EricCosman,Jr.&WilliamWellsArtificialIntelligenceLaboratory
MassachusettsInstitueOfTechnologyCambridge,Massachusetts02139http://www.ai.mit.edu
TheProblem:Weworktowardamethodfortheautomaticregistationofpre-operativeCTand/orMRwithintra-operative2Dx-rayimages(e.g.frombiplanefluoroscopy).Specifically,weinvestigateanintensity-basedapproachtorigid3D-2Dcoregistration,wherebyalignmentisachievedbyoptimizingasimilaritymeasurethatisevaluatedonthejointintensitiesofthex-raydataandofdigitally-reconstructedradiographs(DRR)generatedfromthevolumetricdata.Theeffortisnowaimedatproducingademonstrationsystemthatwillfunctiononplasticpelvismodels.Motivation:Thereareanumberoftherapiestowhichwemightfindapplicationforthefusionofpreoperatively-processedCTand/orMRwithintraoperatively-acquiredfluoroscopy.Currentlytherearemanytypesofsurgeriesthatrequiresuperiorguidancewhich,bycurrentmethods,requirecomplicatedmeasurementandreferenceschemes,useoffiducials,ornear-continuoususeoffluoroscopy(whichnotinfrequentlyresultsinradiation-relatedinjuriessuchashairloss).
ExamplesoftherapiessuitedtoimprovementbyMR-fluoroscopyregistration,involveaccesstosofttissuetargets,suchastumors,bytheirrelationtobonystructuresnearby.Inthesecases,MRcandefinepreoperativelyandin3D,therelationshipofthesofttissuetobonystructureswithconfidence,andthefluoroscopycangiveaccurateinformationaboutthestateofthebonystructures(and,byinference,thesofttissue)duringaprocedure.SpecificexamplesofprodeduresarePallidotomy,braintumorresections,andspinaldiscdecompressions.
TherapiesthatcouldbeimprovedbyCT-fluoroscopyregistrationarethosethatinvolveaccesstostructureswhichcanbevisualizedbyx-raywiththeadditionofcontrastagents,butwhosegeometeriesareambiguousinradiographically-projectedimages.Aspecificexampleistranscatheterproceduresinwhichacatheterisnavigatedthroughcomplicatedvascularstructuresintheheadtocorrectflowproblemsandvascularabnormalities.Anotherexampleis“re-orientingosteotomies”wherebyone’sabilitytowalkcanbeimprovedbymakingcutsinthepelvisaroundtheacetabularcups.PreviousWork:Wehavepreviousexperiencewith3D-3Dregistrationofmedicalimages,aswellas2D-3Dregistra-tionofvideoimages,usingthemethodof“AlignmentbyMaximizationofMutualInformation”.[1,2]Whileothergroupshaveapproached2D-3D,rigidregistrationproblemfromavarietyofdirections(Pizeretal.attheUniversityofNorthCarolina,Duncanetal.atYale,andPenneyetal.[3]),wefeelthatthereisampleremainingopportunity.Approach:Weinvestigateanumberofsimilaritymeasureswhichhavebeenappliedinothermulti-modalregistrationcontexts.Thesimilaritymeasuresarejudgedempiricallybyprobingthespaceoftransformationsbetweenthecoor-dinateframesofanMRorCTvolumeandbiplanarDRRsproducedfromacorrespondingCTvolume.Thismethodallowsustoknowthe“groundtruth”oftheregistration,whichcouldbeestablishedbyprovenmethodsforrigid,3D-3D,CT-MRalignment.[1]Thisevaluationofmeasuresforaccuracyandrobustnesswillbeusedasastartingpointforfurtherapplication-specificalgorithmdevelopment.
Futhermore,weproposeamethodofDRRproductioncalledvoxel-projectionwhichdrasticallyreducesprocessingtimerelativetostandardray-castingmethods.Wehypothesizethatitsspeedmayenable3D-2DregistrationmethodswhichrelyonevaluationofasimilaritymeasureovertheentiretyofaDRRandmodelradiographtoaccountforstructuralorvariationalfeatures(forinstance,patternintensity[3]).
Basedonsimilaritymeasurechacteristicsobservedinourprobingexperiments,andusingthevoxel-projectionmethod,weadapttheuphill-simplexoptimizationalgorithmtoimplementanintensity-basedMR-fluoroscopyrigidregistatra-tionengine.Wealsoconsideraparallelizedversionoftheregistrationalgorithmpresentedin[3].
Difficulty:Theinnerloopofanyintensity-basedregistrationalgorithmofthiskindwillinvolvetheproductionofDRRsfromavolumetricimage.AsevenoptimizedmethodsforfullDRRproductionhaverunningtimesontheorder
1
ofsecondsformedicalimagesoftypicalsize,theefficiencyoftheoptimizationschemeisimportanttomakingthealgorithmserviceablyfast.SomedegreeofparallelizationorsparseevaluationofDRRsmaybeneededforthisreason.Impact:Thecomputationalefficiencyofvoxel-projectionmakesitausefultoolfortheinvestigationofsimilaritymeasuresinthis3D-2Dcontext.ThequalityofDRRsproducedbyvoxel-projectionisclosetothatofthoseproducedbyray-casting,butthespeedofprojectionisincreasedbymorethatanorderofmagnitude,asshowninFigures1and2.
Comparisonsofsimilaritymeasuresonalimiteddatasetsuggestthatpatternintensityiswell-suitedtothistypeof3D-2Dregistration.
Volume with 317x390x29 8−bit pixels300Voxel−projectionNearest−neighborTri−linear 140Volume with 154x191x29 8−bit pixelsVoxel−projectionNearest−neighborTri−linear 120250100Running time (seconds)Running time (seconds)0123456DRR resolution (mm / pixel width)7892008015060100405020000123456DRR resolution (mm / pixel width)789Figure1:Comparisonoftherunningtimesofvoxel-projectionandray-casting(usingtrilinearandnearest-neighborinterpolationtoderiveinter-voxelMRintensities)overaseriesofDRRsizes,givenintermsofDRRpixelresolution.
10317x390x29 volume154x191x29 volume987Running time (seconds)65432100123456DRR resolution (mm / pixel width)789
因篇幅问题不能全部显示，请点此查看更多更全内容