function(s) Such a ROI is called tumorcentered image (TCI) in this paper and Figure 2 shows howa TCI is extracted from a US image

212 Bilateral Filtering Because of diverse interferences (egattenuation speckle shadow and signal dropout) in USimages speckle reduction is necessary to improving thequality of US images Bilateral filter [58] which has provento be an efficient and effective method for speckle reductionis adopted in the MOORGB

213 Histogram Equalization To improve the contrast of USimages histogram equalization is conducted to enhance thefiltered TCI Histogram equalization maps one distribution(the given histogram of intensity values in the filtered TCI)to another distribution (a wider and uniform distributionof intensity values) The classical histogram equalizationmethod [59] is used in the MOORGB

214 Mean Shift Filtering After contrast enhancement weimprove the homogeneity by performing mean shift filteringMean shift filtering is based on mean shift clustering overgrayscale and can well improve the homogeneity of USimages and suppress the speckle noise and tissue-related

textures [60] Figure 3 shows the preprocessing results of theimage

22 RGB Segmentation Method Given an image which isinitially regarded as a graph the RGB method [30] aims tomerge spatially neighboring pixels with similar intensitiesinto a minimal spanning tree (MST) which correspondsto a subgraph (ie a subregion in the image) The imageis therefore divided into several subregions (ie a forest ofMSTs) Obviously the step for merging pixels into a MSTis the key determining the final segmentation results Anovel pairwise region comparison predicate was proposed inthe RGB to determine whether or not a boundary betweentwo subgraphs should be eliminated Given a graph 119866 =(119881 119864) the resulting predicate 119863(1198621 1198622) which comparesintersubgraph differenceswithwithin-subgraph differences isformulated as follows [30]

119863(1198621 1198622)

= false if Dif (1198621 1198622) gt MInt (1198621 1198622)true other

(1)

Dif (1198621 1198622) = 1003816100381610038161003816120583 (1198621) minus 120583 (1198622)1003816100381610038161003816 (2)

MInt (1198621 1198622)= min (120590 (1198621) + 120591 (1198621) 120590 (1198622) + 120591 (1198622))

(3)

120591 (119862) = 119896|119862| sdot (1 + 1

120572 sdot 120573) 120573 = 120583 (119862)120590 (119862) (4)

where Dif(1198621 1198622) is the difference between two subgraphs1198621 and 1198622 isin 119881 MInt(1198621 1198622) represents the smallest internaldifference of1198621 and1198622 120583(119862) denotes the average intensity of119862 120590(119862) is the standard deviation of119862 and 120591(119862) is a thresholdfunction of 119862 while 120572 and 119896 are positive parameters When119896 increases 120591 increases as well and the regions merge moreeasily On the contrary when 120572 increases 120591 decreases andhence the regions are merged less easily


(a) (b)

(c) (d)

Figure 3 Example of preprocessing (a) the TCI image (b) the image after the bilateral filtering (c) the image after histogram equalizationand (d) the image after mean shift filtering

Based on the pairwise region comparison predicate thegeneral procedures of segmenting an image are as follows

Step 1 Construct a graph 119866 = (119881 119864) for the US image tobe segmented In 119866 each pixel corresponds to a vertex andeach edge connects two spatially neighboring vertices Theedge weight is defined by the absolute intensity differencebetween two adjacent pixels Initially each vertex is regardedas a subgraph and all edges constituting the edge set 119864 areinvalid

Step 2 Sort the edges in 119864 in nondescending order accordingto the edge weight and set 119902 = 1Step 3 Pick the 119902th edge in the sorted 119864 If the 119902th edge isan invalid edge (connecting two different subgraphs) and theboundary between these two subgraphs can be eliminatedaccording to the pairwise region comparison predicate asmathematically expressed in (1)ndash(4) then merge these twosubgraphs into a larger subgraph and set this edge valid Let119902 = 119902 + 1Step 4 Repeat Step 3 until all edges in 119864 are traversed

When all edges are traversed a forest including a numberof MSTs can be obtained Each MST corresponds to asubregion in the image However the selection of 120572 and 119896in (4) can significantly influence RGBrsquos segmentation results[30] As shown in Figure 4 it can be seen that inappropriateselections of 120572 and 119896 can lead to under- or oversegmentationIn [30] two significant parameters in RGB segmentationalgorithm were empirically selected and usually manuallyassigned by testing repeatedly to achieve acceptable resultsIt cannot be fixed for real clinical application because goodselections of 120572 and 119896 may be quite different for differentimages due to the diversity of US images

Therefore the PAORGB was proposed to optimize thesetwo parameters and to achieve a good selection of them

automatically for each US image [31] However only region-based information and only one optimization goal (maxi-mizing the difference between target and background) havebeen used Although the PAORGB can obtain good seg-mentation results for some US images its performance isnot adequately stable Therefore we propose the MOORGBwhich usesmultiobjective optimizationmethod (maximizingthe difference between target and background improving theuniformity within the target region considering the edgegradient) to improve the segmentation performance Themethod makes comprehensive consideration of edge-basedand region-based information

23 PSO Optimization of Parameters PSO algorithm is anevolutionary computation techniquemimicking the behaviorof flying birds and their means of information exchange[56 57] In PSO each particle represents a potential solutionand the particle swarm is initialized with a population of ran-domuniform individuals in the search space PSO searchesthe optimal solution by updating positions of particles in anevolutionary manner

Suppose that there are 119899119901 solutions each of whichcorresponds to a particle and the position (ie the solution)and velocity of the 119894th particle (119894 = 1 119899119901) are representedby two 119898-dimensional (119898 = 2 in our study) vectors (ie119909119894 = (1199091198941 1199091198942 119909119894119898) and V119894 = (V1198941 V1198942 V119894119898) resp)Position 119909 is a vector and in our method 119909 = (119896 120572) VelocityV means the varied distance of the position at every itera-tion 1198881 1199031 1198882 1199032 and 119908 are scalars According to specificissues one or more objective functions are used to evaluatefitness of each particle and then the comparison criterionis employed to obtain superior particles Assume that 119901119894 =(1199011198941 1199011198942 119901119894119898) is the best position visited until themomentof the 119894th particle during the update process and the globalbest position of the whole particle swarm obtained so far isindicated as 119901119892 = (1199011198921 1199011198922 119901119892119898) At each generationeach particle updates its velocity and position according to


Originalimages

(a)

k = 2000

훼 = 002훼 = 1훼 = 40

(b)

k = 3000 k = 100 k = 20

훼 = 002

(c)

Figure 4 Influence of 120572 and 119896 the image of (a) is the original image the image of (b) shows the segmentation result with different 120572 and (c)shows the segmentation result with different 119896

the following equations after 119901119894 and 119901119892 are acquired throughfitness evaluation and the comparison criterion [56]

V119905+1119894 = 119908V119905119894 + 11988811199031 (119901119905119894 minus 119909119905119894) + 11988821199032 (119901119905119892 minus 119909119905119894) (5)

119909119905+1119894 = 119909119905119894 + V119905+1119894 (6)

119908119905 = 119908max minus (119908max minus 119908min)119879max

lowast 119905 (7)

where 119905 is the generation number119879max is themaximum itera-tion119908119905 is the value of the 119905th iteration119908 is the inertia weight1198881 and 1198882 are positive parameters known as accelerationcoefficients determining the relative influence of cognitionand social components and 1199031 and 1199032 are independentlyuniformly distributed random variables within the range of(0 1) The value of 119908 describes the influence of historicalvelocity The method with higher 119908 will have stronger globalsearch ability and themethodwith smaller119908will has strongerlocal search ability At the beginning of the optimization

process we initially set 119908 to a large value in order to makebetter global exploration and gradually decrease it to findoptimal or approximately optimal solutions and thus reducethe number of the iterations Hence we let119908 decrease linearlyfrom 1 towards 02 as shown in (7) We set 119908max = 1 119908min =02 and119879max = 200 In (5)119908V119905119894 represents the influence of theprevious velocity on the current one and 11988811199031(119901119905119894 minus 119909119905119894) repre-sents the personal experience while 11988821199032(119901119905119892 minus 119909119905119894) representsthe collaborative effect of particles which pulls particles tothe global best solution thewhole particle swarmhas found sofar As suggested in [41] we set 1198881 = 05 and 1198882 = 05 and makepersonal experience and collaborative effect of particles playthe same important role in optimization as shown in Figure 5

To conclude at each generation the velocity and positionof each particle are updated according to (5) and its positionis updated by (6) At each time any better position is storedfor the next generation Then each particle adjusts its posi-tion based on its own ldquoflyingrdquo experience and the experienceof its companions whichmeans that if one particle arrives at a


Ptg

Xt+1i

Pti

Xti

Vt+1i

Vti

Figure 5 Update of the particle

new promising position all other particles then move closerto it This process is repeated until a satisfactory solution isfound or a predefined number of iterative generations is met

The general procedure is summarized as follows

Step 1 Properly set the size of particle swarm and ran-domlyuniformly initialize them according to the searchspace In this study the size of particle swarm is 119899119901 = 200 andthe particles are uniformly initialized According to the workin [30] 119896 varies from 100 to 4000 and 120572 varies from 0001 to4000 which form the search space

Step 2 Traverse all particles in each traversal (ie at eachgeneration) each particle is evaluated through the objectivefunction and 119901119894 and 119901119892 are acquired according to thecomparison criterion

Step 3 Update the velocity and position of each particleaccording to (5) and (6) As suggested in [57] we set 1198881 = 05and 1198882 = 05 and let 119908 decrease linearly from 1 towards 02

Step 4 Repeat Steps 2 and 3 until all particles converge tothe predefined extent or the iterative number arrives at thepredefined maximum The predefined extent in this study isthat119901119892 does not change for four iterations and themaximumiteration is set to119873 = 200 empirically

24 The Proposed Objective Function in the PSO At eachtime we use RGB to segment the TCI according to theinformation (ie 120572 and 119896) of one particle According tothe a priori knowledge that the focus of interest is locatedin the central part of TCI the central subregion with thecentral pixel of TCI is the possible tumor regionThis centralsubregion is defined as the reference region which varieswiththe setting of 120572 and 119896 and the reference region is the expectedtumor region when 120572 and 119896 are optimally set Figure 6 givesan example of reference region (the original image is shownin Figure 2)

In MOORGB a novel objective function consistingof three parts corresponding to region-based and edge-based information is adopted Based on the above a prioriknowledge these three parts that is between-class variancewithin-class variance and average gradient are defined as

follows Compared with PAORGB we add two objectivefunctions within-class variance and average gradient It isnot enough to optimize parameters just relying on edgeinformation or region information for segmentationWe takethe uniformity of the region and the information of the edgeas the objective contents in the optimization process

241 Between-Class Variance Inspired by the idea of Otsursquosmethod [61] which utilizes the difference between subregionsto quantitatively describe the segmentation result to select anoptimal threshold the between-class variance (119881119861) is definedas follows

119881119861 =119896

sum119894=1

119875 (119862119894) (120583 (119862119894) minus 120583 (119862Ref))2 (8)

where 119881119861 denotes the sum of difference of mean intensitybetween subregion 119862 and the reference region 119896 denotes thenumber of subregions adjacent to the reference region and120583(119862) denotes the mean intensity of subregion 119862 while 119875(119862119894)denotes the proportion of the 119894th subregion in the whole TCIand is expressed as

119875 (119862119894) =10038161003816100381610038161198621198941003816100381610038161003816|TCI| (9)

where |119862119894| is the number of pixels in the 119894th subregion and|TCI| is the number of pixels in the whole TCI

From the definition 119881119861 denotes the difference betweenthe reference region and its adjacent regions Since thereference region corresponds to the interested tumor regionin the US image it is easy to understand that maximizing 119881119861can well overcome oversegmentation By the way this is theonly part adopted in PAORGB [31]

242Within-Class Variance Theaim of image segmentationis to segment a region with uniformity which is alwaysthe target object out of the background [62] Thereforeconsidering the uniformity within the target region wecome up with another part called within-class variance (119881119882)defined as follows

119881119882 = arctan ((1 1003816100381610038161003816119862Ref1003816100381610038161003816) sum|119862Ref |119894=1 (119868119894 minus 120583 (119862Ref))2)119875 (119862Ref)

119875 (119862Ref) =1003816100381610038161003816119862Ref

1003816100381610038161003816|TCI|

(10)

where |119862Ref | is the number of pixels in the reference regionand 119868119894 denotes the intensity of the 119894th pixel while 120583(119862Ref )denotes the mean intensity of the reference region and |TCI|is the number of pixels in the whole TCI Since the min-imizing of pure within-class variance (1|119862Ref |) sum|119862Ref |119894=1 (119868119894 minus120583(119862Ref ))2 will lead to oversegmentation we add 119875(119862Ref ) tosuppress it Since the value range of (1|119862Ref |) sum119862Ref119894=1 (119868119894 minus120583(119862Ref ))2 is much larger than the value range of 119875(119862Ref ) weuse arctan operation to make them comparable From thedefinition 119881119882 denotes the difference within the referenceregion and the undersegmentation problem can be wellovercome by minimizing 119881119882


(a) (b)

Figure 6 Example of reference region (a) a segmented image by the RGB and (b) the reference region of (a)

243 Average Gradient As mentioned above the purpose ofsegmenting US images is to support the latter analysis andclassification in the CAD system and a wealth of useful andsignificant information for classification is contained in thecontour of the focus Accordingly to achieve the objective ofacquiring better tumor contours another part called averagegradient (119866119860) is employed in our objective functionWith theinspiration of the definition of energy in ACM 119866119860 is definedas follows

119866119860 = 1119898119898

sum119894=1

10038161003816100381610038161198661198941003816100381610038161003816 (11)

where 119898 is the number of pixels included in the edge of thereference region and 119866119894 denotes the gradient (calculated bythe Sobel operator) of the 119894th pixel Sobel operator is an edgedetection operator based on 2D spatial gradient measure-ment It can smooth the noise of the image and provide moreaccurate edge direction information than Prewitt andRobertsoperators [59] 119866119860 denotes the average energy of the edge ofthe reference region

Maximizing average gradient 119866119860 obtains more accuratecontour and avoids oversegmentation If there were overseg-mentation the reference region would be included withinthe real target area real target area would be a relativelyhomogeneous region within which every partitioned smallerregion would have smaller 119866119860 Consequently to increase 119866119860would force the contour of the reference region to movetowards that of the real target area Very often the edgesof the targets in the US image are not sufficiently clear andsharp such that we cannot use only 119866119860 in the objectivefunction We take average gradient into account as one of thethree objective functions in optimization process to improvethe segmentation result In ACM the initial edge is forcedto approach the real edge through maximizing the energySimilar to ACM maximizing GA can force the contour of thereference region to approach the real contour of the tumor

244The Final Objective Function Based on the above threeparts the objective function is defined as follows

119865119874 = 119886 lowast 119881119861119891119861 minus 119887 lowast 119881119882

119891119882 + 119888 lowast 119866119860119891119860 (12)

119891119861 = 1119899119901119899119901

sum119894=1

119881119861119894 (13)

119891119882 = 1119899119901119899119901

sum119894=1

119881119882119894 (14)

119891119860 = 1119899119901119899119901

sum119894=1

119866119860119894 (15)


119891119882 + 04 lowast 119866119860119891119860 (16)

where 119886 119887 119888 are the weights of different objective parts(119886 = 03 119887 = 03 119888 = 04 in our experiment they canbe adjusted as needed) The final objective function in theexperiment is defined as (16) 119881119861 119881119882 and 119866119860 are between-class variance within-class variance and average gradientrespectively 119891119861 119891119882 and 119891119860 are normalized factors while119899119901 = 200 is the size of particle swarm Because the valueranges of 119881119861 119881119882 and 119866119860 are quite different they should benormalized to be comparable For each US image 119891119861 119891119882and 119891119860 are calculated once after the uniform initializationof particle swarm but before the first iteration We try tomaximize 119865119874 by the PSO25 Postprocessing After the TCI is segmented by the RGBwith the optimal 120572 and 119896 obtained by the PSO we turnit into a binary image containing the object (tumor) andthe background (tissue) Next morphological opening andclosing are conducted to refine the tumor contour withopening to reduce the spicules and closing to fill the holesA 5 times 5 elliptical kernel is used for both opening and closing

26 The Proposed MOORGB Segmentation Method Assum-ing that the position and velocity of the 119894th particle in our caseare expressed as 119909119894 = (119896119894 120572119894) and V119894 = (V119896119894 V120572119894) respectivelythe general procedure ofMOORGB is summarized as follows

Step 1 Manually delineate TCI from the original US image

Step 2 Use the bilateral filter to do the speckle reduction forTCI

Step 3 Enhance the filtered TCI by histogram equalization toimprove the contrast

Step 4 Improve the homogeneity by performing pyramidmean shift filtering


Step 5 Uniformly initialize the particle swarm within thesearch space and let the iteration count 119902 = 0 and so on

Step 6 Let 119902 = 119902 + 1 traverse all 119899119901 particles in the 119902thtraversal RGB is performed with the position (ie 119909119894 =(119896119894 120572119894)) of each particle then evaluate the segmentation resultwith the objective function 119865119874 and obtain 119901119894 and 119901119892 bycomparing values of 119865119874 for updating each particle (includingposition and velocity) for next iteration

Step 7 Iteratively repeat Step 6 until convergence (ie 119901119892remains stable for 4 generations) or 119902 = 119873 (119873 = 200 in thispaper)

Step 8 After finishing the iteration the position of the glob-ally best particle (ie 119901119892) is namely the optimal setting of 120572and 119896 then get the final segmentation result by performingRGB with the optimal setting

Step 9 Turn the segmentation result into a binary imagethen get the final tumor contour by conducting morphologi-cal opening and closing

27 Experimental Methods We developed the proposedmethod with the C++ language using OpenCV 243 andVisualStudio 2010 and run it on a computer with 340GHzCPU and 120GBRAM To validate ourmethod experimentshave been conducted Our work is approved by HumanSubject Ethics Committee of South ChinaUniversity of Tech-nology In the dataset 100 clinical breast US images and 18clinical musculoskeletal US images with the subjectsrsquo consentforms were provided by the Cancer Center of Sun Yat-senUniversity andwere taken from anHDI 5000 SonoCT System(Philips Medical Systems) with an L12-5 50mm BroadbandLinear Array at the imaging frequency of 71MHzThe ldquotruerdquotumor regions of these US images were manually delineatedby an experienced radiologist who hasworked onUS imagingand diagnosis formore than ten yearsThe contour delineatedby only one doctor is not absolutely accurate because differentdoctors may give different ldquoreal contoursrdquo which is indeeda problem in the research Nevertheless the rich diagnosisexperience of the doctor has fitted the edge of every tumoras accurately as possible This dataset consists of 50 breastUS images with benign tumors 50 breast US images withmalignant tumors and 18 musculoskeletal US images withcysts (including 10 ganglion cysts 4 keratinizing cysts and4 popliteal cysts)

To demonstrate the advantages of the proposed methodbesides PAORGB we also compared the method with theother two well-known segmentation methods (ie DRLSE[43] and MSGC [29]) DRLSE method an advanced level setevolution approach in recent years is applied to an edge-based active contour model for image segmentation It isan edge-based segmentation method that needs to set initialcontour manually The initial contour profoundly affectsthe final segmentation result MSGC is a novel graph-cutmethod whose energy function combines region- and edge-based information to segment US images It also needs tocrop tumor centered ROI Among the three comparative

Cs(i)

Cr(i)

Co

Figure 7 An illustration of computation principle for ARE

methods DRLSE is an edge-based method PAORGB is aregion-basedmethod andMSGC is a compoundmethod Tomake a comparison of computational efficiency the methodsPAORGB and MOORGB were programmed in the samesoftware system As such the fourmethods were run with thesame hardware configurationThe ROI is all the same for thefour segmentation methods

To quantitatively measure the experiment results fourcriteria (ie averaged radial error (ARE) true positive vol-ume fraction (TPVF) false positive volume fraction (FPVF)and false negative volume fraction (FNVF)) were adopted inthis studyTheARE is used for the evaluation of segmentationperformance by measuring the average radial error of asegmented contour with respect to the real contour which isdelineated by an expert radiologist As shown in Figure 7 itis defined as

ARE (119899) = 1119899119899minus1

sum119894=0

1003816100381610038161003816119862119904 (119894) minus 119862119903 (119894)10038161003816100381610038161003816100381610038161003816119862119903 (119894) minus 1198621199001003816100381610038161003816 times 100 (17)

where 119899 is the number of radial rays and set to 180 in ourexperiments while 119862119900 represents the center of the ldquotruerdquotumor region which is delineated by the radiologist and 119862119904(119894)denotes the location where the contour of the segmentedtumor region crosses the 119894th ray while 119862119903(119894) is the locationwhere the contour of the ldquotruerdquo tumor region crosses the 119894thray

In addition TPVF FPVF and FNVF were also used inthe evaluation of the performance of segmentation methodsTPVF means true positive volume fraction indicating thetotal fraction of tissue in the ldquotruerdquo tumor region withwhich the segmented region overlaps FPVF means falsepositive volume fraction denoting the amount of tissuefalsely identified by the segmentation method as a fraction ofthe total amount of tissue in the ldquotruerdquo tumor region FNVFmeans false negative volume fraction denoting the fractionof tissue defined in the ldquotruerdquo tumor region that is missedby the segmentation method In our study the ldquotruerdquo tumorregion is delineated by the radiologist Figure 8 shows the


Table 1 Quantitative segmentation results of 50 breast US images with benign tumors

Methods ARE () TPVF () FPVF () FNVF ()Our method 1109 plusmn 1247 8560 plusmn 1371 451 plusmn 2018 1440 plusmn 1371PAORGB [26] 1647 plusmn 2141 8164 plusmn 2994 1052 plusmn 2940 1836 plusmn 2994DRLSE [46] 1137 plusmn 1304 9360 plusmn 1687 1442 plusmn 2433 640 plusmn 1687MSGC [24] 1576 plusmn 1318 7534 plusmn 1625 251 plusmn 1460 2466 plusmn 1624

Table 2 Quantitative segmentation results of 50 breast US images with malignant tumors

Methods ARE () TPVF () FPVF () FNVF ()Our method 1041 plusmn 1362 8491 plusmn 1639 443 plusmn 1901 1509 plusmn 1639PAORGB 1912 plusmn 2763 7498 plusmn 2749 1016 plusmn 3709 2502 plusmn 2749DRLSE 1584 plusmn 1534 9531 plusmn 1975 2405 plusmn 2068 469 plusmn 1975MSGC 1552 plusmn 2266 7412 plusmn 1512 293 plusmn 1317 2588 plusmn 1512

areas corresponding to TPVF FPVF and FNVF Accordinglysmaller ARE FPVF and FNVF and larger TPVF indicatebetter segmentation performance TPVF FPVF and FNVFare defined by

TPVF = 119860119898 cap 119860119899119860119898

FPVF = 119860119899 minus 119860119898 cap 119860119899119860119898

FNVF = 119860119898 minus 119860119898 cap 119860119899119860119898

(18)

where 119860119898 is the area of the ldquotruerdquo tumor region delineatedby the radiologist and 119860119899 is the area of the tumor regionobtained by the segmentation algorithm

3 Experimental Results and Discussion

31 Qualitative Analysis In this paper we present the seg-mentation results for five tumors Five US images withthe segmentation results are shown in Figures 9ndash13 Thequantitative segmentation results on US images are shown inTables 1 2 3 and 4 Figures 9(a) 10(a) 11(a) 12(a) and 13(a)show original B-mode US images for two benign tumors twomalignant tumors and onemusculoskeletal cyst respectivelyAfter preprocessing the original images the segmentationresults using the MOORGB are illustrated in Figures 9(b)10(b) 11(b) 12(b) and 13(b) those using the PAORGB inFigures 9(d) 10(d) 11(d) 12(d) and 13(d) those using theDRLSE in Figures 9(e) 10(e) 11(e) 12(e) and 13(e) and thoseusing the MSGC in Figures 9(f) 10(f) 11(f) 12(f) and 13(f)

In Figures 9ndash13 we can see that our method achievedthe best segmentation results compared with the other threemethods and the contour generated by our method isquite close to the real contour delineated by the radiologistUndersegmentation happens in Figures 9(d) and 10(d) butnot in Figures 9(b) and 10(b) and oversegmentation hap-pens in Figures 12(d) and 13(d) but not in Figures 12(b)

FPVFTPVFFNVF

AnAm

Figure 8 The areas corresponding to TPVF FPVF and FNVFrespectively119860119898 indicates the ldquotruerdquo contour delineated by the radi-ologist and 119860119899 denotes the contour obtained by the segmentationalgorithm

and 13(b) Comparing with the PAORGB the MOORGBimproves the segmentation results obviously avoiding theundersegmentation and oversegmentation more effectivelyRegional uniformity has been significantly improved andthe edge has been smoother The reason is that the within-class variance and average gradient are introduced into theobjective function of MOORGB by combining region- andedge-based information The segmentation results of theMSGC are better than those of the PAORGB and DRLSEsince MSGC is also a compound method (region energy andboundary energy are both included in its energy function)andmany preprocessing techniques are adopted As shown inFigures 9(e) 10(e) 11(e) 12(e) and 13(e) the DRLSE can onlyroughly detect the tumor contour and the detected contoursare irregular The reason is that it depends on edge-basedinformation and is sensitive to speckle noise and sharp edgehence it captures sharp edge easily and leads to boundaryleakage and undersegmentation


(a) (b) (c)

(d) (e) (f)

Figure 9 Segmentation results for the first benign breast tumor (a) A breast US image with a benign tumor (the contour is delineated by theradiologist) (b) The result of MOORGB (c) The final result of MOORGB (d)The result of PAORGB (e) The result of DRLSE (f) The resultof MSGC

Table 3 Quantitative segmentation results of 18 musculoskeletal US images with cysts


Table 4 Overall quantitative segmentation results of 118 US images

Methods ARE () TPVF () FPVF () FNVF () Averaged computingtime (s)

Our method 1077 plusmn 1722 8534 plusmn 1669 448 plusmn 3426 1467 plusmn 1667 5054PAORGB 1827 plusmn 3703 7889 plusmn 3024 1051 plusmn 3574 2110 plusmn 3023 71978DRLSE 1284 plusmn 1682 9407 plusmn 1851 1797 plusmn 2803 592 plusmn 2378 593MSGC 1546 plusmn 2627 7561 plusmn 1774 29 plusmn 4241 2437 plusmn 2463 0123

32 Quantitative Analysis Table 4 shows the quantitativecomparisons of different segmentation approaches on thewhole dataset Similarly we show the quantitative segmen-tation results of the benign tumors malignant tumors andcysts in Tables 1 2 and 3 respectively

Comparing Tables 1 and 3 with Table 2 it is shown thatall four segmentation methods perform better on benigntumors and musculoskeletal cysts than on malignant tumorson thewhole indicating that the boundaries of benign tumorsand musculoskeletal cysts are more significant than those ofmalignant tumors The shape of the benign tumor is moreregular and similar to circle or ellipse The shape of themalignant tumor is irregular and usually lobulated with burrsin the contour The segmentation result of the malignanttumor is worse than benign tumors because the contour of

malignant tumor is less regular and less homogenous thanthat of benign tumor

FromTable 4 it is seen that ourmethod achieved the low-est ARE (1077) Due to the undersegmentation the DRLSEgot the highest TPVF (9407) and FPVF (1797) indicatingthe high ratio of false segmentationTheMSGCgot the lowestFPVF (29) which indicates the low ratio of false segmen-tation and is the fastest method (0123 s) among the fourmethods However it got the lowest TPVF (7561) showingoversegmentation in a way Comparing with the originalPAORGB (as shown in Table 4) our method improves thesegmentation result obviously achieving higher TPVF andlower ARE FPVF and FNVF Although MOORGB couldnot achieve the best performance in all evaluation indices itgot the best overall performance Comparing with DRLSE


(a) (b) (c)

(d) (e) (f)

Figure 10 Segmentation results for the second benign breast tumor (a) A breast US image with a benign tumor (the contour is delineatedby the radiologist) (b)The result of MOORGB (c)The final result of MOORGB (d)The result of PAORGB (e)The result of DRLSE (f)Theresult of MSGC

(a) (b) (c)

(d) (e) (f)

Figure 11 Segmentation results for the first malignant breast tumor (a) A breast US image with a malignant tumor (the contour is delineatedby the radiologist) (b)The result of MOORGB (c)The final result of MOORGB (d)The result of PAORGB (e)The result of DRLSE (f)Theresult of MSGC

our method was 645 lower than it in TPVF but 875better than it in FPVF obtaining better overall performanceComparing with MSGC our method was 158 higher thanit in FPVF but 993 better than it in TPVF obtaining better

overall performance and avoiding oversegmentation in a wayAs shown in Table 4 our method is faster than the PAORGBIt is because the convergence condition in our method is thatldquo119901119892 remains stable for 4 generationsrdquo rather than that ldquothe


(a) (b) (c)

(d) (e) (f)

Figure 12 Segmentation results for the second malignant breast tumor (a) A breast US image with a malignant tumor (the contour isdelineated by the radiologist) (b) The result of MOORGB (c) The final result of MOORGB (d) The result of PAORGB (e) The result ofDRLSE (f) The result of MSGC

(a) (b) (c)

(d) (e) (f)

Figure 13 Segmentation results for the keratinizing cyst (a) A musculoskeletal US image with a cyst (the contour is delineated by theradiologist) (b) The result of MOORGB (c) The final result of MOORGB (d)The result of PAORGB (e) The result of DRLSE (f) The resultof MSGC

updating of 119896 is below 1 and that of 120572 is below 000001 for allthe particles in an experimentrdquo in the PAORGB [31]

33 The Influence of the Weight Our method synthesizesthree optimization objective functions (between-class vari-ance 119881119861 within-class variance 119881119882 and average gradient119866119860) Thus the weight values of three objective parts (ie 119886

119887 and 119888) are introduced Figure 14 shows the comparisonof experimental results with different weight values FromFigures 14(a) 14(b) and 14(c) and Table 5 we can see thatwhen the weight values of the three objective functions arealmost the same three optimization objectives play nearlyequal roles in the optimization process making the algorithmnot only region-based but also edge-based When one of


(a) The original image

(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1

Figure 14 The segmentation results with different weights (119886 119887 119888)


Table 5 Quantitative segmentation results of 15 US images with different weight values

Methods ARE () TPVF () FPVF () FNVF ()119886 = 04 119887 = 03 119888 = 03 1067 8561 452 1470119886 = 03 119887 = 04 119888 = 03 1071 8560 451 1469119886 = 03 119887 = 03 119888 = 04 1069 8560 452 1469119886 = 06 119887 = 02 119888 = 02 1047 8551 449 1491119886 = 02 119887 = 06 119888 = 02 1059 8552 450 1473119886 = 02 119887 = 02 119888 = 06 1074 8564 480 1475119886 = 08 119887 = 02 119888 = 02 1112 8497 544 1539119886 = 02 119887 = 08 119888 = 02 1125 8503 578 1541119886 = 02 119887 = 02 119888 = 08 1123 8477 561 1528119886 = 1 119887 = 0 119888 = 0 1292 8339 678 1607119886 = 0 119887 = 1 119888 = 0 6984 9886 15472 046119886 = 0 119887 = 0 119888 = 1 897 8779 1049 1793

the three weight values is overlarge it would not be able toreflect the three optimization results evenly hence leading tooversegmentation or undersegmentation As shown in Fig-ures 14(k) 14(l) and 14(m) if one of the three weight valuesequals one the proposed method degenerated into a singleobjective optimization algorithm and the optimization goalis only one role which cannot avoid oversegmentation andundersegmentation effectively Through analyzing the influ-ence of different weight values of objective functions and ourrepeated experiments we can set the system parameters as119886 = 119887 = 03 119888 = 04 The final objective function is describedas (16) which can make segmentation system work well

4 Conclusions

In this paper we propose a novel segmentation scheme forUS images based on the RGB and PSO methods In thisscheme the PSO is used to optimally set the two significantparameters in the RGB for determining the segmentationresult automatically To combine region-based and edge-based information we consider the optimization as a mul-tiobjective problem First because of the low contrast andspeckle noises of US images the ROI image is filteredby a bilateral filter and contrast-enhanced by histogramequalization and then pyramid mean shift is executed on theenhanced image to improve homogeneity A novel objectivefunction consisting of three parts corresponding to region-based and edge-based information is adopted by PSO Thebetween-class variance denotes the difference between thereference region and its adjacent regions The within-classvariance denotes the difference within the reference regionand the undersegmentation problem can be well overcomeby minimizing it Between-class variance and within-classvariance reflect the regional information The average gradi-ent denotes the average energy of the edge of the referenceregion and maximizing it can force the contour of thereference region to approach the real contour of the tumorAverage gradient reflects the edge-based information of theimage Three optimization objectives play important rolesin the optimization process making the algorithm achieve

the corresponding segmentation effect not only region-based but also edge-based With the optimization of PSORGB is performed to segment the ROI image Finally thesegmented image is processed by morphological openingand closing to refine the tumor contour To validate ourmethod experiments have been conducted on 118 clinicalUS images including breast US images and musculoskeletalUS images The segmentation results of our method havebeen compared with the corresponding results obtained bythree existing approaches and four metrics have been usedto measure the segmentation performanceThe experimentalresults show that our method could successfully segment USimages and achieved the best segmentation results comparedwith the other three methods MSGC PAORGB and DRLSEThe contour generated by our method was closer to thereal contour delineated by the radiologist The MOORGBobtained the lowest ARE and better overall performance inTPVF FPVF and FNVF avoiding the undersegmentationand oversegmentation more effectively

However the step to obtain TCI (as shown in Figure 2)requires userrsquos participation which may result in significantinfluence on the following segmentation To obtain accept-able segmentation results the operator should be well expe-rienced in examining US images and identifying suspiciouslesions in clinical practices Moreover the TCI should becarefully delineated to achieve the full lesion region withpartial surrounding tissues and the interested lesion regionmust be located in the central part Consequently how toautomatically extract TCI from the BUS image is one ofour future studies In addition the computation time is stillfar away from real-time applications Accordingly makingefforts to reduce the computation time by adopting parallelprocessing techniques is also part of our future work Besidesadopting our segmentation method in real CAD systems tovalidate the whole performance will be included in our futurework

Conflicts of Interest

The authors declare that they have no conflicts of interest


Authorsrsquo Contributions

Yaozhong Luo and Longzhong Liu have contributed equallyto this work

Acknowledgments

This work was partially supported by National Natu-ral Science Foundation of China (nos 61372007 and61571193) Guangzhou Key Lab of Body Data Science (no201605030011) and Guangdong Provincial Science and Tech-nology Program-International Collaborative Projects (no2014A050503020)

References

[1] B Sahiner H-P Chan M A Roubidoux et al ldquoMalignant andbenign breast masses on 3D US volumetric images effect ofcomputer-aided diagnosis on radiologist accuracyrdquo Radiologyvol 242 no 3 pp 716ndash724 2007

[2] C-M Chen Y-H Chou K-C Han et al ldquoBreast lesionson sonograms computer-aided diagnosis with nearly setting-independent features and artificial neural networksrdquo Radiologyvol 226 no 2 pp 504ndash514 2003

[3] K Drukker M L Giger K Horsch M A Kupinski C JVyborny and E B Mendelson ldquoComputerized lesion detectionon breast ultrasoundrdquo Medical Physics vol 29 no 7 pp 1438ndash1446 2002

[4] Q Li W Zhang X Guan Y Bai and J Jia ldquoAn improvedapproach for accurate and efficient measurement of commoncarotid artery intima-media thickness in ultrasound imagesrdquoBioMed Research International vol 2014 Article ID 740328 8pages 2014

[5] B O Anderson R Shyyan A Eniu et al ldquoBreast cancer inlimited-resource countries an overview of the breast healthglobal initiative 2005 guidelinesrdquo Breast Journal vol 12 no 1pp S3ndashS15 2006

[6] V Naik R S Gamad and P P Bansod ldquoCarotid artery segmen-tation in ultrasound images and measurement of intima-mediathicknessrdquo BioMed Research International vol 2013 Article ID801962 10 pages 2013

[7] Y L Huang D R Chen and Y K Liu ldquoBreast cancer diag-nosis using image retrieval for different ultrasonic systemsrdquo inProceedings of the International Conference on Image ProcessingICIP pp 2957ndash2960 Institute of Electrical and ElectronicsEngineers Singapore 2004

[8] H D Cheng J Shan W Ju Y Guo and L Zhang ldquoAutomatedbreast cancer detection and classification using ultrasoundimages a surveyrdquo Pattern Recognition vol 43 no 1 pp 299ndash317 2010

[9] Y Li W Liu X Li Q Huang and X Li ldquoGA-SIFT a newscale invariant feature transform for multispectral image usinggeometric algebrardquo Information Sciences vol 281 pp 559ndash5722014

[10] J Shi S Zhou X Liu Q Zhang M Lu and T Wang ldquoStackeddeep polynomial network based representation learning fortumor classification with small ultrasound image datasetrdquoNeurocomputing vol 194 pp 87ndash94 2016

[11] J A Noble and D Boukerroui ldquoUltrasound image segmenta-tion a surveyrdquo IEEE Transactions on Medical Imaging vol 25no 8 pp 987ndash1010 2006

[12] J Peng J Shen and X Li ldquoHigh-Order Energies for StereoSegmentationrdquo IEEE Transactions on Cybernetics vol 46 no 7pp 1616ndash1627 2016

[13] M Xian Y Zhang H-D Cheng F Xu and J Ding ldquoNeutro-connectedness cutrdquo IEEE Transactions on Image Processing vol25 no 10 pp 4691ndash4703 2016

[14] P N T Wells andM Halliwell ldquoSpeckle in ultrasonic imagingrdquoUltrasonics vol 19 no 5 pp 225ndash229 1981

[15] G Xiao M Brady J A Noble and Y Zhang ldquoSegmentationof ultrasound B-mode images with intensity inhomogeneitycorrectionrdquo IEEE Transactions on Medical Imaging vol 21 no1 pp 48ndash57 2002

[16] S Y Joo W K Moon and H C Kim ldquoComputer-aided diag-nosis of solid breast nodules on ultrasound with digital imageprocessing and artificial neural networkrdquo in Proceedings of the26th Annual International Conference of vol 2 pp 1397ndash1400San Francisco CA USA 2004

[17] K Horsch M L Giger C J Vyborny and L A Venta ldquoPerfor-mance of Computer-Aided Diagnosis in the Interpretation ofLesions on Breast Sonographyrdquo Academic Radiology vol 11 no3 pp 272ndash280 2004

[18] M H Yap E A Edirisinghe and H E Bez ldquoFully automaticlesion boundary detection in ultrasound breast imagesrdquo inMedical Imaging 2007 Image Processing vol 6512 of Proceedingsof SPIE p I5123 San Diego Calif USA 2007

[19] N A M Isa S Sabarudin U K Ngah and K Z ZamlildquoAutomatic detection of breast tumours fromultrasound imagesusing the modified seed based region growing techniquerdquo inProceedings of the 9th International Conference on Knowledge-Based Intelligent Information and Engineering Systems vol3682 pp 138ndash144 Springer La Trobe University MelbourneAustralia 2005

[20] J Shan H D Cheng and Y Wang ldquoA novel segmentationmethod for breast ultrasound images based on neutrosophic l-means clusteringrdquoMedical Physics vol 39 no 9 pp 5669ndash56822012

[21] H B Kekre and P Shrinath ldquoTumour delineation using statisti-cal properties of the breast us images and vector quantizationbased clustering algorithmsrdquo International Journal of ImageGraphics and Signal Processing vol 5 no 11 pp 1ndash12 2013

[22] W K Moon C-M Lo R-T Chen et al ldquoTumor detection inautomated breast ultrasound images using quantitative tissueclusteringrdquo Medical Physics vol 41 no 4 Article ID 0429012014

[23] D Boukerroui O Basset N Guerin and A Baskurt ldquoMultires-olution texture based adaptive clustering algorithm fo breastlesion segmentationrdquoEuropean Journal of Ultrasound vol 8 no2 pp 135ndash144 1998

[24] C-M Chen Y-H Chou C S K Chen et al ldquoCell-competitionalgorithm a new segmentation algorithm for multiple objectswith irregular boundaries in ultrasound imagesrdquo Ultrasound inMedicine and Biology vol 31 no 12 pp 1647ndash1664 2005

[25] B Deka and D Ghosh ldquoUltrasound image segmentation usingwatersheds and regionmergingrdquo inProceedings of the IETVisualInformation Engineering (VIE rsquo06) p 6 Bangalore India 2006

[26] L Zhang andM Zhang ldquoA fully automatic image segmentationusing an extended fuzzy setrdquo in Proceedings of the InternationalWorkshop on Computer Science for Environmental Engineeringand EcoInformatics vol 159 pp 412ndash417 Springer KunmingPeoplersquos Republic of China 2011


[27] WGomezA RodriguezWCA Pereira andA F C InfantosildquoFeature selection and classifier performance in computer-aided diagnosis for breast ultrasoundrdquo in Proceedings of the 10thInternational Conference and Expo on Emerging Technologies fora Smarter World CEWIT IEEE Melville NY USA 2013

[28] J Zhang S K Zhou S Brunke C Lowery and D ComaniciuldquoDatabase-guided breast tumor detection and segmentation in2D ultrasound imagesrdquo in Medical Imaging 2010 Computer-AidedDiagnosis vol 7624 ofProceedings of SPIE p 7 SanDiegoCalif USA February 2010

[29] Z ZhouWWu SWu et al ldquoSemi-automatic breast ultrasoundimage segmentation based on mean shift and graph cutsrdquoUltrasonic Imaging vol 36 no 4 pp 256ndash276 2014

[30] Q-H Huang S-Y Lee L-Z Liu M-H Lu L-W Jin and A-H Li ldquoA robust graph-based segmentation method for breasttumors in ultrasound imagesrdquo Ultrasonics vol 52 no 2 pp266ndash275 2012

[31] Q Huang X Bai Y Li L Jin and X Li ldquoOptimized graph-based segmentation for ultrasound imagesrdquo Neurocomputingvol 129 pp 216ndash224 2014

[32] Q Huang F Yang L Liu and X Li ldquoAutomatic segmentationof breast lesions for interaction in ultrasonic computer-aideddiagnosisrdquo Information Sciences vol 314 pp 293ndash310 2015

[33] H Chang Z Chen Q Huang J Shi and X Li ldquoGraph-based learning for segmentation of 3D ultrasound imagesrdquoNeurocomputing vol 151 no 2 pp 632ndash644 2015

[34] Q Huang B Chen J Wang and T Mei ldquoPersonalized videorecommendation through graph propagationrdquo in Transactionson Multimedia Computing Communications and Applicationsvol 10 of 4 pp 1133ndash1136 ACM 2012

[35] Y Luo S Han and Q Huang ldquoA novel graph-based segmenta-tion method for breast ultrasound imagesrdquo in Proceedings of theDigital Image Computing Techniques andApplications (DICTA)pp 1ndash6 IEEE 2016

[36] R-F Chang W-J Wu C-C Tseng D-R Chen and W KMoon ldquo3-D Snake for US in Margin Evaluation for Malig-nant Breast Tumor Excision Using Mammotomerdquo in Proceed-ings of the IEEE Transactions on Information Technology inBiomedicine vol 7 of 3 pp 197ndash201 2003

[37] A K Jumaat W E Z W A Rahman A Ibrahim and R Mah-mud ldquoSegmentation of masses from breast ultrasound imagesusing parametric active contour algorithmrdquo inProceedings of theInternational Conference on Mathematics Education ResearchICMER pp 640ndash647 Malacca Malaysia 2010

[38] A Sarti C Corsi E Mazzini and C Lamberti ldquoMaximumlikelihood segmentation of ultrasound images with rayleighdistributionrdquo IEEE Transactions on Ultrasonics Ferroelectricsand Frequency Control vol 52 no 6 pp 947ndash960 2005

[39] B Liu H D Cheng J Huang J Tian X Tang and J LiuldquoProbability density difference-based active contour for ultra-sound image segmentationrdquo Pattern Recognition vol 43 no 6pp 2028ndash2042 2010

[40] L Gao X Liu and W Chen ldquoPhase- and GVF-based level setsegmentation of ultrasonic breast tumorsrdquo Journal of AppliedMathematics vol 2012 Article ID 810805 pp 22ndashID 8108052012

[41] B Wang X Gao J Li X Li and D Tao ldquoA level set methodwith shape priors by using locality preserving projectionsrdquoNeurocomputing vol 170 pp 188ndash200 2015

[42] B N Li J Qin R Wang and M Wang ldquoSelective level setsegmentation using fuzzy region competitionrdquo IEEE Access vol4 pp 4777ndash4788 2016

[43] C Li C Xu C Gui and M D Fox ldquoDistance regularized levelset evolution and its application to image segmentationrdquo IEEETransactions on Image Processing vol 19 no 12 pp 3243ndash32542010

[44] I Lang M Sklair-Levy and H Spitzer ldquoMulti-scale texture-based level-set segmentation of breast B-mode imagesrdquo Com-puters in Biology amp Medicine vol 72C pp 30ndash42 2016

[45] AMadabhushi andDNMetaxas ldquoCombining low- high-leveland empirical domain knowledge for automated segmentationof ultrasonic breast lesionsrdquo IEEE Transactions on MedicalImaging vol 22 no 2 pp 155ndash169 2003

[46] D R Chen R F ChangW J KuoM C Chen and Y L HuangldquoDiagnosis of breast tumors with sonographic texture analysisusing wavelet transform and neural networksrdquo Ultrasound inMedicine and Biology vol 28 no 10 pp 1301ndash1310 2002

[47] Y Guo A Sengur and J W Tian ldquoA novel breast ultrasoundimage segmentation algorithm based on neutrosophic similar-ity score and level setrdquo Computer Methods and Programs inBiomedicine vol 123 pp 43ndash53 2016

[48] A T Stavros D Thickman C L Rapp M A Dennis S HParker and G A Sisney ldquoSolid breast nodules use of sonog-raphy to distinguish between benign and malignant lesionsrdquoRadiology vol 196 no 1 pp 123ndash134 1995

[49] W Leucht andD Leucht ldquoTeachingAtlas of BreastUltrasoundrdquopp 24ndash38 Thieme Medical Stuttgart Germany 2000

[50] R F Chang W J WuW K Moon W M Chen W Lee and DR Chen ldquoSegmentation of breast tumor in three-dimensionalultrasound images using three-dimensional discrete active con-tour modelrdquoUltrasound in Medicine and Biology vol 29 no 11pp 1571ndash1581 2003

[51] Y LHuang andDRChen ldquoWatershed segmentation for breasttumor in 2-D sonographyrdquoUltrasound in Medicine and Biologyvol 30 no 5 pp 625ndash632 2004

[52] Y L Huang and D R Chen ldquoAutomatic contouring for breasttumors in 2-D sonographyrdquo in Proceedings of the 27th AnnualInternational Conference of the IEEE Engineering in Medicineand Biology Society (IEEE) pp 3225ndash3228 IEEE ShanghaiChina 2005

[53] Y L Huang Y R Jiang D R Chen andW K Moon ldquoLevel setcontouring for breast tumor in sonographyrdquo Journal of DigitalImaging vol 20 no 3 pp 238ndash247 2007

[54] M Aleman-Flores L Alvarez and V Caselles ldquoTexture-oriented anisotropic filtering and geodesic active contours inbreast tumor ultrasound segmentationrdquo Journal of Math Imag-ing Vision vol 28 no 1 pp 81ndash97 2007

[55] W M Wang L Zhu J Qin Y P Chui B N Li and P A HengldquoMultiscale geodesic active contours for ultrasound imagesegmentation using speckle reducing anisotropic diffusionrdquoOptics and Lasers in Engineering vol 54 pp 105ndash116 2014

[56] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of the International Conference on Neural Networks(ICNN rsquo95) vol 4 pp 1942ndash1948 IEEE Perth Australia 1995

[57] K E Parsopoulos and M N Vrahatis ldquoParticle swarm opti-mization method in multiobjective problemsrdquo pp 603ndash607Proceedings of the ACM symposium Madrid Spain March2002

[58] M Elad ldquoOn the origin of the bilateral filter and ways toimprove itrdquo IEEE Transactions on Image Processing vol 11 no10 pp 1141ndash1151 2002

[59] R C Gonzalez and R E Woods Digital Image ProcessingPrentice Hall Upper Saddle River NJ USA 2nd edition 2001


[60] D Comaniciu and P Meer ldquoMean shift analysis and applica-tionsrdquo in Proceedings of the 7th IEEE International Conferenceon Computer Vision vol 2 pp 1197ndash1203 Kerkyra GreeceSeptember 1999

[61] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 pp 23ndash27 1975

[62] N R Pal and D Bhandari ldquoImage thresholding some newtechniquesrdquo Signal Processing vol 33 no 2 pp 139ndash158 1993

Submit your manuscripts athttpswwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


thresholding-based methods [16ndash18] clustering-based meth-ods [19ndash23] watershed-based methods [24ndash27] graph-based methods [28ndash35] and active contour models [36ndash42]Thresholding is one of the frequently used segmentationtechniques for themonochrome image Yap et al [18] adoptedthe thresholding segmentation to separate the lesion regionfrom the background before detecting the initial boundaryvia edge detection Clustering is a classification technique andhas been successfully applied to image segmentation basedon similarity between image regions or pixels Isa et al [19]used the moving 119896-means clustering to automatically selectthe seed and proposed a modified seed based region growingalgorithm to detect the edge Shan et al [20] used a novel neu-trosophic clustering approach to detect the lesion boundaryMoon et al [22] used the fuzzy C-means (FCM) clusteringto extract the tumor candidates in their CAD system Thewatershed transformation which is frequently used in thesegmentation of grey scale images considers the gradientmagnitude of an image as a topographic surface Chen etal [24] employed the two-pass watershed transformationsto generate the cells and proposed a region-based approachcalled cell-competition algorithm to simultaneously segmentmultiple objects in a sonogram L Zhang and M Zhang [26]used an extended fuzzy watershed method to segment USimages fully automatically The experiments showed that theproposedmethod could get good results on blurryUS images

In the last few years graph-based segmentation hasbecome a research hotspot due to the simple structure andsolid theories In graph-based segmentation the image ismodeled as a weighted undirected graph Zhang et al [28]applied the discriminative graph-cut approach to segment-ing tumors after discrimination between tumors and thebackground via a trained classifier In 2014 Zhou et al [29]proposed a novel US image segmentation method based onmean shift and graph cuts (MSGC) It uses mean shift filterto improve the homogeneity and applies graph-cut methodwhose energy function combines region- and edge-basedinformation to segment US images The result showed thatthe method is rapid and efficient Huang et al [30] designeda novel comparison criterion for pairwise subregions whichtakes local statistics into account to make their method morerobust to noises and hence it was named as robust graph-based (RGB) segmentationmethodThe experimental resultsshowed that accurate segmentation results can be obtained bythis method However two significant parameters determin-ing the segmentation result should be set empirically and fordifferent images they need to be adjusted by repeated teststo obtain good segmentation results In 2013 Huang et al[31] proposed an improvement method for RGB by usingPSO algorithm to optimize the two significant parametersautomatically The between-class variance which denotesthe difference between the reference region and its adjacentregions was introduced as the objective function and themethod was named as parameter-automatically optimizedrobust graph-based (PAORGB) segmentation

The active contour model (ACM) more widely knownas snake is another very popular segmentation method forUS images and has been massively used as an edge-basedsegmentation method This approach attempts to minimize

the energy associated with the initial contour as the sum ofthe internal and external energies During the deformationprocess the force is calculated from the internal energyand external energy The internal energy derived from thecontour model is used to control the shape and regularityof the contour and the external energy derived from theimage feature is used to extract the contour of the desiredobject A 3D snake technique was used by Chang et al [36]to obtain the tumor contour for the pre- and postoperativemalignant tumor excision Jumaat et al [37] applied theBalloon Snake to segment the mass in the US image takenfrom Malaysian population To overcome the curvature andtopology problems in the ACM level set has been employedto improve the US image segmentation Sarti et al [38]used a level set formulation to search the minimal value ofACM and the segmentation results showed that their modelis efficient and flexible Gao et al [40] combined an edgestopping term and an improved gradient vector flow snakein the level set framework to robustly cope with noise and toaccurately extract the low contrast andor concave ultrasonictumor boundaries Liu et al [39] proposed a novel probabilitydensity difference-based active contour method for ultra-sound image segmentation In 2010 Li et al [43] proposed thenew level set evolution model Distance Regularized Level SetEvolution (DRLSE) in which it adds a distance regularizationterm over traditional level set evolution to eliminate theneed for reinitialization in evolution process and improve theefficiency Some researchers combined texture informationwith other methods for US images segmentation [44ndash47]In 2016 Lang et al [44] used a multiscale texture identifierintegrated in a level set framework to capture the spiculatedboundary and showed improved segmentation result

However most of the above methods are purely region-based or edge-based For region-based methods they usehomogeneity statistics and low-level image features likeintensity texture and histogram to assign pixels to objectsTwo pixels would be assigned to the same object if theyare similar in value and connected to each other in somesense The problem of applying these approaches to USimages is that without considering any shape informationthey would classify pixels within the acoustic shadow asbelonging to the tumor while posterior acoustic shadowingis a common artifact in US images [48 49] For edge-based methods (ACM) they are used to handle only theROI not the entire image Although they can obtain theprecise contour of the desired object they are sensitive tonoise and heavily rely on the suitable initial contour whichis very difficult to generate properly Also the deformationprocedure is very time-consuming Therefore segmentationapproaches which integrate region-based techniques andedge-based techniques have been proposed to obtain accuratesegmentation results for US images [50ndash55] Chang et al[50] introduced the concepts of 3D stick 3D morphologicprocess and 3D ACM The 3D stick is used to reducethe speckle noise and enhance the edge information in 3DUS images Then the 3D morphologic process is used toobtain the initial contour of the tumor for the 3D ACMHuang and Chen [51 52] utilized the watershed transformand ACM to overcome the natural properties of US images


Original USimage


Coping TCI

Bilateral filtering



Particles traversal

RGBsegmentation


objective function

Update particle


segmentation result

Postprocessing

Final result

Preprocess

YesNo






2 Methods



(a) (b)


21 Preprocessing







119863(1198621 1198622)


(1)

Dif (1198621 1198622) = 1003816100381610038161003816120583 (1198621) minus 120583 (1198622)1003816100381610038161003816 (2)

MInt (1198621 1198622)= min (120590 (1198621) + 120591 (1198621) 120590 (1198622) + 120591 (1198622))

(3)

120591 (119862) = 119896|119862| sdot (1 + 1

120572 sdot 120573) 120573 = 120583 (119862)120590 (119862) (4)



(a) (b)

(c) (d)











Originalimages

(a)

k = 2000

훼 = 002훼 = 1훼 = 40

(b)

k = 3000 k = 100 k = 20

훼 = 002

(c)



V119905+1119894 = 119908V119905119894 + 11988811199031 (119901119905119894 minus 119909119905119894) + 11988821199032 (119901119905119892 minus 119909119905119894) (5)

119909119905+1119894 = 119909119905119894 + V119905+1119894 (6)


lowast 119905 (7)





Ptg

Xt+1i

Pti

Xti

Vt+1i

Vti












119881119861 =119896

sum119894=1

119875 (119862119894) (120583 (119862119894) minus 120583 (119862Ref))2 (8)


119875 (119862119894) =10038161003816100381610038161198621198941003816100381610038161003816|TCI| (9)





119875 (119862Ref) =1003816100381610038161003816119862Ref

1003816100381610038161003816|TCI|

(10)



(a) (b)



119866119860 = 1119898119898

sum119894=1

10038161003816100381610038161198661198941003816100381610038161003816 (11)





119891119882 + 119888 lowast 119866119860119891119860 (12)

119891119861 = 1119899119901119899119901

sum119894=1

119881119861119894 (13)

119891119882 = 1119899119901119899119901

sum119894=1

119881119882119894 (14)

119891119860 = 1119899119901119899119901

sum119894=1

119866119860119894 (15)


119891119882 + 04 lowast 119866119860119891119860 (16)















Cs(i)

Cr(i)

Co




ARE (119899) = 1119899119899minus1

sum119894=0










TPVF = 119860119898 cap 119860119899119860119898

FPVF = 119860119899 minus 119860119898 cap 119860119899119860119898

FNVF = 119860119898 minus 119860119898 cap 119860119899119860119898

(18)





FPVFTPVFFNVF

AnAm




(a) (b) (c)

(d) (e) (f)












(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Original USimage


Coping TCI

Bilateral filtering



Particles traversal

RGBsegmentation


objective function

Update particle


segmentation result

Postprocessing

Final result

Preprocess

YesNo






2 Methods



(a) (b)


21 Preprocessing







119863(1198621 1198622)


(1)

Dif (1198621 1198622) = 1003816100381610038161003816120583 (1198621) minus 120583 (1198622)1003816100381610038161003816 (2)

MInt (1198621 1198622)= min (120590 (1198621) + 120591 (1198621) 120590 (1198622) + 120591 (1198622))

(3)

120591 (119862) = 119896|119862| sdot (1 + 1

120572 sdot 120573) 120573 = 120583 (119862)120590 (119862) (4)



(a) (b)

(c) (d)











Originalimages

(a)

k = 2000

훼 = 002훼 = 1훼 = 40

(b)

k = 3000 k = 100 k = 20

훼 = 002

(c)



V119905+1119894 = 119908V119905119894 + 11988811199031 (119901119905119894 minus 119909119905119894) + 11988821199032 (119901119905119892 minus 119909119905119894) (5)

119909119905+1119894 = 119909119905119894 + V119905+1119894 (6)


lowast 119905 (7)





Ptg

Xt+1i

Pti

Xti

Vt+1i

Vti












119881119861 =119896

sum119894=1

119875 (119862119894) (120583 (119862119894) minus 120583 (119862Ref))2 (8)


119875 (119862119894) =10038161003816100381610038161198621198941003816100381610038161003816|TCI| (9)





119875 (119862Ref) =1003816100381610038161003816119862Ref

1003816100381610038161003816|TCI|

(10)



(a) (b)



119866119860 = 1119898119898

sum119894=1

10038161003816100381610038161198661198941003816100381610038161003816 (11)





119891119882 + 119888 lowast 119866119860119891119860 (12)

119891119861 = 1119899119901119899119901

sum119894=1

119881119861119894 (13)

119891119882 = 1119899119901119899119901

sum119894=1

119881119882119894 (14)

119891119860 = 1119899119901119899119901

sum119894=1

119866119860119894 (15)


119891119882 + 04 lowast 119866119860119891119860 (16)















Cs(i)

Cr(i)

Co




ARE (119899) = 1119899119899minus1

sum119894=0










TPVF = 119860119898 cap 119860119899119860119898

FPVF = 119860119899 minus 119860119898 cap 119860119899119860119898

FNVF = 119860119898 minus 119860119898 cap 119860119899119860119898

(18)





FPVFTPVFFNVF

AnAm




(a) (b) (c)

(d) (e) (f)












(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)


21 Preprocessing







119863(1198621 1198622)


(1)

Dif (1198621 1198622) = 1003816100381610038161003816120583 (1198621) minus 120583 (1198622)1003816100381610038161003816 (2)

MInt (1198621 1198622)= min (120590 (1198621) + 120591 (1198621) 120590 (1198622) + 120591 (1198622))

(3)

120591 (119862) = 119896|119862| sdot (1 + 1

120572 sdot 120573) 120573 = 120583 (119862)120590 (119862) (4)



(a) (b)

(c) (d)











Originalimages

(a)

k = 2000

훼 = 002훼 = 1훼 = 40

(b)

k = 3000 k = 100 k = 20

훼 = 002

(c)



V119905+1119894 = 119908V119905119894 + 11988811199031 (119901119905119894 minus 119909119905119894) + 11988821199032 (119901119905119892 minus 119909119905119894) (5)

119909119905+1119894 = 119909119905119894 + V119905+1119894 (6)


lowast 119905 (7)





Ptg

Xt+1i

Pti

Xti

Vt+1i

Vti












119881119861 =119896

sum119894=1

119875 (119862119894) (120583 (119862119894) minus 120583 (119862Ref))2 (8)


119875 (119862119894) =10038161003816100381610038161198621198941003816100381610038161003816|TCI| (9)





119875 (119862Ref) =1003816100381610038161003816119862Ref

1003816100381610038161003816|TCI|

(10)



(a) (b)



119866119860 = 1119898119898

sum119894=1

10038161003816100381610038161198661198941003816100381610038161003816 (11)





119891119882 + 119888 lowast 119866119860119891119860 (12)

119891119861 = 1119899119901119899119901

sum119894=1

119881119861119894 (13)

119891119882 = 1119899119901119899119901

sum119894=1

119881119882119894 (14)

119891119860 = 1119899119901119899119901

sum119894=1

119866119860119894 (15)


119891119882 + 04 lowast 119866119860119891119860 (16)















Cs(i)

Cr(i)

Co




ARE (119899) = 1119899119899minus1

sum119894=0










TPVF = 119860119898 cap 119860119899119860119898

FPVF = 119860119899 minus 119860119898 cap 119860119899119860119898

FNVF = 119860119898 minus 119860119898 cap 119860119899119860119898

(18)





FPVFTPVFFNVF

AnAm




(a) (b) (c)

(d) (e) (f)












(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c) (d)











Originalimages

(a)

k = 2000

훼 = 002훼 = 1훼 = 40

(b)

k = 3000 k = 100 k = 20

훼 = 002

(c)



V119905+1119894 = 119908V119905119894 + 11988811199031 (119901119905119894 minus 119909119905119894) + 11988821199032 (119901119905119892 minus 119909119905119894) (5)

119909119905+1119894 = 119909119905119894 + V119905+1119894 (6)


lowast 119905 (7)





Ptg

Xt+1i

Pti

Xti

Vt+1i

Vti












119881119861 =119896

sum119894=1

119875 (119862119894) (120583 (119862119894) minus 120583 (119862Ref))2 (8)


119875 (119862119894) =10038161003816100381610038161198621198941003816100381610038161003816|TCI| (9)





119875 (119862Ref) =1003816100381610038161003816119862Ref

1003816100381610038161003816|TCI|

(10)



(a) (b)



119866119860 = 1119898119898

sum119894=1

10038161003816100381610038161198661198941003816100381610038161003816 (11)





119891119882 + 119888 lowast 119866119860119891119860 (12)

119891119861 = 1119899119901119899119901

sum119894=1

119881119861119894 (13)

119891119882 = 1119899119901119899119901

sum119894=1

119881119882119894 (14)

119891119860 = 1119899119901119899119901

sum119894=1

119866119860119894 (15)


119891119882 + 04 lowast 119866119860119891119860 (16)















Cs(i)

Cr(i)

Co




ARE (119899) = 1119899119899minus1

sum119894=0










TPVF = 119860119898 cap 119860119899119860119898

FPVF = 119860119899 minus 119860119898 cap 119860119899119860119898

FNVF = 119860119898 minus 119860119898 cap 119860119899119860119898

(18)





FPVFTPVFFNVF

AnAm




(a) (b) (c)

(d) (e) (f)












(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Originalimages

(a)

k = 2000

훼 = 002훼 = 1훼 = 40

(b)

k = 3000 k = 100 k = 20

훼 = 002

(c)



V119905+1119894 = 119908V119905119894 + 11988811199031 (119901119905119894 minus 119909119905119894) + 11988821199032 (119901119905119892 minus 119909119905119894) (5)

119909119905+1119894 = 119909119905119894 + V119905+1119894 (6)


lowast 119905 (7)





Ptg

Xt+1i

Pti

Xti

Vt+1i

Vti












119881119861 =119896

sum119894=1

119875 (119862119894) (120583 (119862119894) minus 120583 (119862Ref))2 (8)


119875 (119862119894) =10038161003816100381610038161198621198941003816100381610038161003816|TCI| (9)





119875 (119862Ref) =1003816100381610038161003816119862Ref

1003816100381610038161003816|TCI|

(10)



(a) (b)



119866119860 = 1119898119898

sum119894=1

10038161003816100381610038161198661198941003816100381610038161003816 (11)





119891119882 + 119888 lowast 119866119860119891119860 (12)

119891119861 = 1119899119901119899119901

sum119894=1

119881119861119894 (13)

119891119882 = 1119899119901119899119901

sum119894=1

119881119882119894 (14)

119891119860 = 1119899119901119899119901

sum119894=1

119866119860119894 (15)


119891119882 + 04 lowast 119866119860119891119860 (16)















Cs(i)

Cr(i)

Co




ARE (119899) = 1119899119899minus1

sum119894=0










TPVF = 119860119898 cap 119860119899119860119898

FPVF = 119860119899 minus 119860119898 cap 119860119899119860119898

FNVF = 119860119898 minus 119860119898 cap 119860119899119860119898

(18)





FPVFTPVFFNVF

AnAm




(a) (b) (c)

(d) (e) (f)












(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Ptg

Xt+1i

Pti

Xti

Vt+1i

Vti












119881119861 =119896

sum119894=1

119875 (119862119894) (120583 (119862119894) minus 120583 (119862Ref))2 (8)


119875 (119862119894) =10038161003816100381610038161198621198941003816100381610038161003816|TCI| (9)





119875 (119862Ref) =1003816100381610038161003816119862Ref

1003816100381610038161003816|TCI|

(10)



(a) (b)



119866119860 = 1119898119898

sum119894=1

10038161003816100381610038161198661198941003816100381610038161003816 (11)





119891119882 + 119888 lowast 119866119860119891119860 (12)

119891119861 = 1119899119901119899119901

sum119894=1

119881119861119894 (13)

119891119882 = 1119899119901119899119901

sum119894=1

119881119882119894 (14)

119891119860 = 1119899119901119899119901

sum119894=1

119866119860119894 (15)


119891119882 + 04 lowast 119866119860119891119860 (16)















Cs(i)

Cr(i)

Co




ARE (119899) = 1119899119899minus1

sum119894=0










TPVF = 119860119898 cap 119860119899119860119898

FPVF = 119860119899 minus 119860119898 cap 119860119899119860119898

FNVF = 119860119898 minus 119860119898 cap 119860119899119860119898

(18)





FPVFTPVFFNVF

AnAm




(a) (b) (c)

(d) (e) (f)












(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)



119866119860 = 1119898119898

sum119894=1

10038161003816100381610038161198661198941003816100381610038161003816 (11)





119891119882 + 119888 lowast 119866119860119891119860 (12)

119891119861 = 1119899119901119899119901

sum119894=1

119881119861119894 (13)

119891119882 = 1119899119901119899119901

sum119894=1

119881119882119894 (14)

119891119860 = 1119899119901119899119901

sum119894=1

119866119860119894 (15)


119891119882 + 04 lowast 119866119860119891119860 (16)















Cs(i)

Cr(i)

Co




ARE (119899) = 1119899119899minus1

sum119894=0










TPVF = 119860119898 cap 119860119899119860119898

FPVF = 119860119899 minus 119860119898 cap 119860119899119860119898

FNVF = 119860119898 minus 119860119898 cap 119860119899119860119898

(18)





FPVFTPVFFNVF

AnAm




(a) (b) (c)

(d) (e) (f)












(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
























Cs(i)

Cr(i)

Co




ARE (119899) = 1119899119899minus1

sum119894=0










TPVF = 119860119898 cap 119860119899119860119898

FPVF = 119860119899 minus 119860119898 cap 119860119899119860119898

FNVF = 119860119898 minus 119860119898 cap 119860119899119860119898

(18)





FPVFTPVFFNVF

AnAm




(a) (b) (c)

(d) (e) (f)












(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of






















TPVF = 119860119898 cap 119860119899119860119898

FPVF = 119860119899 minus 119860119898 cap 119860119899119860119898

FNVF = 119860119898 minus 119860119898 cap 119860119899119860119898

(18)





FPVFTPVFFNVF

AnAm




(a) (b) (c)

(d) (e) (f)












(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c)

(d) (e) (f)












(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)





(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c)

(d) (e) (f)


(a) (b) (c)

(d) (e) (f)







(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















(b) 119886 = 04 119887 = 03 119888 = 03 (c) 119886 = 03 119887 = 04 119888 = 04 (d) 119886 = 03 119887 = 03 119888 = 04

(e) 119886 = 06 119887 = 02 119888 = 02 (f) 119886 = 02 119887 = 06 119888 = 02 (g) 119886 = 02 119887 = 02 119888 = 06

(h) 119886 = 08 119887 = 01 119888 = 01 (i) 119886 = 01 119887 = 08 119888 = 01 (j) 119886 = 01 119887 = 01 119888 = 08

(k) 119886 = 1 119887 = 0 119888 = 0 (l) 119886 = 0 119887 = 1 119888 = 0 (m) 119886 = 0 119887 = 0 119888 = 1






4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of




















4 Conclusions









Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



















Acknowledgments


References






































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



























































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















A Novel Segmentation Approach Combining Region- and Edge...

Documents

Transcript of A Novel Segmentation Approach Combining Region- and Edge...