research article an interactive method based on...

Research ArticleAn Interactive Method Based on the Live Wire forSegmentation of the Breast in Mammography Images

Zhang Zewei Wang Tianyue Guo Li Wang Tingting and Xu Lu

School of Medical Imaging Tianjin Medical University Tianjin 300203 China

Correspondence should be addressed to Guo Li gl6290126com

Received 11 February 2014 Accepted 26 May 2014 Published 15 June 2014

Academic Editor Volkhard Helms

Copyright copy 2014 Zhang Zewei 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

In order to improve accuracy of computer-aided diagnosis of breast lumps the authors introduce an improved interactivesegmentation method based on Live Wire This paper presents the Gabor filters and FCM clustering algorithm is introducedto the Live Wire cost function definition According to the image FCM analysis for image edge enhancement we eliminate theinterference of weak edge and access external features clear segmentation results of breast lumps through improving Live Wire ontwo cases of breast segmentation data Compared with the traditional method of image segmentation experimental results showthat the method achieves more accurate segmentation of breast lumps and provides more accurate objective basis on quantitativeand qualitative analysis of breast lumps

1 Introduction

According to the Cancer Registration Annual Report 2013statistics [1] breast cancer has the highestmorbidity andmor-tality in women all over the world Over one-quarter (29) ofnew cancer cases in females are attributed to breast cancerin Americans Breast cancer is always the most commonmalignancy in China and the incidence has been increasingin recent years Therefore early detection has an impact onprognosis Because of the advantages of noninvasion lowcost simplicity and high resolution Mammography X-rayimaging is widely used in medical clinical diagnosis

At present it mainly depends on the professional physi-cian manual operation to distinguish the difference betweenbenign and malignant breast neoplasm and make the borderwhich leads to heavy workload poor efficiency With the helpof computer segmentation for the breast mass the doctor cananalyze quantitatively or qualitatively the lesions in ROI

However because breast tissues are complex and hetero-geneous and Mammography X-ray imaging has disadvan-tages of low contrast noises and artifacts the results areimperfect for traditional segmentation such as thresholdingregion growing watershed and FCM

In recent years a large number of researches have beendone inMammography X-ray imaging to improve the quality

of segmentation Paliwal presented a new approach forpreprocessing and segmenting out the infiltration and tumorregions from digital mammograms by using two techniquesinvolving iterative and noniterative algorithms of Delaunaytriangulation [2] Kurt et al have implemented mammogra-phy image enhancement using wavelet transform CLAHEand anisotropic diffusion filter then rough pectoral muscleextraction for false region reduction and better segmentationNext they have used Rough entropy to define a thresh-old and then fuzzy based microcalcification enhancementafter these microcalcifications have been segmented usingan iterative detection algorithm [3] Marungo and Taylorcombined thresholding with seeded region growing whichcan perform rapid automated breast region segmentationwith acceptable results in 9815 of the datasetrsquos 10411 images[4] Rahmati et al presented a computer-aided approach tosegmenting suspicious lesions in digitalmammograms basedon a novel maximum likelihood active contour model usinglevel sets (MLACMLS) [5] Yuvaraj and Ragupathy proposeda new computer aided mass segmentation and classificationscheme Five statistical features were selected for the classifi-cation of mass [6] Ullah et al proposed a new thresholdingmethod for segmentation that is based on DyWT momentpreserving principal and Gamma function [7] Sheet et alpresented a completely automatic system for detection and

Hindawi Publishing CorporationComputational and Mathematical Methods in MedicineVolume 2014 Article ID 954148 9 pageshttpdxdoiorg1011552014954148

2 Computational and Mathematical Methods in Medicine

segmentation of breast lesions in 2D ultrasound imagesThey employ random forests for learning of tissue specificprimal to discriminate breast lesions from surroundingnormal tissues This enables it to detect lesions of multipleshapes and sizes as well as discriminate between hypoechoiclesion from associated posterior acoustic shadowing [8]Tunali and Kilic presented a method for segmentation ofbenign and malignant masses found onmammograms Afteradding low valued pixels outside of mammogram Chan-Vese active contour algorithm with new stopping criteria wasimplemented [9] Song et al developed a new segmentationmethod by use of plane fitting and dynamic programmingwhich achieved a relatively high performance level Thenew segmentation method would be useful for improvingthe accuracy of computerized detection and classification ofbreast cancer in mammography [10] Kuo et al evaluated a3D lesion segmentation method which they had previouslydeveloped for breast CT and investigated its robustness onlesions on 3D breast ultrasound images [11] Cordeiro et alpresented the GrowCut technique to segment tumor regionsof digitizedmammograms available in theminimias database[12]

There are three kinds of interactive segmentation meth-ods [13] One is based on closed initial contour segmenta-tion method active contour model and level set methodsare widely used in computer vision they have the strongcapability of noise immunity but they are sensitive to theimage contrast Therefore these methods are still not perfectfor Mammography X-ray imaging Another is based on theinitial target points and background points randomwalk andgraph cut methods are the typical representative of this kindof medical segmentation methodsThe results are affected bythe location and the number of the initial target points andbackground points the result has significant influence on theuser experience and the repeatability is lowThe third is basedon the marked points on the edge Live Wire is one of theimportant representatives it has been widely implementedin medical image segmentation Live Wire method has theless manual intervention and high-precision for the blurboundary of Mammography X-ray imaging

This paper presents a semiautomatic segmentationmethod based on Gabor FCM algorithm for the costfunction in Live Wire algorithm In recent years theGabor FCM as a mathematical tool has been widely usedin image segmentation We present the Gabor filters andFCM clustering algorithm is calculated to the Live Wirecost function definition and according to the image FCManalysis for image edge enhancement to eliminate theinterference of weak edge access external features clearsegmentation results of breast lumps The procedure of theproposed algorithm is shown in Figure 1

2 Live Wire Algorithm

Live Wire is a famous classic interactive algorithm firstproposed by Mortensen and Barrett [14ndash16] The algorithmproduces reasonable cost function to construct the optimalpath between the starting point and ending point provided bythe user so as to achieve the purpose of edge segmentation

Achieve Mammography X-ray imaging

Obtain the ROI

Preprocessing (Gabor-FCM cost function)

Segmentation of breast lump by Live Wire

The results

Figure 1 The procedure flow of the proposed algorithm

The optimal path from each pixel is determined atinteractive speeds by computing an optimal spanning treeof the image using an efficient implementation of Dijkstrarsquosgraph searching algorithmThe basic idea is to formulate theimage as a weighted 356 graph where pixels represent nodeswith directed weighted edges connecting each pixel with its8 adjacent neighbors

The key issue of Live Wire algorithms is the constructionof the cost function Theoretically the cost function shouldcontain all the information about the lesion edges includinggradient magnitude function gradient direction functionand Laplace zero crossing point the cost function of classicalLive Wire algorithms is defined as follows

119862(119901 119902) = 119908119885119891119885(119902) + 119908

119866119891119866(119902) + 119908

119863119891119863(119901 119902) (1)

where each 119908 is the weight of the corresponding featurefunction Note that empirically weights of 119882

119885= 0 43

119882119866

= 0 43 119882119863

= 0 14 119891119885represents formulation

of Laplace zero crossing point 119891119866represents formulation of

gradient magnitude function and 119891119863represents formulation

of gradient direction functionLaplace zero crossing point (119891

119885) is a binarization Eigen

function which is used to localize edge The Laplacian imagezero-crossing corresponds to points of maximal gradientmagnitude Thus Laplacian zero-crossing represents ldquogoodrdquoedge properties and should have a low local cost If 119871(119902) is theLaplacian of an image at the pixel 119902 then

119891119885=

0 119871 (119902) = 0

1 119871 (119902) = 0(2)

where 119871(119902) is the Laplacian of the original image at a pixel 119902If a pixel is on a zero-crossing then the component cost forall links to that pixel is low otherwise it is high Howevergradient magnitude provides a direct correlation betweenedge strength and local cost Multiscale gradient magnitude(119891119866) is as follows

119891119866(119902) = 1 minus

119866 (119902)

max (119866) (3)

Computational and Mathematical Methods in Medicine 3

where 119866 is the gradient magnitude of the 119902 pixel If 119868119883

and 119868119884represent the partials of an image I in 119883 and 119884

respectively then the gradient magnitude is 119866 = radic1198682119909+ 1198682119910

Max(119866) represents themaximumvalue in the image119891119863(119901 119902)

represents the local cost on the directed gradient from pixel119901 to the neighboring pixel 119902

The formulation of the gradient direction feature cost iscalculated as follows

119891119863(119901 119902) =

2

3120587cos [119889

119901(119901 119902)]

minus1

+ cos [119889119902(119901 119902)]

minus1

(4)

where

119889119901(119901 119902) = 119863

1015840(119901) sdot 119871 (119901 119902)

119889119902(119901 119902) = 119871 (119901 119902) sdot 119863

1015840(119902)

(5)

where 119863(119901) is a unit vector of the gradient direction at apoint 119901 and defining 1198631015840(119901) as the unit vector perpendicularto119863(119901) which is defined as follows

1198631015840(119901) = (119868

119910(119901) minus119868

119909(119901)) (6)

where direction of the unit vector is defined by 119868119909and 119868119910

at a pixel 119901 The unit edge vector between pixels 119901 and 119902 isdefined as

119871 (119901 119902) = 119902 minus 119901 119863

1015840(119901) sdot (119902 minus 119901) ge 0

119901 minus 119902 1198631015840(119901) sdot (119902 minus 119901) lt 0

(7)

3 Improve the Cost Function

Classical Live Wire algorithm used the Laplace zero-crossingas the cost function Its segmentation results are sensitiveto the Gauss functionrsquos parameter 120575 Image edge becomesmore blurry with the value of 120575 increasing However whenthe value of 120575 is small the segmentation results cannotfully suppress the noise We cannot distinguish between astrong edge and a weak edge basis of above cost functionIn 2005 Yang et al [17] proposed canny operator instead ofLaplace zero-crossing in the cost function of Live Wire Thesegmentation results have improved however the use of edgedetection operator as the cost function make appear edgepoint error tracking pseudoedge discontinuous edges andedge information loss

Texture analysis plays an important role in variable ofapplications of medical image analysis for tumors segmen-tation based on local spatial variations of intensity Thepathological changes in lesion area result in the differencebetween the lesion and the surrounding normal tissue basedon the texture properties in medical image In this paper weintroduce aGabor FCMapproach to improving the LiveWirealgorithm to segment the breast lumps

31 Gabor Filter The Gabor filter is used as an edge detectorbased on texture to detect the texture and gradient of the

image First consider the Gabor filter which is defined asfollows

119866 = exp(minus1199062+ 120574V2

21205902) sdot cos (2120587119891119906 + 120593)

119906 = (119909119895minus 119909119894) sdot cos 120579 + (119910

119895minus 119910119894) sdot sin 120579

V = minus (119909119895minus 119909119894) sdot sin 120579 + (119910

119895minus 119910119894) sdot cos 120579

(8)

where (119909119894 119910119894) is the center location of the filter 120590 standard

deviation is utilized to normalize the size of filter and 119891 isthe bet space frequency bandwidth of the filter

The selection ofGabor parameters has been a focus for thelong-term research of Gabor based image processing A largenumber of works have been done on this important issue Inthis paper according to the symmetry in the filter we willchoose spatial orientations (120579) of 0∘ 30∘ 60∘ 90∘ 120∘ and150∘ The radial frequency (119891) is often selected as follows

119891119867= 025 +

2119894minus05

119873025 le 119891

119867lt 05

119891119871= 025 minus

2119894minus05

1198730 lt 119891 lt 025

119894 = 1 2 log(1198738)2

(9)

where 119891119867

is higher frequencies 119891119871is lower frequencies

and the frequency is normalized by the image size N butaccording to the symmetry of Fourier spectrum only halfpart [0 05] is considered in the paper

It can be seen that the curve of the selection ismuch flatterin the intermediate frequency This indicates that choice ofcentral frequency does have finer resolutions in the inter-mediate frequency (around 025) The most of the spectralenergy of natural image often centers at low frequency sothe choice of Gabor filters does not consider the very highfrequency band of image

32 Fuzzy c-Means Fuzzy c-means (FCM) is a methodof clustering which allows one piece of data to belong totwo or more clusters This algorithm works by assigningmembership to each data point corresponding to each clustercenter on the basis of distance between the cluster center andthe data pointThemore the data is near to the cluster centerthe more is its membership towards the particular clustercenter It is based on minimization of the following objectivefunction

119869FCM =

119888

sum

119894=1

119899

sum

119896=1

120583119898

119894119896

1003817100381710038171003817119909119896 minus V119894

1003817100381710038171003817

2 (10)

where V119894isin 119881 119881 = (V

1 V2 V

119888) isin 119877119888times119901 is the cluster center

of c 119880 = 120583119894119896 isin 119877

119888times119899 (120583119894119896

isin [0 1] sum119888

119894=1120583119894119896

= 1) is themembership matrix 119898 isin (1infin) is the weighting exponentas usual 119898 = 2 Membership matrix 119880 and cluster centroids119881 can be obtained after minimization of the object function


Membership matrix 119880 and cluster centroids 119881 are calculatedas follows

120583119894119896= (

119888

sum

119895=1

(

1003817100381710038171003817V119894 minus 119909119896

100381710038171003817100381710038171003817100381710038171003817V119895minus 119909119896

10038171003817100381710038171003817

)

2(119898minus1)

)

minus1

V119894=sum119899

119896=1120583119898

119894119896119909119896

sum119899

119896=1120583119898

119894119896

(11)

The condition of termination is as follows1003816100381610038161003816119869119905 minus 119869

119905minus1

1003816100381610038161003816 lt 120576 (12)

where 119869119905and 119869119905minus1

are the current objective function and lastobjective function 120576 gt 0 is a small number which our choiceis 120576 = 10

minus5 In our experiment we found that ten clustersare sufficient to separate the breast lumps centers the fuzzyboundary and others clearly

4 Experimental Results and the Analysis

41 Experimental Data In order to verify the effectivenessof the algorithm we applied the algorithm on two parts ofexperimental data the first part of data includes 16 caseswhich are provided by Tianjin Medical University GeneralHospital Radiology Department The images are clear andvoxels of each image are isotropous which ensure theaccuracy of the volume segmentationThe second part of thedata consists of 68 cases which are extracted from a publicdatabase of Digital Database for Screening Mammography(DDSM) provided by the US Army Medical Research andMateriel Command Breast Cancer Research Program Fourradiological diagnosticians on breast manually outlined thelumps on these data and obtained the diameter and thevolume of the lumps

42The Segmentation Results The results of our Gabor FCMLive Wire algorithm used in this paper proved that GaborFCM fits better than traditional Laplace zero crossing pointas cost function of Live Wire algorithm

Figure 2(a) is the original image of breast X-ray Mam-mography In this paper we captured the region of interestof breast lumps from each image to reduce the number ofunnecessary pixels for segregating out the lumpswhich couldimprove the segmentation accuracy of the algorithm and thecomputational efficiency

Figure 2(b) is the edge curves obtained by using tradi-tional LiveWire algorithm Laplace zero-crossing point is notgood at edge extraction of breast lumps with complex andchangeable organizational structure strong noise and lowcontrast which make the curves extracted by Laplace zero-crossing point discontinuous and complex and have a lot offalse contourThus traditional LiveWire is susceptible to thefalse contour

Figure 2(c) shows the results of edge extraction foreach lump image by applying Canny edge operator andthere are many edge curves in the results Particularly thesecond and third cases have invasive breast cancer which

was characterized by irregular shape and dim edge andsurrounded by funicular tufted or trabecular structuresand their segmentation results were discursive The detectedmultiple edge curves of the lumps using canny edge operatorhave great interference in extracting the accurate edge curvesto accurately locate the target boundary

Figure 2(d) is the processed result of Gabor FCM Wecan see that there are several gray value layers from thecenter of the breast lumps to the outside which can accuratelyreflect the general shape and size of the lump as well as thelayers and area of the halo around the lump Thus it providesconvenience for evaluating clinical breast tumor at differentrespects

Figure 2(e) shows the obtained edge curves extractedfrom Gabor FCM processed images These edge curves wereregularly closed around the breast lump region The curvesreplaced the results of using Laplace zero-crossing pointand were used as the improved cost function in Live Wirealgorithm by which we acquired the target contour which iscloser to the essence of tumor center or the halo around thetumor and improved the accuracy of the segmentation

Figure 3(a) is the segmentation results of Snake algo-rithm Figure 3(b) is the segmentation results of Level Setalgorithm Both of the algorithmsmistakenly segregated partof the normal breast tissue into the diseased tissue andcould not converge to the real edges Figure 3(c) is thesegmentation results of Graph cut algorithm Figure 3(d)is the segmentation results of Random Walk algorithmAlthough the segmentation results are ideal it requires alot of manual operations It needs 20 target points and 20background points to get ideal results In the fourth row wecan see that it failed to converge to the correct position at theedge of concave which has a large curvature Figure 3(e) is thesegmentation results of our improved algorithm in this paper

The normal tissues which were missegmented in Figures3(a) and 3(b) were segmented correctly Using the improvedalgorithm in this paper it only needs tomark 10 points on theedge of lumpwhich can greatly reduce themanual operationsand the impact of operational personnel experience onsegmentation results and improve the accuracy and efficiencyof segmentation

The top row in Figure 4 is the segmentation resultsof traditional Live Wire algorithm the second row is thesegmentation results of the improved algorithm in this paperIt can be seen that the segmentation results of the breastlump edges obtained by traditional Live Wire algorithm arenot ideal and are insensitive to fuzzy halo around the breastlump From the segmentation results of traditional Live Wirealgorithm in Figure 4(a) we can see that the obtained contouris too restrained to extract the whole focus areas and part ofthe lesions areas are divided into normal tissues by mistakewhich may have bad effect on clinical diagnosis

Figure 4(b) shows that traditional Live Wire algorithmcan only segment out the substantial part of the cancer-ous tissue and fail to classify the fuzzy halo around thesubstantial part of the cancerous tissue into focus lesionsareas which may result in omission of lesions tissue andreducing the algorithm value of assisted diagnose breastcancer In Figure 4(c) inconsistent with the smooth edge


(a) (b) (c) (d) (e)

Figure 2 Preprocess for the improvement of the cost function of Live Wire

of breast lumps the contour obtained from traditional LiveWire algorithm showed break angles which may be relatedto the false contour and discontinuous curves due to theshortcoming of Laplace zero-crossing point in detecting theedge of breast lumps which have complex and changeableorganizational structure strong noise and low contrast In

Figure 4(d) we can see that the edge of the lesion areaobtained by using traditional Live Wire algorithm does notconverge completely and thus diverges outwards excessivelyThe segmentation results contain some normal tissue withlow density which may exaggerate the patientrsquos condition inclinical diagnosis From the four figures we can see that the


(a) (b) (c) (d) (e)

Figure 3 The comparison with other interactive segmentation methods

segmentation results obtained by using improved algorithmin this study are more fitting to the edges of the lesion areasand highly sensitive to the lesion area having weak edgesThegenerated target contours are smooth and accurate which canmeet the demands of clinical diagnosis and improve the valueof the computer-aided diagnosis of the breast cancer

Figure 5 shows the image segmentation results of a set ofbreast cases by using the improved Live Wire algorithm inthis paper From the images of breast cases we can see thatthe lump shape of most of the breast cases is irregular andthe boundaries are fuzzy which cannot meet the request ofimage segmentation for automatic segmentationmethods Byusing the Live Wire interactive segmentation method in this

paper we quickly obtained the smooth and accurate lumpedges according to the gray information of the images Sothe improved Live Wire interactive segmentation methodcan reduce the doctorrsquos interactive operation and provideassistance to the doctors to analyze the breast cases

These are the examples of segmentation failures for breastlump The failures were due to a removal of a part of thelumps (Figure 6) It is more difficult to segment breastlumps due to the large variations in the intensities andvague boundaries such as Figure 6(a) Figure 6(b) shows thesegmentation of breast lump For this case different from theprevious a large wrong result of Live Wire occurs makinglumps identification more difficult For this breast lump in


(a) (b) (c) (d)

Figure 4 The comparison between the traditional Live Wire and the improved Live Wire

Figure 5 The segmentation results of the improved Live Wire algorithm


(a) (b) (c)

Figure 6 Examples of segmentation failures for breast lump

Figure 6(c) a part of the lumps was missed in the finalsegmentation Through Figure 6 we can obtain the reasonwhich led to failure in the detection of this nonhomogeneitylumps Their inhomogeneous density distribution made ithard to obtain an exact edge as possible

43 Assessment of Segmentation Performance In order toquantitatively evaluate the accuracy of the segmentationresults we have carried out several sets of experimentsto compare the computer automatic segmentation resultswith the standard lesion edge manually outlined by clinicalphysicians at the pixel level According to the clinical testingcriterion the accuracy in terms of true positive ratio (TP)false positive ratio (FP) false negative ratio (FN) andmisclassification error (ME) was used as quantitative indiceswhich are defined as follows

TP =Area 119878

119860cap 119878119898

Area 119878119861

FP =Area 119878

119860cup 119878119861minus 119878119861

Area 119878119861

FN =Area 119878

119860cup 119878119861minus 119878119860

Area 119878119861

ME =Area 119878

119860cup 119878119861 minus Area 119878

119860cap 119878119861

Area 119878119860cup 119878119861

(13)

where 119878119860is the area of computer automatic segmentation and

119878119861is the area of manually describeWhen TP is high it indicates that the segmentation

result will be less missed lesions with high-value of clinicaldiagnosisWhen FP is high it indicates that the segmentationresults diverge outwardly some of the lesions classified asnormal tissue region the area of segmentation results islarger than the area of actual lesion This causes physicianto assess disease excessively excise part of normal tissue inthe operation and give patients greater body injury WhenFN is high it indicates that the segmentation results inexcessive convergence which causes the physician to ignore

Table 1 Comparison between our algorithms and other algorithmsin quantitative indices

Method TP () FP () FN () ME ()Our algorithm 9664 197 407 591Level set 9276 3517 761 2950Graph cut 8965 365 1086 1395Random walk 7746 133 2348 2456Snake 9485 3686 622 2695

some lesions and thus affects clinical diagnosis The smallerthe ME value the more ideal the segmentation result Thesegmentation results are more consistent with the actuallesion WhenME = 0 the computer automatic segmentationresults are fully consistent with the manual segmentationresults

The comparison between our algorithm and other algo-rithms in quantitative indices is represented in Table 1 Itcan be clearly seen that the average TP (9664) of ouralgorithm exceeded that of other algorithms and the averageFN (407) of our algorithm is much less than that of otheralgorithms It manifests that our algorithm can recognizethe lesion accurately and misses lesions rarely However asfar as the average FP is concerned it can be seen that ouralgorithm is slightly worse than the random walk algorithmbut the average FN (407) of our algorithm is far lessthan that (2348) of the random walk algorithm Theresults of experiment show that the random walk algorithmis not sensitive to the weak edges surrounding the breastlumpsThe segmentation results show excessive convergencewhich results in ignoring some lesions and missing somepathological tissue during the procedure of operation

Data in Table 2 demonstrates the comparison betweenour algorithm and the classical Live Wire algorithm inquantitative indices It can be seen that our algorithm ismore satisfactory than the classical Live Wire algorithm inall quantitative indices From the observation of averageME our algorithm is superior to the classical Live Wirealgorithm which indicates that the segmentation result is


Table 2 Comparison between our algorithm and the classical LiveWire algorithm

Method TP () FP () FN () ME ()Our algorithm 9664 197 407 591Live Wire 9207 455 865 1242

more consistent with the actual lesion and can provideaccurate reference for clinical physicians

5 Conclusions

The edge information of breast mass lumps which wasobtained from computer automatic segmentation methodprovides important data for the diagnosis of breast cancerOwing to edge of breast lumps which is very fuzzy muchnoise and low contrast the classical Live Wire algorithm andsome other widely used algorithms are not able to accuratelyidentify the edge of breast lumps This paper presents animproved Live Wire interactive segmentation method whichuses Gabor-FCM as the cost function through the interactivesegmentation to extract the edge of breast lumps Basedon a lot of experiments we can arrive at the conclusionthat our algorithm can accurately extract the boundary ofthe lesions as well as its capability of location detectionand segmentation has more significant improvement thanthe classical Live Wire algorithm Our algorithm sufficientlysatisfies the requirements of clinical diagnosis and comes upto the auxiliary diagnostic standardMoreover our algorithmprovides accurate and reliable data for the physicians Sothe method used by this paper will be widely applied to theclinical diagnosis

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (Project no 81000639) and the ChinaPostdoctoral Science Foundation (Project no 20100470791and Project no 201104307) and the Tianjin Medical Uni-versityrsquos Undergraduate Research Opportunities Program(TMUUROP)

References

[1] R Siegel DNaishadham andA Jemal ldquoCancer statistics 2013rdquoCA Cancer Journal for Clinicians vol 63 no 1 pp 11ndash30 2013

[2] A Paliwal S Bisen M R Ghalib and N Sharmila ldquoAnimproved segmentation technique based on Delaunay triangu-lations for breast infiltrationtumor detection from mammo-gramsrdquo International Journal of Engineering andTechnology vol5 no 3 pp 2565ndash2574 2013

[3] B Kurt V V Nabiyev and K Turhan ldquoAutomatic microcalcifi-cation segmentation using rough entropy and fuzzy approachrdquo

in Information Technology in Bio-and Medical Informatics pp103ndash105 Springer Berlin Germany 2013

[4] F Marungo and P Taylor ldquoStorage and breast region segmen-tation for a non-distributed approach to clinical scale content-based image retrieval in mammographyrdquo in Medical Imaging2013 Advanced PACS-Based Imaging Informatics andTherapeu-tic Applications vol 8674 of Proceedings of SPIE InternationalSociety for Optics and Photonics February 2013

[5] P Rahmati A Adler and G Hamarneh ldquoMammography seg-mentation with maximum likelihood active contoursrdquoMedicalImage Analysis vol 16 no 6 pp 1167ndash1186 2012

[6] K Yuvaraj and U S Ragupathy ldquoComputer aided segmentationand classification of mass in mammographic images usingANFISrdquo The European Journal for Biomedical Informatics vol9 no 2 pp 37ndash41 2013

[7] I Ullah M Hussain and H A Aboalsamh ldquoSegmentation ofmammogram using gamma and dyadic wavelet transformsrdquohttpfacultyksuedusamhussainCADProjectJP6 Segmen-tationIhsan2012pdf

[8] D Sheet A Karamalis S Kraft et al ldquoRandom forest learningof ultrasonic statistical physics and object spaces for lesiondetection in 2D sonomammographyrdquo inMedical Imaging 2013Ultrasonic Imaging Tomography and Therapy Proceedings ofSPIE International Society for Optics and Photonics February2013

[9] I Tunali and E Kilic ldquoMass segmantation on mammogramsusing active contoursrdquo inProceedings of the 21st Signal Processingand Communications Applications Conference (SIU rsquo13) pp 1ndash4April 2013

[10] E Song L Jiang R Jin L Zhang Y Yuan and Q Li ldquoBreastmass segmentation in mammography using plane fitting anddynamic programmingrdquo Academic Radiology vol 16 no 7 pp826ndash835 2009

[11] H C Kuo M L Giger I Reiser K Drukker A Edwards andC A Sennett ldquoAutomatic 3D lesion segmentation on breastultrasound imagesrdquo in Medical Imaging 2013 Computer-AidedDiagnosis Proceedings of SPIE International Society for Opticsand Photonics February 2013

[12] F R Cordeiro W P Santos and A G Silva-Filhoa ldquoSegmenta-tion ofmammography by applying growcut formass detectionrdquoin Proceedings of the 14thWorld Congress onMedical and Health(Medinfo rsquo13) vol 192 of 87 IOS Press 2013

[13] G LiThe research and application for the interactive segmenta-tion algorithm [PhD thesis] Northeastern University 2009

[14] E N Mortensen and W A Barrett ldquoInteractive segmentationwith intelligent scissorsrdquo Graphical Models and Image Process-ing vol 60 no 5 pp 349ndash384 1998

[15] E N Mortensen and W A Barrett ldquoIntelligent scissors forimage compositionrdquo in Proceedings of the 22nd Annual ACMConference on Computer Graphics and Interactive Techniquespp 191ndash198 August 1995

[16] OGerard TDeschampsMGreff and LD Cohen ldquoReal-timeinteractive path extraction with on-the-fly adaptation of theexternal forcesrdquo in Proceedings of the 7th European Conferenceon Computer Vision (ECCV rsquo02) Copenhagen Denmark

[17] J-Z Yang S-Y Zhao YHu J-H Liu andX-H Xu ldquoAlgorithmof segmentation of medical series image 3D reconstructsrdquoJournal of SystemSimulation vol 17 no 12 pp 2896ndash2900 2005

Submit your manuscripts athttpwwwhindawicom

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


segmentation of breast lesions in 2D ultrasound imagesThey employ random forests for learning of tissue specificprimal to discriminate breast lesions from surroundingnormal tissues This enables it to detect lesions of multipleshapes and sizes as well as discriminate between hypoechoiclesion from associated posterior acoustic shadowing [8]Tunali and Kilic presented a method for segmentation ofbenign and malignant masses found onmammograms Afteradding low valued pixels outside of mammogram Chan-Vese active contour algorithm with new stopping criteria wasimplemented [9] Song et al developed a new segmentationmethod by use of plane fitting and dynamic programmingwhich achieved a relatively high performance level Thenew segmentation method would be useful for improvingthe accuracy of computerized detection and classification ofbreast cancer in mammography [10] Kuo et al evaluated a3D lesion segmentation method which they had previouslydeveloped for breast CT and investigated its robustness onlesions on 3D breast ultrasound images [11] Cordeiro et alpresented the GrowCut technique to segment tumor regionsof digitizedmammograms available in theminimias database[12]

There are three kinds of interactive segmentation meth-ods [13] One is based on closed initial contour segmenta-tion method active contour model and level set methodsare widely used in computer vision they have the strongcapability of noise immunity but they are sensitive to theimage contrast Therefore these methods are still not perfectfor Mammography X-ray imaging Another is based on theinitial target points and background points randomwalk andgraph cut methods are the typical representative of this kindof medical segmentation methodsThe results are affected bythe location and the number of the initial target points andbackground points the result has significant influence on theuser experience and the repeatability is lowThe third is basedon the marked points on the edge Live Wire is one of theimportant representatives it has been widely implementedin medical image segmentation Live Wire method has theless manual intervention and high-precision for the blurboundary of Mammography X-ray imaging

This paper presents a semiautomatic segmentationmethod based on Gabor FCM algorithm for the costfunction in Live Wire algorithm In recent years theGabor FCM as a mathematical tool has been widely usedin image segmentation We present the Gabor filters andFCM clustering algorithm is calculated to the Live Wirecost function definition and according to the image FCManalysis for image edge enhancement to eliminate theinterference of weak edge access external features clearsegmentation results of breast lumps The procedure of theproposed algorithm is shown in Figure 1

2 Live Wire Algorithm

Live Wire is a famous classic interactive algorithm firstproposed by Mortensen and Barrett [14ndash16] The algorithmproduces reasonable cost function to construct the optimalpath between the starting point and ending point provided bythe user so as to achieve the purpose of edge segmentation

Achieve Mammography X-ray imaging

Obtain the ROI

Preprocessing (Gabor-FCM cost function)

Segmentation of breast lump by Live Wire

The results

Figure 1 The procedure flow of the proposed algorithm

The optimal path from each pixel is determined atinteractive speeds by computing an optimal spanning treeof the image using an efficient implementation of Dijkstrarsquosgraph searching algorithmThe basic idea is to formulate theimage as a weighted 356 graph where pixels represent nodeswith directed weighted edges connecting each pixel with its8 adjacent neighbors

The key issue of Live Wire algorithms is the constructionof the cost function Theoretically the cost function shouldcontain all the information about the lesion edges includinggradient magnitude function gradient direction functionand Laplace zero crossing point the cost function of classicalLive Wire algorithms is defined as follows

119862(119901 119902) = 119908119885119891119885(119902) + 119908

119866119891119866(119902) + 119908

119863119891119863(119901 119902) (1)

where each 119908 is the weight of the corresponding featurefunction Note that empirically weights of 119882

119885= 0 43

119882119866

= 0 43 119882119863

= 0 14 119891119885represents formulation

of Laplace zero crossing point 119891119866represents formulation of

gradient magnitude function and 119891119863represents formulation

of gradient direction functionLaplace zero crossing point (119891

119885) is a binarization Eigen

function which is used to localize edge The Laplacian imagezero-crossing corresponds to points of maximal gradientmagnitude Thus Laplacian zero-crossing represents ldquogoodrdquoedge properties and should have a low local cost If 119871(119902) is theLaplacian of an image at the pixel 119902 then

119891119885=

0 119871 (119902) = 0

1 119871 (119902) = 0(2)

where 119871(119902) is the Laplacian of the original image at a pixel 119902If a pixel is on a zero-crossing then the component cost forall links to that pixel is low otherwise it is high Howevergradient magnitude provides a direct correlation betweenedge strength and local cost Multiscale gradient magnitude(119891119866) is as follows

119891119866(119902) = 1 minus

119866 (119902)

max (119866) (3)








119891119863(119901 119902) =

2

3120587cos [119889

119901(119901 119902)]

minus1

+ cos [119889119902(119901 119902)]

minus1

(4)

where

119889119901(119901 119902) = 119863

1015840(119901) sdot 119871 (119901 119902)

119889119902(119901 119902) = 119871 (119901 119902) sdot 119863

1015840(119902)

(5)


1198631015840(119901) = (119868

119910(119901) minus119868

119909(119901)) (6)



119871 (119901 119902) = 119902 minus 119901 119863

1015840(119901) sdot (119902 minus 119901) ge 0


(7)






119866 = exp(minus1199062+ 120574V2

21205902) sdot cos (2120587119891119906 + 120593)

119906 = (119909119895minus 119909119894) sdot cos 120579 + (119910

119895minus 119910119894) sdot sin 120579


119895minus 119910119894) sdot cos 120579

(8)




119891119867= 025 +

2119894minus05

119873025 le 119891

119867lt 05

119891119871= 025 minus

2119894minus05

1198730 lt 119891 lt 025

119894 = 1 2 log(1198738)2

(9)

where 119891119867





119869FCM =

119888

sum

119894=1

119899

sum

119896=1

120583119898

119894119896

1003817100381710038171003817119909119896 minus V119894

1003817100381710038171003817

2 (10)

where V119894isin 119881 119881 = (V

1 V2 V


of c 119880 = 120583119894119896 isin 119877

119888times119899 (120583119894119896

isin [0 1] sum119888

119894=1120583119894119896




120583119894119896= (

119888

sum

119895=1

(

1003817100381710038171003817V119894 minus 119909119896

100381710038171003817100381710038171003817100381710038171003817V119895minus 119909119896

10038171003817100381710038171003817

)

2(119898minus1)

)

minus1

V119894=sum119899

119896=1120583119898

119894119896119909119896

sum119899

119896=1120583119898

119894119896

(11)


119905minus1

1003816100381610038161003816 lt 120576 (12)

where 119869119905and 119869119905minus1

















(a) (b) (c) (d) (e)





(a) (b) (c) (d) (e)







(a) (b) (c) (d)




(a) (b) (c)




TP =Area 119878

119860cap 119878119898

Area 119878119861

FP =Area 119878

119860cup 119878119861minus 119878119861

Area 119878119861

FN =Area 119878

119860cup 119878119861minus 119878119860

Area 119878119861

ME =Area 119878

119860cup 119878119861 minus Area 119878

119860cap 119878119861

Area 119878119860cup 119878119861

(13)













5 Conclusions




Acknowledgments


References
























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of























119891119863(119901 119902) =

2

3120587cos [119889

119901(119901 119902)]

minus1

+ cos [119889119902(119901 119902)]

minus1

(4)

where

119889119901(119901 119902) = 119863

1015840(119901) sdot 119871 (119901 119902)

119889119902(119901 119902) = 119871 (119901 119902) sdot 119863

1015840(119902)

(5)


1198631015840(119901) = (119868

119910(119901) minus119868

119909(119901)) (6)



119871 (119901 119902) = 119902 minus 119901 119863

1015840(119901) sdot (119902 minus 119901) ge 0


(7)






119866 = exp(minus1199062+ 120574V2

21205902) sdot cos (2120587119891119906 + 120593)

119906 = (119909119895minus 119909119894) sdot cos 120579 + (119910

119895minus 119910119894) sdot sin 120579


119895minus 119910119894) sdot cos 120579

(8)




119891119867= 025 +

2119894minus05

119873025 le 119891

119867lt 05

119891119871= 025 minus

2119894minus05

1198730 lt 119891 lt 025

119894 = 1 2 log(1198738)2

(9)

where 119891119867





119869FCM =

119888

sum

119894=1

119899

sum

119896=1

120583119898

119894119896

1003817100381710038171003817119909119896 minus V119894

1003817100381710038171003817

2 (10)

where V119894isin 119881 119881 = (V

1 V2 V


of c 119880 = 120583119894119896 isin 119877

119888times119899 (120583119894119896

isin [0 1] sum119888

119894=1120583119894119896




120583119894119896= (

119888

sum

119895=1

(

1003817100381710038171003817V119894 minus 119909119896

100381710038171003817100381710038171003817100381710038171003817V119895minus 119909119896

10038171003817100381710038171003817

)

2(119898minus1)

)

minus1

V119894=sum119899

119896=1120583119898

119894119896119909119896

sum119899

119896=1120583119898

119894119896

(11)


119905minus1

1003816100381610038161003816 lt 120576 (12)

where 119869119905and 119869119905minus1

















(a) (b) (c) (d) (e)





(a) (b) (c) (d) (e)







(a) (b) (c) (d)




(a) (b) (c)




TP =Area 119878

119860cap 119878119898

Area 119878119861

FP =Area 119878

119860cup 119878119861minus 119878119861

Area 119878119861

FN =Area 119878

119860cup 119878119861minus 119878119860

Area 119878119861

ME =Area 119878

119860cup 119878119861 minus Area 119878

119860cap 119878119861

Area 119878119860cup 119878119861

(13)













5 Conclusions




Acknowledgments


References
























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















120583119894119896= (

119888

sum

119895=1

(

1003817100381710038171003817V119894 minus 119909119896

100381710038171003817100381710038171003817100381710038171003817V119895minus 119909119896

10038171003817100381710038171003817

)

2(119898minus1)

)

minus1

V119894=sum119899

119896=1120583119898

119894119896119909119896

sum119899

119896=1120583119898

119894119896

(11)


119905minus1

1003816100381610038161003816 lt 120576 (12)

where 119869119905and 119869119905minus1

















(a) (b) (c) (d) (e)





(a) (b) (c) (d) (e)







(a) (b) (c) (d)




(a) (b) (c)




TP =Area 119878

119860cap 119878119898

Area 119878119861

FP =Area 119878

119860cup 119878119861minus 119878119861

Area 119878119861

FN =Area 119878

119860cup 119878119861minus 119878119860

Area 119878119861

ME =Area 119878

119860cup 119878119861 minus Area 119878

119860cap 119878119861

Area 119878119860cup 119878119861

(13)













5 Conclusions




Acknowledgments


References
























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c) (d) (e)





(a) (b) (c) (d) (e)







(a) (b) (c) (d)




(a) (b) (c)




TP =Area 119878

119860cap 119878119898

Area 119878119861

FP =Area 119878

119860cup 119878119861minus 119878119861

Area 119878119861

FN =Area 119878

119860cup 119878119861minus 119878119860

Area 119878119861

ME =Area 119878

119860cup 119878119861 minus Area 119878

119860cap 119878119861

Area 119878119860cup 119878119861

(13)













5 Conclusions




Acknowledgments


References
























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c) (d) (e)







(a) (b) (c) (d)




(a) (b) (c)




TP =Area 119878

119860cap 119878119898

Area 119878119861

FP =Area 119878

119860cup 119878119861minus 119878119861

Area 119878119861

FN =Area 119878

119860cup 119878119861minus 119878119860

Area 119878119861

ME =Area 119878

119860cup 119878119861 minus Area 119878

119860cap 119878119861

Area 119878119860cup 119878119861

(13)













5 Conclusions




Acknowledgments


References
























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c) (d)




(a) (b) (c)




TP =Area 119878

119860cap 119878119898

Area 119878119861

FP =Area 119878

119860cup 119878119861minus 119878119861

Area 119878119861

FN =Area 119878

119860cup 119878119861minus 119878119860

Area 119878119861

ME =Area 119878

119860cup 119878119861 minus Area 119878

119860cap 119878119861

Area 119878119860cup 119878119861

(13)













5 Conclusions




Acknowledgments


References
























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c)




TP =Area 119878

119860cap 119878119898

Area 119878119861

FP =Area 119878

119860cup 119878119861minus 119878119861

Area 119878119861

FN =Area 119878

119860cup 119878119861minus 119878119860

Area 119878119861

ME =Area 119878

119860cup 119878119861 minus Area 119878

119860cap 119878119861

Area 119878119860cup 119878119861

(13)













5 Conclusions




Acknowledgments


References
























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of




















5 Conclusions




Acknowledgments


References
























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of





















of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















research article an interactive method based on...

Documents