模式识别英文经典论文--Non-rigid imageregistrationofbrainmagneticresonanceimagesusing

模式识别英文经典论文

PatternRecognition44(2011)2450–2467

ContentslistsavailableatScienceDirect

PatternRecognition

journalhomepage:doc.guandang.net

Non-rigidimageregistrationofbrainmagneticresonanceimagesusinggraph-cuts

RonaldW.K.SoÃ,TommyW.H.Tang,AlbertC.S.Chung

LoKwee-SeongMedicalImageAnalysisLaboratory,DepartmentofComputerScienceandEngineering,TheHongKongUniversityofScienceandTechnology,ClearWaterBay,HongKong

articleinfo

Articlehistory:

Received12February2010Receivedinrevisedform6April2011

Accepted10April2011

Availableonline21April2011Keywords:

Non-rigidimageregistrationGraph-cuts

abstract

Wepresentagraph-cutsbasedmethodfornon-rigidmedicalimageregistrationonbrainmagneticresonanceimages.Inthispaper,thenon-rigidmedicalimageregistrationproblemisreformulatedasadiscretelabelingproblem.Basedonavoxel-to-voxelintensitysimilaritymeasure,eachvoxelinthesourceimageisassignedadisplacementlabel,whichrepresentsadisplacementvectorindicatingwhichpositioninthe oatingimageitisspatiallycorrespondingto.Intheproposedmethod,asmoothnessconstraintbasedonthe rstderivativeisusedtopenalizesharpchangesintheadjacentdisplacementlabelsacrossvoxels.Theimageregistrationproblemisthereforemodeledbytwoenergytermsbasedonintensitysimilarityandsmoothnessofthedisplacement eld.Theseenergytermsaresubmodularandcanbeoptimizedbyusingthegraph-cutsmethodviaa-expansions,whichisapowerfulcombinatorialoptimizationtoolandcapableofyieldingeitheraglobalminimumoralocalminimuminastrongsense.UsingtherealisticbrainphantomsobtainedfromtheSimulatedBrainDatabase,wecomparetheregistrationresultsoftheproposedmethodwithtwostate-of-the-artmedicalimageregistrationapproaches:free-formdeformationbasedmethodanddemonsmethod.Inaddition,theregistrationresultsarealsocomparedwiththatofthelinearprogrammingbasedimageregistrationmethod.Itisfoundthattheproposedmethodismorerobustagainstdifferentchallengingnon-rigidregistrationcaseswithconsistentlyhigherregistrationaccuracythanthosethreemethods,andgivesrealisticrecovereddeformation elds.

&2011ElsevierLtd.Allrightsreserved.

1.Introduction

Imageregistrationhasbeenactivelyandextensivelystudiedandappliedinthe eldofmedicalimageanalysis.Bythenatureoftheimagetransformation,imageregistrationcanbeclassi edintorigidimageregistrationandnon-rigidimageregistration.Rigidimagetransformationconsistsofatranslationfollowedbyarotationinthe3Dspace.Assuch,inarigidtransformation,thedegreeoffreedomisrelativelylowandthetechniquesforrigidimageregistrationarebecomingmature.Methodsbasedonthemaximizationofmutualinformation[1,2]andnormalizedmutualinformation[3]arewidelyusedinrigidmedicalimageregistra-tion.Arecentoverviewandmoreapplicationsofthemutual-information-basedmedicalimageregistrationmethodscanbefoundin[4].Incontrast,non-rigidimageregistrationisanill-posedproblemduetoitssupernormalhighdegreeoffreedomandinherentrequirementofsmoothnessinthedeformation eld.Thetopicisstillunderactiveresearchbecausetherearemanyusefulapplicationsincluding(1)constructionofatlasacrossa

ÃCorrespondingauthor.

populationoraspeci cpatientgroup[5],(2)recoveryofdeforma-tion eldduringsurgerywithrespecttoapre-treatmentimage[6],(3)analysisofanatomicalvariationsacrosspatientsorduringgrowth[7],and(4)atlas-basedsegmentation[8].Overviewsofthenon-rigidimageregistrationmethodscanbefoundin[9–12].1.1.Basicsforimageregistrationproblem

Thetaskofimageregistrationisto ndspatialcorrespon-dencesbetweentwoimages,IandJ.Thisisusuallydoneby ndingatransformationTsuchthatIandT(J)arespatiallymatched,accordingtoanimage-to-imagedissimilaritymeasure,C(I,T(J)).IandJarereferredasthesourceimageandthe oatingimage,respectively,andT(J)referstotheresultantimageafterapplyingTtoJ.Mathematically,theregistrationproblemcanbede nedas ndingtheoptimaltransformationTÃsuchthatTüargminCðI,TðJÞÞ:

T

ð1Þ

E-mailaddresses:csswk@cse.ust.hk(R.W.So),

cstommy@cse.ust.hk(T.W.Tang),achung@cse.ust.hk(A.C.Chung).NotethataminimizationisusedsinceweassumeCðÁ,ÁÞincreaseswiththedegreeofdissimilarity.

Fornon-rigidimageregistration,thereisavarietyofmethodsproposedintheliteratureregardinghowthetransformationTshouldbemodeled.SomeparametricmodelsrestrictTtobeof