Segmentingwholeheartfromcardiaccomputedtomography(CTimagescanprovideanimportantbasisfortheevaluationofcardiacfunctionandhelpimprovetheaccuracyofclinicaldiagnosis.Manualsegmentationisthemostaccuratemethodforcardiacsegmentation.Butitistimeconsumingandnotsufficientlyreproducible.However,cliniciansstillrelyonthismethodinpracticalapplications.Soafullyautomaticmethodisneededtoimprovethesegmentationefficiency.Thispapeproposesaregistration-basedautomaticapproachforthree-dimensional(3DsegmentationofcardiacCTimages.TheproposedmethodutilizesthesimilarityocardiacCTimagesbetweendifferentinpiduals,andusesregistrationtoachievethesegmentation.Affinetransformationisfirstlyimplementedtoachieveglobalcoarseregistration.Then,cubicB-splinesareusedtorefinethelocaldetailsinlocallyaccurateregistration.Mutualinformation(Ml)isusedasthesimilaritymeasure,andadaptivestochasticgradientdescent(ASGD)astheoptimizationalgorithm.Oumethodisappliedtothedual-sourcecardiacCTimagestosegmentwholeheartExperimentalresultsshowthattheproposedmethodcanautomaticallysegmentwholeheartfromcardiacCTimages.
