ComputerizedMedicalImagingandGraphics35 (2011) 383–397
ContentslistsavailableatScienceDirect
ComputerizedMedicalImagingand

Graphics
journalhomepage:doc.guandang.net

g
Amodi edpossibilisticfuzzyc-meansclusteringalgorithmforbias eldestimationandsegmentationofbrainMRimage
Ze-XuanJi,Quan-SenSun ,De-ShenXia
TheSchoolofComputerScienceandTechnology,NanjingUniversityofScienceandTechnology,No.200,XiaoLingWeiStreet,Nanjing210094,China
articleinfoabstract
Amodi edpossibilisticfuzzyc-meansclusteringalgorithmispresentedforfuzzysegmentationofmag-neticresonance(MR)imagesthathavebeencorruptedbyintensityinhomogeneitiesandnoise.Byintroducinganoveladaptivemethodtocomputetheweightsoflocalspatialintheobjectivefunction,thenewadaptivefuzzyclusteringalgorithmiscapableofutilizinglocalcontextualinformationtoimposelocalspatialcontinuity,thusallowingthesuppressionofnoiseandhelpingtoresolveclassi cationambi-guity.Toestimatetheintensityinhomogeneity,theglobalintensityisintroducedintothecoherentlocalintensityclusteringalgorithmandtakesthelocalandglobalintensityinformationintoaccount.Theseg-mentationtargetthereforeisdrivenbytwoforcestosmooththederivedoptimalbias eldandimprovetheaccuracyofthesegmentationtask.Theproposedmethodhasbeensuccessfullyappliedto3T,7T,syntheticandrealMRimageswithdesirableresults.Comparisonswithotherapproachesdemonstratethesuperiorperformanceoftheproposedalgorithm.Moreover,theproposedalgorithmisrobusttoinitialization,therebyallowingfullyautomaticapplications.
© 2010 Elsevier Ltd. All rights reserved.
Articlehistory:
Received7July2010
Receivedinrevisedform27October2010Accepted9December2010
Keywords:
BrainMRimagesegmentationIntensityinhomogeneity
Possibilisticfuzzyc-meansalgorithmAnisotropicweight
Coherentlocalandglobalintensity
1.Introduction
Magneticresonance(MR)imaginghasseveraladvantagesoverothermedicalimagingmodalities,includinghighcontrastamongdifferentsofttissues,relativelyhighspatialresolutionacrosstheentire eldofviewandmulti-spectralcharacteristics.Therefore,ithasbeenwidelyusedinquantitativebrainimagingstudies.Quantitativevolumetricmeasurementandthree-dimensional(3D)visualizationofbraintissuesarehelpfulforpathologicalevolutionanalyses,whereimagesegmentationplaysanimportantrole.
Thesizealterationsinbraintissuesoftenaccompanyvariousdiseases,suchasschizophrenia[1].Thus,estimationoftissuesizeshasbecomeanextremelyimportantaspectoftreatmentwhichshouldbeaccomplishedaspreciselyaspossible.ThiscreatestheneedtoproperlysegmentthebrainMRimagesintograymatter(GM),whitematter(WM)andcerebrospinal uid(CSF)andalsotoidentifytumorsorlesions,ifpresent[2].
Themaindif cultiesinbrainsegmentationaretheintensityinhomogeneitiesandnoise.Infact,intensityinhomogeneityoccursinmanyreal-worldimagesfromdifferentmodalities[3,4].Inparticular,itisoftenseeninmedicalimages,suchasX-rayradio-graphy/tomographyandMRimages.Forexample,theintensityinhomogeneityinMRimagesoftenappearsasanintensityvaria-
Correspondingauthor.Tel.:+8613611507508.E-mailaddresses:jizexuan@hotmail.com(Z.-X.Ji),sunquansen@mail.njust.edu.cn(Q.-S.Sun).
0895-6111/$–seefrontmatter© 2010 Elsevier Ltd. All rights reserved.doi:10.1016/j.compmedimag.2010.12.001
tionacrosstheimage,whicharisesfromradio-frequency(RF)coilsoracquisitionsequences.Thustheresultantintensitiesofthesametissuevarywiththelocationsintheimage.ThenoiseinMRimagesisRiciandistributedandcanaffectsigni cantlytheperformancesofclassi cationmethods.Thebestsolutionsconsistofeither lteringtheimagepriortoclassi cationorembeddingspatialregularizationinsidetheclassi eritself.
Bias eldcorrectionreferstoaproceduretoestimatethebias eldfromthemeasuredimagesothatitseffectcanbeeliminated.MethodsforintensityinhomogeneitycorrectioninMRimageshavebeenreviewed[5,6].Currently,therearetwopopularsegmen-tationbasedmodelsforintensityinhomogeneitycorrection:theexpectation-maximization(EM)algorithm[7,8]andthefuzzyC-mean(FCM)algorithm[9–12].Inthesemethods,thetasksofbias eldcorrectionandsegmentationareinterleavedinaniterativeprocesssuchthattheybene tfromeachothertoyieldbetterresults.Wellsetal.[7]developedanapproachbasedonEMalgo-rithmforinterleavedbias eldcorrectionandimagesegmentation.However,therearetwomaindisadvantagesofthisEMapproach.First,theEMalgorithmisextremelycomputationallyintensive.Second,theEMalgorithmrequiresagoodinitialguessforeitherthebias eldorfortheclassi cationestimate.Otherwise,theEMalgorithmcouldbeeasilytrappedinalocalminimum,resultinginanunsatisfactorysolution.BasedontheEMframeworkin,Leem-putetal.[8]proposedanexplicitparametricmodelofthebias eld.Insteadofmanualintervention,theirmethodusedadig-italbrainatlasthatprovidesaprioriprobabilitymapsforWM,GM,andCSF.Althoughthismethodisclaimedtobemorerobust
Kernel based fuzzy and possibilistic c-means clustering_材料科学_工程科技_...R p , the original FCM algorithm partitions X into c fuzzy subsets by ...
Kernel based fuzzy and possibilistic c-means clustering_材料科学_工程科技_...R p , the original FCM algorithm partitions X into c fuzzy subsets by ...
热门文档