TransientBiometricsusingFingerNailsIgorBarrosBarbosaTheoharisTheoharis - PDF document

TransientBiometricsusingFingerNailsIgorBarrosBarbosaTheoharisTheoharisChristianSchellewaldDepartmentofComputerandInformationScienceNorwegianUniversityofScienceandTechnologyChamAthwalSchoolofDigitalMediaTechnologyBirminghamCityUniversityAbstractTransientbiometrics,anewconceptforbiometricrecog-nition,isintroducedinthispaper.Atraditionalperspectiveofbiometricrecognitionsystemsconcentratesonbiometriccharacteristicsthatareasconstantaspossible(suchastheeyeretina),givingaccuracyovertimebutatthesametimeresultinginresistancetotheirusefornon-criticalappli-cationsduetothepossibilityofmisuse.Incontrast,tran-sientbiometricsisbasedonbiometriccharacteristicsthatdochangeovertimeaimingatincreasedacceptanceinnon-criticalapplications.Weshowthatthengernailisatran-sientbiometricwithalifetimeofapproximatelytwomonths.Ourevaluationdatasetsareavailabletotheresearchcom-munity.1.IntroductionBiometricrecognitionsystemsofferuniqueadvantagewhencomparedtoconventionalrecognitionsystems,suchassmartcardsorpasswords.Byusingabiometricrecogni-tionsystem,thesubjectdoesnotneedtocarryorrememberanyidorpassword,andthereislessriskoflossordis-closureoftherecognitiontoken.Biometricrecognitionisthusgainingsupportandacceptanceincriticalrecognitionsituationssupportedbygovernmentsorotherlargeorgani-zations.Despitetheadvantagesofbiometricrecognitionsystems,amajorconcernofindividualsisthepossibilityofmisuseoftheirbiometricdata.Acardorpasswordcanbecanceled,butwhathappensifyourbiometricdatafallsintothewronghands?Anindividual'sprivacymaybecompromised(e.g.throughtheiruseforunauthorizedrecognitionpurposes)ordiscriminationmaybeenabled(e.g.throughunauthorizedusebyinsuranceagents).Cancelablebiometrics[7,8]attemptstoanswerthiscon-cernbypre-transforming(distorting)thebiometricdatabe-forethebiometricsignatureisextracted.Thetransforma-tionisnon-reversible.Thus,thepotentialformisuseislim-itedbythefactthatthemisusercannotretrievetheoriginalbiometricdata,andthetransformationcanbechangedatanytime.However,cancelablebiometricsrequiresthatthesubjecttruststhebiometricscapturepointandalsothatthemisuseisdetectedinordertoactivateatransformchange.Thereisplentyofscopeforbiometricrecognitionsys-temstobecomemoresociallyacceptable,inthesensethatsocietycouldacceptandusesuchsystemsinday-to-dayscenarios.Theacceptabilityissueremainsparticularlyopenwhendealingwithnon-criticalscenariosandcollaborativesubjects.Forinstance,individualswillnothappilyoffertheirngerprintsjusttohaveaccesstotheirhotelroom.Thepointsraisedabovelimitstheuseofbiometrictechnologiesinamultitudeofnoncriticalsituations.Inthispaperweintroducetransientbiometrics.Tran-sientbiometricsisdenedasbiometricrecognitiontech-nologieswhichrelyonbiometriccharacteristicsthatareproventochangeovertime.Thus,theyautomaticallycan-celthemselvesoutafteraknownperiodoftime.AtransientbiometricapproachforthevericationtaskisshowninFig.1.Incontrasttocancelablebiometrics,itistheactualbio-metricdatathatarenaturallychangingovertime.Asacon-sequenceitwillpresumptivelyhelpinthecreationofmoresociableacceptablerecognitionssystems.Weshowthatim-agesofthengernailconstituteatransientbiometricwithalifetimeoftwomonths.Theremainingofthepaperisorganisedasfollows.Sec-tion2brieypresentsthebiometricliteraturewhichem-ploysngernails.Section3detailsourapproachfollowedbySection4thatshowsexperimentalresults.Finally,Sec-tion5concludesthepaper,envisagingsomefutureperspec-tives. Subjects:(B)(C)(D) (A) TODAY DB Biometric feature extraction Enrollment (A) MATCH? After 1 week YES (A) MATCH? After 2 months NO Matching Procedure Figure1.Exampleofavericationtaskemployingtransientbio-metrics.2.PreviousWorkTheuseofngernailsinbiometricsapplicationshasbeenthetopicofafewdifferentlinesofresearch.Acom-plexacquisitionsystememployingtailoredlightingequip-menthasbeendesignedtoacquireimagesofthenailbed,whichistheskinunderthenailplate[9].Suchimagesarethenusedforindividualauthenticationbyexploringfea-turesfromthenailbedgrooves.Thisispossiblebecausethenailbedisuniquetoeachindividual[3].Recently,acancelablebiometricapproachhasdevel-opedstickers,whichcanbeplacedoverngernailsforanidenticationprocess[4].Inthepresentedpaper,thumbim-ageswereacquiredagainstablackbackground,foritmadeiteasiertocomputetheboundariesofthethumb.Thestick-ersgluedoverthengernailsprovidetwolandmarkswhichareusedinthefeatureextractionprocess.Thengeroutlineispursued,andthedistancefromtheoutlinetotheland-markscreatesadistanceprole.Thenalmatchingproce-dureisdonebycomputingcorrelationcoefcientsbetweendistanceproles.Wheneversuchstickerisremovedorre-positionedoverthengernail,thedistanceprolechanges.Therefore,itispossibletocancelthebiometricbyreplacingthelandmarksorbyshiftingthestickerposition.Thengernailwouldonlyplayaroleintherecognitionsystemifthenailoutgrewthenger.Althoughitisavalidcancelablebiometricapproach,thisworkdoesnottrulyexplorengernailsforbiometricrecognition.Theinformationofthenailsurfacehasbeenexploredbyabiometricauthenticationsystemproposedin[2].Thisap-proachusesimagesofhandsinordertoextractinformationfromthengernails.First,thengersaresegmentedbyasmartcontoursegmentationalgorithm,thenthenailsaresegmentedbygreyscalethresholding.Thissimplesegmen-tationapproachislikelytoworkgiventhattheemployeddatasetwasbiasedwithrespecttosubject'sskintones.TheindividualauthenticationprocessisbuiltuponthehammingdistanceofhighfrequencyHaarwaveletcoefcients.Ex-perimentalresultsshowreasonablerecognitionratesusingthreesampleimagespersubject.Thersttwoimagesareemployedintrainingwhilethelastimageisusedfortest-ing.Despitethepositiveresults,thisworkdoesnotexplorehowtherecognitionratebehaveswithrespecttothegrowthofthengernails;theauthorsdonotprovidethetimedif-ferencebetweenacquisitions,soweassumethatallimageswereacquiredonthesamedate.3.TheProposedApproachThetransientbiometricapproachpresentedinthisworkaddressestheidenticationproblem.Therefore,ourobjec-tiveistoidentifyasubjectbycomparingabiometricsig-natureagainstadatasetofpreviouslycollectedsamples.Tothisend,theproposedsolutioncanbedividedintothreephases.Therstphasedealswithimagesegmentationandpre-processing.Thesecondphaseextractsthebiometricsignature,whilethethirdphaseaddressessignaturematch-ing.3.1.Nailimagepre­processingImagesoftherightindexngerarethesourceofbio-metricinformation.Sincetheimagesweretakenondif-ferentdaysandsometimeswithdifferentcameras,thepre-processingofinputimagesisakeystepfortheoverallpro-cess.Itassuresthattheimagesdeliveredtothesignatureextractionalgorithmfulltherequisitesregardingcolourcorrection,nailplateregistrationandimagesize.Pre-processingstartsbysegmentingthenailfromthen-gerimages.Suchsegmentationisdonebyanactiveshapemodel(ASM)see[11].Theactiveshapemodelrequiresasetoftrainingimageswherethesegmentationhasbeenmanuallyperformed(contourdrawn).Thealgorithmem-ploysPrincipalComponentAnalysis(PCA)tondeigensegmentationcontours,withveryaccurateresults.TheASMalsodescribestheimagearoundeachcontrolpointwithagrey-levelappearancemodel.Thisgrey-levelappearancemodeliscomputedusinglinesperpendiculartoeachcon-trolpoint,anditisbuiltusingtherstderivativeofgrey-levelimages.Thisappearancemodelwillbeusedlaterinaniterativefashiontocorrectthepositionofcontrolpointswhilesearchingforthebestsegmentationcontour.TheASMrequirestrainingdataandforthispurposethedatasetD01wasused.Thistrainingdatasetrepresentstherstacquisitionday.Itcontainsanimageoftherightindexngerforeveryenrolledindividual,withatotalof32im-ages(seemoreinformationonthedatasetinSec.4).TheASMwastrainedwithtwomainlandmarksand20controlpointsbetweenthem.Therstlandmarkisplacedatthebaseofthenailplatejustbytheintersectionwiththenger skin.Meanwhile,theotherlandmarkisplacedoppositetoit,bytheendofthenailplate.Aninputsampleisshownas[A]inFig.2,andtheresultingsegmentationisshownas[B].Imagepre-processingcontinuesbycomputingtheboundingbox.Next,theboundingboxisconvertedtogreyscale,makingtheinputmorerobusttochanges.Thesechangesarelikelytohappenduetowrongwhitebalanceorevenduetotheuseofdifferentcameras.Theoverallpre-processedimageisgivenbyresizingtheboundedboxtoawidthandheightof128pixels.Theresultingimageisshownas[C]inFig.2 [A] Subjects {1} ASM segmentation.[B] Segmented {2} Image Registration and color convertion. [C] Gray scale Initial LandmarkSecond Landmark Figure2.Sampleresultsfromtheimagepre-processingpipeline.Tomakethetrainingprocessmorerobusteachimageisusedtocreatemultiplevariations.ThesearegiventhroughtheapplicationofWienerFilters,byshiftingthesegmentedregion-of-interestbyafewpixelsandbytheapplicationofhistogramequalisation.Whenalltheseimagesmodicationarecombinedinachain,everyinputimagegenerates810variations.3.2.SignatureextractionbasedonuniformLBPIthasbeenobservedthatanailplateiscategorisedbyauniquetexturewhichisinuencedbypatternsinthenailbed[3].Thenailplatetextureisalsodependentoninter-actionwithexternalfactors.Hence,itiscommontono-ticewhitespotsandmarksoriginatingfromscratchesorbumps.Sincethenailplatepossessessuchrichtexture,wehaveoptedtobasesignatureextractiononLocalBinaryPatterns(LBP),whichisasuccessfulandrobusttexturede-scriptor[5].LBPsareknownfortheircomputationalef-ciencyandtheircapacitytodiscriminatemicro-patterns.Theyhavealsobeensuccessfullyemployedinawideva-rietyofapplications,rangingfromtextureclassication[6](theiroriginalpurpose),tofacialrecognition[1].Thus,LBPhasbeenselectedforthesignatureextractionprocess.LBParecomputedpixelwise,relyingonthepixelneigh-bourhoodinformation.ThecomputationstartsbydeninganeighbouringcirclewitharadiusofRpixelsandPevenlyspacedsamplepoints.Bilinearinterpolationisusedtocom-putethevalueofasamplepointifitdoesnotfallonapixelcenter.Fig.3illustratestwopossiblecircularneighbour-hoods.LBPiscomputedforthepixelgc,locatedinthecenterofthecircle,usingthethresholdoperationofEq.1. gc g2 g3 g4 g5 g6 g7 g8 g1 gc g1 g2 g3 g4 g8 g7 g6g5 Figure3.Sampleneighbourhoodsof(P;R)=(8;1)and(P;R)=(8;2).Intheseexamplesgparethesamplepoints,whereprangesfrom1toP.LPBP;R=PXp=1(gpgc)2p1(x)=1ifx00ifx0(1)AnLBPisuniformwheneverthecodedvalueiscom-posedofzero,oneortwobit-wisetransitions.Thus,thepat-terns11111111and00001111areuniformsincetheyhavezeroandtwobit-wisetransitions,respectively.Ontheotherhand,thepattern10101010isnon-uniformsinceitiscom-posedofeighttransitions.Ifacodedpixeluseseightsamplepoints,itispossibletogenerate256patterns,outofwhich58arecalleduniformLBPpatterns.Theworkof[1,6]con-rmsthatuniformLocalBinaryPatterns(LBP2)accountforthevastmajorityofencounteredpatterns.Therefore,signatureextractionemploysuniformLBPfordescribingthenailplatetexture.Thesignatureextractionprocessstartsbydividingthepre-processednailplateimageintosmallerimageblocks.A44gridisusedforthedivision,generating16blocksof3232pixels.AhistogramofthevaluesofLBP28;2isthencomputedforeachblock.Thehistogramiscomposedof59bins,58ofthemusedforuniformpatternsandthelastbinfornon-uniformones.Thesignatureisthencreatedbyconcatenatingthe16histograms,thusformingaglobalde-scriptorofthenailplate.ThisprocessisillustratedinFig.4andfollowsthemethodologyproposedin[1].Adescriptorcapableofdescribingtextureanditsspatialrelationshipsisthuscreated,whichisverysuitableforthenailplate.How-ever,theresultingsignaturehas944features.Thedimen-sionalitycurseisavoidedbytheuseofKarhunen-Loeve Analysis.ThedatafromD01(seeSec.4),isanalysedbyKarhunen-Loevedecompositionwhichallowsustomaptheextractedsignatureintoasubspaceof200dimensions. 59 59 Division in squareregions Region LBPextraction 944 01 Concatenation (x16)LBP based Signature Figure4.Signatureextractionpipeline3.3.SignaturematchingSignaturematchingessentiallyidentiespatternswithinadataset.Thesepatternsshouldideallyhavesmallvari-ationbetweenobjectsofthesameclass(e.g.nailimagesofthesamesubject)whiletheyshouldhavelargevari-ationacrossdifferentclasses.ToidentifysuchpatternsBayesianclassicationisemployed.ABayesianclassierestimatestheboundariesbetweenclassesassuringthattheBayesrisk/errorisminimal.TheBayesrulestatesthattheprobabilityofasubjectbelongingtoaclass!kgivenanobservationz(signatureafterdimensionalityreduction)isgivenbytheposteriorprobabilityasshowninEq.2.P(!kjz)=p(zj!k)P(!k) p(z):(2)whereP(!kjz)istheposteriorprobability,p(zj!k)istheprobabilitydistributionofzcomingfromasubjectwithknownclass!k,P(!k)isthepriorprobabilityofhavingtheclass!kandp(z)isthedistributionoftheobservation.Ifaunitarycostisassumedforeverywrongclassication,theminimisationoftheBayesriskbecomesequivalenttothemaximisationofposteriorprobability[10].Therefore,theclassiercanbere-writtenas^!map(z)=argmaxjfp(zj!j)P(!j)g:(3)Ourassumptionisthattheconditionalprobabilitydensityfunction,p(zj!j),canbemodelledasnormal.ThereforetheobservationsareassumedtohaveanexpectationvectorkandacovariancematrixCk,yieldingtothefunctionshowninEq.4p(zj!j)=1 p (2)NjCkje(zk)TC1k(zk) 2(4)Suchaconditionalprobabilitydensityfunctionresultsinaquadraticclassier.Itwasdeterminedexperimentallythatifweassumethatthecovariancedoesnotdependontheclass,e.g.Ck=Cforallpossibleclasses,weendupwithalinearBayesnormalclassierwhichoutperformesthequadraticclassier.ThelinearBayesnormalclassieristhemappliedtoev-erycomputedzandthenalclassicationisgivenbytheconjunctionofthe810imageswhichwerearticiallycre-ated.Togetthenalclassicationofeachinputimage,analdistancemeasureiscreatedbyEq.5.DZ810;!k=810Xn=1logcZ810n;!k(5)whereZ810representstheconjunctionof810observa-tionsderivedfromasingleinputimage,cZ810n;!krepre-sentsthecondenceofthenthobservationofZbeingfromclass!k.Whenthisdistancemeasureisemployed,themostprobableclassisgivenbythesmallestcomputeddistance.4.ExperimentsThissectionwilldescribetheexperimentaldatasetandtheeffectofnailplategrowthonidenticationperformance.4.1.DatasetCreationOurdatasetiscomposedofthreedifferentsetsofdata.Allthreesetshavefollowedthesameacquisitionprocess.Firsttherightindexngerofthesubjectisplacedoverawhitesheetofpaper,whichissupportedbyaatsurface.Thengerisplacedoverthissurfacewithoutputtingpres-sureagainstit,aspressurechangesthecolourofthengernail.Thenadiffuselightisplacedsothelightsourcepointstothetopofthenger.Therefore,thengerisvirtuallypointingtothelightsource.Suchlightingconditionavoidhighlightsandhelpachieveproperexposure.Finallythe imageisacquiredbyframingonlythecontentsofthewhitepaperandmaintainingproperfocusonthengernailplate.Twothingsdifferentiatethethreesetsofdata:acquisi-tiondateandnumberofsubjects.SetD01consistsofthecollectionofimagesfromtherstacquisitionday.Notsur-prisingly,thisrepresentsthelargestsetintermsofnumberofsubjects,consistingofdatafrom32individuals.Thesec-ondsetD08iscomposedofimagesacquiredsevendaysaftertheinitialacquisition.Thissetiscomposedof24indi-vidualswhowerealsopartofD01.Thethirdandsmall-estset,D70,containsimagesof17subjects,allpartofD08.Theimagesofthisthirdsetwhereacquiredseventydaysaftertheinitialacquisitionday.Figure5showssam-plesoffourindividualswhowererepresentedinallthreedatasets.ThenaildatasetanditsfutureextensionswillbemadeavailableatNTNU'sVisualComputinggroupwebsite[http://www.idi.ntnu.no/grupper/vis/]. D01 Figure5.Imagessamplesfromavailabledatasets.EachRowrep-resentsadatasetwhileeachcolumnrepresentsadifferentsubject.4.2.IdenticationperformanceanalysisInallexperimentstheclassierwastrainedusingonlyinformationfromD01.TheclassierwasthenappliedtoD08andD70toevaluatethedecayofidenticationperfor-mance,asexpectedforatransientbiometricsolution.AsD70containsonly17subjects,theclassierK17istrainedonD01usingonlytheinformationfromsubjectsavailableinD70.TheclassieristhenappliedtoD08andD70.Thecumulativematchingcurve(CMC)isusedasastandardisedevaluationgraph.Itassessestheclassicationperformanceinidenticationproblems.CMCmodelstheprobabilityofasignaturefromatestdataset,inthiscaseD08andD70,beingcorrectlymatchedintherstPrankedsubjectsfromthetrainingdatasetD01.Suchrankisderivedfromthecom-puteddistances,asspeciedinSection3.3.TheCMCcurveforbothD08andD70areplottedinFig.6.Theperfor-mancedecayobservedinD70isevidencethatthebiometricsignatureextractedfromthenailplatebiometricisofshortpersistence.Thusthenailplateisagoodcandidatefortran-sientbiometricsolutions.Ifthenormalisedareaunderthecurveistakenasaperformancemeasure,thechangesinthenailplateduringthe62daysbetweentheacquisitionofD70andD08accountfora9:32%decay.Ifrankonerecognitionistakenasameasureofperformance,theresultsareevenmoreconclusive:thetwomonthintervalrepresentsadecayof58:82%intheprobabilityofidentifyingtheindividualinarstguess. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 0 10 20 30 40 50 60 70 80 90 100 Rank[k]Recognition Rate [%]Classification Results for K17 D08 D70 Figure6.CumulativematchingcurvesforclassierK17evaluatedonD08andD70.Theclassierwastrainedfor17subjectsandthenappliedtoimagesacquired8dayslater(D08)and70dayslater(D70).ThedecayinperformancefromD08toD70argu-mentsinfavourthatngernailsaretransientbiometrics.There-fore,thebiometricinformationchangesduringtime,makingtheidenticationprocessunreliableaftertwomonths.AsecondclassierK24wastrainedonD01usingonlytheinformationfromsubjectsavailableinD08.Thisclassi-eriscomposedof24subjectsandrepresentsaharderclas-sicationtaskthantheoneK17wasassigned.Theobjectivewastoshowthatpositiveidenticationispossiblewithnailbiometricsoverashortperiodoftime.TheclassicationresultsareshownasaCMCcurveinFig.7.Finally,Table1summarisestheresultsofthethreeclas-sicationproblemspresented. ClassierK17 TestDataset nAUCRank1Rank2 D08 100:00%17/1717/17D70 90:657%7/1711/17 ClassierK24 D08 99:479%22/2423/24 Table1.Classicationperformance5.ConclusionSofar,biometricsresearchhasproducedsignicantre-sultsintermsofuniversality,distinctivenessandperma-nence.Acceptabilitystillremainsasanimportantissueandthemainreasonbehindthisisthefearofmisuseofone's 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 0 10 20 30 40 50 60 70 80 90 100 Rank[k]Recognition Rate [%]Classification Results for K24 D08 Figure7.CumulativematchingcurvesforclassierK24evaluatedonD08.Theclassierwastrainedfor24subjectsandthenappliedtoimagesacquired8dayslater(D08).TheoverallperformanceachievedinD08argumentsinfavourthatngernailsaretransientbiometricswhichaweekintervalhaslittleeffectintherecognitioncapabilities.permanentbiometricdata.Individualsarethusreluctanttovolunteertheirbiometriccharacteristicswherepossible,andtheleapinusabilitythatbiometrictechnologyoffers(i.e.password-anddevicefreeaccesstoresources),cannotberealised.Thisworkintroducesanewideatoaddresstheaccept-abilityissueinherenttobiometricsolutions.Thisapproach,designedforcollaborativeindividuals,insteadofrecordingpermanentdata,recordstransientdata,i.e.datathatdochangeovertimeandarethuscancelledbynature.Users,whoknowthatthebiometricdatatheyofferisgoingtobeuselessforrecognitionpurposesafteracertainamountoftime,arelikelytobemorewillingtoofferit,evenforday-to-dayapplications.Thisapproachistermedtransientbio-metrics;theideaistousefeatureswithashortpermanence,givingadiminutiveperiodofrecognition.Atransientbiometricsolutiontotheidenticationtaskwaspresented,whichexploitstexturefeatures,extractedfromngernailimages,investigatingdifferentacquisitionintervals.Identicationperformancewashighwithinaweekbutdegradedconsiderableafteratwomonthperiod.Thisindicatesthatngernailimagesareavalidtransientbiometricsolution.Insubsequentworkweintendtoexpandourdatasetwithmoresubjects,morerealisticcaptureconditionsanddiffer-entskintones.References[1]T.Ahonen,A.Hadid,andM.Pietik¨ainen.Facedescriptionwithlocalbinarypatterns:applicationtofacerecognition.IEEEtransactionsonpatternanalysisandmachineintelli-gence,28(12):2037–41,Dec.2006.[2]S.Garg,A.Kumar,andM.Hanmandlu.Biometricauthen-ticationusingngernailsurface.In201212thInternationalConferenceonIntelligentSystemsDesignandApplications(ISDA),pages497–502.IEEE,Nov.2012.[3]R.Krstic.HumanMicroscopicAnatomy:AnAtlasforStu-dentsofMedicineandBiology.Springer,1991.[4]N.NishiuchiandH.Soya.CancelableBiometricIdentica-tionbyCombiningBiologicalDatawithArtifacts.In2011InternationalConferenceonBiometricsandKanseiEngi-neering,numberii,pages61–64.IEEE,Sept.2011.[5]T.Ojala,M.Pietik¨ainen,andD.Harwood.Acomparativestudyoftexturemeasureswithclassicationbasedonfea-tureddistributions.PatternRecognition,29(1):51–59,Jan.1996.[6]T.Ojala,M.Pietikainen,andT.Maenpaa.Multiresolutiongray-scaleandrotationinvarianttextureclassicationwithlocalbinarypatterns.IEEETransactionsonPatternAnalysisandMachineIntelligence,24(7):971–987,July2002.[7]N.K.Ratha,J.H.Connell,andR.M.Bolle.Enhancingsecu-rityandprivacyinbiometrics-basedauthenticationsystems.IBMSystemsJournal,40(3):614–634,2001.[8]C.RathgebandA.Uhl.Asurveyonbiometriccryptosys-temsandcancelablebiometrics.EURASIPJournalonInfor-mationSecurity,2011(1):3,2011.[9]A.Topping,V.Kuperschmidt,andA.Gormley.UnitedStatesPatentUS005751835A,1998.[10]F.vanderHeijden,R.P.W.Duin,D.deRidder,andD.M.J.Tax.Classication,parameterestimationandstateestima-tion-anengineeringapproachusingMatlab.JohnWiley&Sons,Chichester,2004.[11]B.vanGinneken,A.F.Frangi,J.J.Staal,B.M.terHaarRomeny,andM.A.Viergever.Activeshapemodelsegmen-tationwithoptimalfeatures.IEEEtransactionsonmedicalimaging,21(8):924–33,Aug.2002.