Intraoperative Segmentation and Nonrigid Registrationfor Image Guided Therapy

SimonK. Warfield�, Arya Nabavi

��� �, TorstenButz

�, KemalTuncali

�, StuartG.

Silverman�, PeterMcL. Black

�, FerencA. Jolesz

�, andRonKikinis

��SurgicalPlanningLaboratoryhttp://www.spl.harvard.edu

�Departmentof

Radiology,�Departmentof Surgery, HarvardMedicalSchoolandBrighamandWomen’s

Hospital,75FrancisSt.,Boston,MA 02115USA,

SignalProcessingLaboratory, SwissFederalInstituteof Technologyat Lausanne,1015Lausanne,Switzerland.

warfield,arya,butz,ktuncali,jolesz,kikinis� @bwh.harvard.edu,pmblack@bics.bwh.harvard.edu,sgsilverman@partners.org

Abstract. Our goalwasto improve imageguidanceduring minimally invasiveimageguidedtherapy by developinganintraoperativesegmentationandnonrigidregistrationalgorithm.Thealgorithmwasdesignedto allow for improved navi-gationandquantitative monitoringof treatmentprogressin orderto reducethetime requiredin theoperatingroomandto improveoutcomes.Thealgorithmhasbeenappliedto intraoperative imagesfrom cryotherapy of theliver and from surgery of the brain. Empirically the algorithmhasbeenfoundto be robust with respectto imagingcharacteristicssuchasnoiseandintensityinhomogeneityandrobustwith respectto parameterselection.Serialandparal-lel implementationsof the algorithmare sufficiently fast to be practicalin theoperatingroom.Thecontributionsof this work areanalgorithmfor intraoperative segmentationandintraoperative registration,amethodfor quantitative monitoringof cryother-apy from real-timeimaging,quantitative monitoringof braintumorresectionbycomparisonto a preoperative treatmentplan andan extensive validationstudyassessingthereproducibilityof the intraoperative segmentation.We have evalu-atedour algorithmwith six neurosurgical casesandtwo liver cryotherapy caseswith promisingresults.Furtherclinical validationwith larger numbersof caseswill benecessaryto determineif our algorithmsucceedsin improving intraoper-ative navigationandintraoperative therapy delivery andhenceimprovestherapyoutcomes.1

1 Introduction

Imageguidedsurgical techniquesareusedin operatingroomsequippedwith specialpurposeimagingequipment.Thedevelopmentof imageguidedsurgical methodsoverthepastdecadehasprovideda majoradvancein minimally invasive therapy delivery.Earlywork suchasthatreviewedby Jolesz[1] hasestablishedtheimportanceandvalueof imageguidancethroughbetterdeterminationof tumormargins,betterlocalizationoflesions,andoptimizationof thesurgicalapproach.

1 Thispaperwill appearatMICCAI 2000:Third InternationalConferenceonMedicalRobotics,ImagingAnd ComputerAssistedSurgery

Researchin imageguidedtherapy hasbeendrivenby theneedfor improvedvisu-alization.Qualitative judgementsby expertsin clinical domainshave beenrelieduponasquantitative andautomatedassessmentof intraoperative imagingdatahasnot beenpossiblein thepast.In orderto provide thesurgeonor interventionalradiologistwithasrich a visualizationenvironmentaspossiblefrom which to derive suchjudgements,existing work hasbeenconcernedprimarily with imageacquisition,visualizationandregistrationof intraoperativeandpreoperativedata.Intraoperativesegmentationhasthepotentialto be a significantaid to the intraoperative interpretationof imagesand toenablepredictionof surgicalchanges.

Earlier work hasbeena steadyprogressionof improving imageacquisitionandintraoperative imageprocessing.This hasincludedincreasinglysophisticatedmulti-modalityimagefusionandregistration.Clinical experiencewith imageguidedtherapyin deepbrainstructuresandwith largeresectionshasrevealedthelimitationsof existingrigid registrationandvisualizationapproaches[1]. Thedeformationsof anatomythattake placeduringsuchsurgeryareoften betterdescribedasnonrigidandsuitableap-proachesto capturesuchdeformationsarebeingactively developedby severalgroups(describedbelow).

A numberof imagingmodalitieshavebeenusedfor imageguidance.Theseinclude,amongstothers,computedtomography(CT), ultrasound,digital subtractionangiogra-phy (DSA), andmagneticresonanceimaging(MRI). Intraoperative MR imagingcanacquirehigh contrastimagesof soft tissueanatomywhich hasprovento bevery use-ful for image-guidedtherapy [2]. Multi-modality registrationallows preoperative datathatcannotbeacquiredintraoperatively, suchasfMRI or nuclearmedicinescans,to bevisualizedtogetherwith intraoperativedata.

Geringet al. [3] describedan integratedsystemallowing the readyvisualizationof intraoperative imageswith preoperative data,includingsurfacerenderingof previ-ously preparedtriangle modelsand arbitrary interactive resamplingof 3D grayscaledata.Multiple imageacquisitionswerepresentedin a combinedvisualizationthroughrigid registrationandtrilinear interpolation.The systemalsoallows for visualizationof virtual surgical instrumentsin the coordinatesystemof the patientandpatientim-ageacquisitions.Thesystemsupportsqualitativeanalysisbasedonexpertinspectionofimagedataandthesurgeonsexpectationof whatshouldbepresent(normalanatomy,patient’sparticularpathology, currentprogressof thesurgeryetc.)

Severalgroupshave investigatedintraoperative nonrigidregistration,primarily forneurosurgicalapplications.Theapproachescanbecategorizedby thosethatusesomeform of biomechanicalmodel(recentexamplesinclude[4–6]) andthosethat apply aphenomenologicalapproachrelying uponimagerelatedcriteria (recentexamplesin-clude[7,8].)

We aimedto demonstratethatintraoperativesegmentationis possibleandaddssig-nificantly to thevalueof intraoperativeimaging.Comparedto registrationof preopera-tive imagesandinspectionof intraoperative imagesalone,intraoperativesegmentationenablesidentificationof structuresnot presentin previous images(examplesof suchstructuresincludetheregionof cryoablationor radiofrequency treatmentarea,surgicalprobesandchangesdueto resection),quantitativemonitoringof theprogressof therapy

(includingtheability to comparequantitatively with a preoperatively determinedtreat-mentplan)andintraoperativesurfacerenderingfor rapid3D interactivevisualization.

2 Method

Preoperative data:MRI,fMRI, SPECT,PET and/or CT

Manual or AutomatedSegmentation

SpatialLocalization

Preparation for Image Guided Therapy

Intraoperative MRI

Rigid Registration

During Image Guided Therapy

SpatiallyVaryingClassification

NonrigidRegistration

Segmented intraoperative MRIand registered preoperative data

Fig. 1. Schemafor Intraoperative SegmentationandRegistration

In orderto successfullysegmentintraoperative images,we have developedan imagesegmentationalgorithmthattakesadvantageof anexistingpreoperativeMR acquisitionandsegmentationto generatea patient-specificmodelfor thesegmentationof intraop-erative data.The algorithmusesthe segmentationof preoperative dataasa templatefor thesegmentationof intraoperativedata.Figure1 illustratestheprocessingstepsthattakeplacebeforeandduringthetherapy procedure.

We have experimentedfor several yearswith a generalimagesegmentationap-proachthat usesa 3D digital anatomicalatlasto provide automaticlocal context forclassification[9–11].Thework describedhereextendsour previouswork to ensureitssuitability for intraoperativesegmentation.Ratherthana genericdigital anatomicatlasweproposeheretouseasegmentedpreoperativepatientscanto deriveapatient-specificanatomicalmodelfor intraoperativesegmentationof new scans.

Sincepreoperative datais acquiredbeforesurgery, the time availablefor segmen-tation is longer. This meanswe canusesegmentationapproachesthat arerobust andaccuratebut aretime consumingandhenceimpracticalto usein the operatingroom.In our laboratory, preoperative datais segmentedwith a variety of manual[3], semi-automated[12] or automated[13,11] approaches.We attemptto selectthemostrobustandaccurateapproachavailablefor agivenclinical application.Eachsegmentedtissueclassis thenconvertedinto anexplicit 3D volumetricspatiallyvaryingmodelof thelo-cationof thattissueclass,by computingasaturateddistancetransform[14] of thetissue

class.Thismodelis usedto providerobustautomaticlocalcontext for theclassificationof intraoperativedatain thefollowing way.

During surgery, intraoperative datais acquiredandthe preoperative data(includ-ing any MRI/fMRI/PET/SPECT/MRAthatis appropriate,thetissueclasssegmentationandthespatiallocalizationmodelderivedfrom it) is alignedwith theintraoperativedatausinganMI basedrigid registrationmethod[15,3]. Theintraoperativeimagedatathentogetherwith the spatiallocalizationmodel forms a multichannel3D dataset.Eachvoxel is thena vectorhaving componentsfrom the intraoperative MR scan,the spa-tially varyingtissuelocationmodelandif relevantto theparticularapplication,any oftheotherpreoperativeimagedatasets.For thefirst intraoperativescanto besegmenteda statisticalmodelfor theprobabilitydistribution of tissueclassesin the intensityandanatomicallocalizationfeaturespaceis built. Thestatisticalmodelis encodedimplicitlyby selectinggroupsof prototypicalvoxelswhich representthetissueclassesto beseg-mentedintraoperatively (lessthanfiveminutesof userinteraction).Thespatiallocationof theprototypevoxelsis recordedandis usedto updatethestatisticalmodelautomati-cally whenfurtherintraoperativeimagesareacquiredandregistered.Thismultichanneldatasetis thensegmentedwith a spatiallyvaryingclassification[10,13,16].

Segmentationof intraoperative datahelpsto establishexplicitly the regionsof tis-suesthatcorrespondin the preoperative andintraoperative data.It is thenstraightfor-wardtoapplyourpreviouslydescribed[17,18]andvalidated[19] multi-resolutionelas-tic matchingalgorithm.Oncethenonrigidtransformationmappingfrom thepreopera-tive to the intraoperative datahasbeenestablished,the mappingis appliedto eachoftherelevantpreoperativedatasetsto bring theminto alignmentwith theintraoperativescan.

3 Results

In thissectionillustrative segmentationsandtwo validationexperimentsarepresented.Duringinterventionalproceduresin theliverandbrain,intraoperativeMRI (IMRI) datasetswereacquiredandstored.Our segmentationandnonrigid registrationalgorithmwasappliedto thesedatasetsafter therapy delivery in orderto allow us to assesstherobustness,accuracy andtime requirementsof the approach.In the future we intendto carry out segmentation,nonrigid registrationandvisualizationusingthe approachdescribedhereduringthe interventionalprocedureswith thegoalof improving imageguidedtherapy outcomes.

3.1 Intraoperative Segmentation for Neurosurgery

Figure2 illustratesthesegmentationof six neurosurgerycasesusingour intraoperativesegmentationalgorithm.In eachcase,several volumetricMRI scanswerecarriedoutduringsurgery. Thefirst scanwasacquiredatthebeginningof theprocedurebeforeanychangesin theshapeof thebraintook place,andthenover thecourseof surgeryotherscanswereacquiredasthesurgeoncheckedtheprogressof tumorresection.Thefinalscanin eachsequenceexhibits significantnonrigiddeformationandlossof tissuedueto tumor resection.In order to testour segmentationapproacheachsubsequentscan

wasalignedto thefirst by maximizationof mutualinformationandthefirst scanwasmanuallysegmentedto actasanindividualizedanatomicalmodel.Thelastscanin eachsequencewas then segmentedwith our new segmentationapproach.The segmentedbrainandIMRI is shown in Figure2. In orderto checkthequalityof thesegmentation,eachsegmentationwasvisually comparedwith theMRI from which it wasderived.Ineachcasethesegmentedbraincloselymatchedtheexpectedlocation.

Fig. 2. Visualizationsof intraoperative segmentationof braintissuefrom six neurosurgerycases.Lessthanfive minutesof userinteractionwasrequiredfor eachsegmentation.The segmentedbraintissueis shown in white with surfacerenderingandtheIMRI is texturemappedin planesalongthecoordinateaxes.This allows readycomparisonof thepositionof thesegmentedbrainandtheIMRI (skinappearsbright,brainis graycloselymatchingthesegmentedbrainborder).

User Interaction and ComputationalRequirementsEachbrain segmentationof Fig-ure2 involvesthesegmentationof theentire3D IMRI scanof 256x256x60= 3,932,160voxels(with voxel size0.9375x0.9375x2.5mm

�.) Suchavolumeis acquiredin approx-

imately10 minutesintraoperatively (a secondscanningmodecanacquirea 2D imagein approximately2 seconds).Lessthanfive minutesof userinteractionwasrequiredfor eachbrainsegmentationshown in Figure2. OnaSunMicrosystemsUltra-10work-stationwith a 440MHzUltraSPARC-IIi CPUand512MB RAM eachbraintissueseg-mentation(excludinggenerationof spatiallocalizationmodelswhich requiresapprox-imately 200 secondsandcanbe donepreoperatively andexcluding rigid registrationwhich requiresapproximately30 secondsusingmaximizationof mutual information[15]) requiredlessthan330secondsto complete.As wehavepreviouslydescribed,par-allel tissueclassificationcanachieve excellentspeedups[20]. On a SunMicrosystemsUltra-80serverwith 4 x 450MHzUltraSPARC-II CPUsand2GBRAM, eachbraintis-

suesegmentationrequiredlessthan130secondsto complete.Thiscanbecomparedtoa typical manualsegmentationthatcantake 1800–3600secondsandhassignificantlylessreproducibility.

3.2 Intraoperative Segmentation for Liver Cryotherapy

(a)Segmentation (b) IMRI (c) Tumorlocalization (d) Liver localization

Fig. 3. IMRI of theliverandthesegmentationof theliverandtumor. Threeof thefeaturechannelsusedto carryout thesegmentationshown in (a)areshown in (b), (c) and(d).

Figure3 shows IMRI of theliver andthesegmentationof theliver andtumor. Intraop-erativeMRI hasbeenusedto guidepercutaneouscryotherapy of liver tumors[21]. Thisfigureillustratesthespatiallocalizationof liverandtumorfrom a3D volumetricpreop-erative segmentation(not shown) andindicatesthat isointensebut differentstructuresin the IMRI canbesuccessfullysegmentedin the joint featurespaceformedwith theIMRI intensityandspatiallocalizationinformation.

3.3 Intraoperative Monitoring of Cryotherapy Iceball Formation

Tumortarget Freezing Freezing Freezing Thawing

Fig. 4. Intraoperative imagingof iceballformation.Thelesionis bright in thefirst imageandtheiceball grows to cover it. The iceball appearsasa dark region in the liver, which grows whilefreezingandshrinksduringthawing. The intraoperative iceball segmentationobtainedwith ourmethodis indicatedby thewhiteoutline.

Figure4 shows the intraoperative appearanceof an iceball during cryotherapy of an-othercase.Ourintraoperativesegmentationalgorithmallowedrapid,robustandstraight-forwardsegmentationof theiceball.By comparingthesegmentationof theiceballwith

a preoperativeplanof thedesirediceballsizeandlocationthetherapy progresscanbemonitoredquantitatively.

3.4 Validation Experiments

Key parametersin oursegmentationalgorithmaretheprototypevoxelswhichimplicitlymodeltheprobabilitydistributionof theintensitiesof tissueswhichareto besegmentedandthealignmentof thespatiallocalizationmodelswhichformpartof thefeaturespacein whichthetissueclassificationtakesplace.Westudiedtheeffectof variationsin thesekey parametersuponthesegmentation.

Reproducibility: Variations in prototype selection Table1 recordsthevariability ofbrain segmentationfrom a singleneurosurgerycasewhenthe setof prototypevoxelsmodelingthetissuecharacteristicsis varied.Thesetof prototypesusedfor thesegmen-tationwassubsampledby randomlyselecting90%of theprototypes100 times.Eachof the100subsetssimulatesdifferentuserprototypeselection.EachsubsetwasusedtosegmenttheIMRI usingthenew methodandthevolumeof thebrainsegmentationwasrecorded.Themean,minimumandmaximumvolumerecordedareshown in thetablealongwith thecoefficient of variation(C.V.) of thevolumeof segmentation(which islessthan1%). This indicatesthe segmentationis extremelyrobust in the presenceofvariability in theprototypevoxel selection.

Minimum volumeMaximumvolume Meanvolume C.V. (%)401074voxels 422440voxels 414440voxels 0.97

Table 1. Measuresof variability of thevolumeof thebrain(unitsarevoxels)in repeatedsegmen-tations,with differentselectionsof prototypevoxels,from brainIMRI.

Reproducibility: Variations in preoperative model alignment In orderto determinetheinfluenceof thealignmentof thepreoperativesegmentationupontheintraoperativesegmentation,we selecteda neurosurgerycaseandsegmentedthe brain asdescribedabove.Wethenappliedasetof translationsandrotationsto thepreoperativesegmenta-tion sothatit wasno longercorrectlyregisteredto theIMRI to besegmented.For eachtranslationandrotationweappliedoursegmentationmethodandobtaineda segmenta-tion of thebrain.We thencomparedthevolumeof tissuesegmentedasbrain for eachmisalignedpositionwith thatobtainedwith thecorrectalignmentandrecordedtheratioof thenew segmentationvolumeto theoriginal segmentationvolume.Thevariationinsegmentationwith translationsalongeachof thecoordinateaxesis shown in Figure5.Similar results(not shown) wereobtainedfor rotationsaroundeachof thecoordinateaxes.

Intrapatientregistrationwith maximizationof mutualinformationhasa typical ac-curacy smallerthanonevoxel (in this case0.9375x0.9375x2.5mm

�). Theperturbation

of theregistrationof thepreoperativesegmentationdoesnotcausea significantchangein the intraoperative segmentationuntil this misalignmentreachesaround �� �� mm,which indicatesour intraoperative segmentationmethodis robust to misalignmenter-rorsandalsoto errorsin thepreoperativesegmentation.

6065707580859095

100

-40 -30 -20 -10 0 10 20 30 40

Bra

in v

olum

e (V

_x/V

_0 %

)

Translation along x axis (mm)

Segmented brain volume.

ratio (%)

6065707580859095

100

-40 -30 -20 -10 0 10 20 30 40

Bra

in v

olum

e (V

_y/V

_0 %

)

Translation along y axis (mm)

Segmented brain volume.

ratio (%)

6065707580859095

100

-40 -30 -20 -10 0 10 20 30 40

Bra

in v

olum

e (V

_z/V

_0 %

)

Translation along z axis (mm)

Segmented brain volume.

ratio (%)

Fig. 5. Reproducibilityof braintissuesegmentationasthespatiallocalizationmodelis translated.The low variability in segmentationaroundthe correctalignmentindicatesthe segmentationisrobustto misregistrationandspatiallocalizationerrors.

4 Discussion and Conclusion

Ourearlyexperiencewith two livercasesandsix neurosurgerycasesindicatesthatourintraoperativesegmentationalgorithmis arobustandreliablemethodfor intraoperativesegmentation.It requireslittle userinteraction,is robust to variationin theparametersthatrequireinteraction,andis sufficiently fastto beusedintraoperatively.

Theapplicationof ourpreviouslydescribedandvalidatednonrigidregistrationalgo-rithm is enabledby intraoperativesegmentationestablishingthecorrespondingtissuesin thedatasetsto bealigned.Furtherwork is neededto characterizetheaccuracy androbustnessof ournonrigidregistrationfor intraoperativedata,especiallyin theliver.

Intraoperative segmentationaddssignificantlyto thevalueof intraoperative imag-ing. Comparedto registrationof preoperative imagesand inspectionof intraopera-tive imagesalone,intraoperative segmentationenablesidentificationof structuresnotpresentin previousimages(examplesof suchstructuresincludetheregion of cryoab-lation or RF treatmentarea,surgicalprobesandchangesdueto resection),quantitativemonitoringof therapy applicationincludingtheability to comparequantitativelywith apreoperativelydeterminedtreatmentplanandintraoperativesurfacerenderingfor rapid3D interactivevisualization.

Thecontributionsof thiswork areanalgorithmfor intraoperativesegmentationandintraoperative registration,a methodfor quantitative monitoringof cryotherapy fromreal-timeimaging,quantitative monitoringof brain tumor resectionby comparisontoa preoperative treatmentplan and an extensive validationstudy assessingthe repro-ducibility of theintraoperativesegmentation.Empiricallythealgorithmhasbeenfoundto berobustwith respectto imagingcharacteristicssuchasnoiseandintensityinhomo-geneityandrobustwith respectto parameterselection.Serialandparallelimplementa-tionsof thealgorithmaresufficiently fastto bepracticalin theoperatingroom.

Wehaveevaluatedouralgorithmwith sixneurosurgicalcasesandtwo livercryother-apy cases.Furtherclinical validationwith largernumbersof caseswill benecessarytodetermineif our new approachsucceedsin improving intraoperative navigation andintraoperativetherapy deliveryandhenceimprovestherapy outcomes.

Acknowledgements:ThisinvestigationwassupportedbyNIH P41RR13218,NIH P01CA67165andNIH R01RR11747.

References

1. F. Jolesz,“Image-guidedProceduresand the OperatingRoom of the Future,” Radiology,vol. 204,pp.601–612,May 1997.

2. P. M. Black,T. Moriarty, E. Alexander, P. Stieg, E. J.Woodard,P. L. Gleason,C. H. Martin,R. Kikinis, R. B. Schwartz, and F. A. Jolesz,“The Developmentand ImplementationofIntraoperativeMRI andits NeurosurgicalApplications,” Neurosurgery, vol. 41,pp.831–842,April 1997.

3. D. Gering,A. Nabavi, R. Kikinis, W. Grimson,N. Hata,P. Everett,F. Jolesz,andW. Wells,“An IntegratedVisualizationSystemfor Surgical PlanningandGuidanceusingImageFu-sionandInterventionalImaging,” in MICCAI 99: Proceedingsof theSecondInternationalConferenceonMedicalImageComputingandComputerAssistedIntervention, pp.809–819,SpringerVerlag,1999.

4. A. Hagemann,K. Rohr, H. Stiel,U. Spetzger, andJ.Gilsbach,“Biomechanicalmodelingofthehumanheadfor physicallybased,non-rigid imageregistration,” IEEE TransactionsOnMedicalImaging, vol. 18,no.10,pp.875–884,1999.

5. O. SkrinjarandJ. Duncan,“Real time 3D brainshift compensation,” in IPMI’99, pp. 641–649,1999.

6. M. Miga, K. Paulsen,J. Lemery, A. Hartov, andD. Roberts,“In vivo quantificationof ahomogeneousbrain deformationmodel for updatingpreoperative imagesduring surgery,”IEEETransactionsOnMedicalImaging, vol. 47,pp.266–273,February1999.

7. D. Hill, C. Maurer, R. Maciunas,J. Barwise,J. Fitzpatrick,andM. Wang,“Measurementof intraoperative brain surface deformationundera craniotomy,” Neurosurgery, vol. 43,pp.514–526,1998.

8. N. Hata,Rigid anddeformablemedicalimage registration for image-guidedsurgery. PhDthesis,Universityof Tokyo, 1998.

9. S. Warfield,J. Dengler, J. Zaers,C. R. Guttmann,W. M. Wells III, G. J. Ettinger, J. Hiller,andR. Kikinis, “Automatic identificationof Grey Matter Structuresfrom MRI to ImprovetheSegmentationof WhiteMatterLesions,” Journalof Image GuidedSurgery, vol. 1, no.6,pp.326–338,1995.

10. S. K. Warfield,M. Kaus,F. A. Jolesz,andR. Kikinis, “Adaptive TemplateModeratedSpa-tially VaryingStatisticalClassification,” in MICCAI 98: First InternationalConferenceonMedical Image Computingand Computer-AssistedIntervention, pp. 231–238,Springer-Verlag,Heidelberg, Germany, October11–131998.

11. M. R.Kaus,S.K. Warfield,A. Nabavi, E.Chatzidakis,P. M. Black,F. A. Jolesz,andR.Kiki-nis,“Segmentationof MRI of meningiomasandlow gradegliomas,” in MICCAI 99: SecondInternationalConferenceonMedicalImageComputingandComputer-AssistedIntervention(C. TaylorandA. Colchester, eds.),pp.1–10,Springer-Verlag,Heidelberg, Germany, 1999.

12. R.Kikinis, M. E.Shenton,G.Gerig,J.Martin,M. Anderson,D. Metcalf,C.R.G.Guttmann,R. W. McCarley, W. E. Lorenson,H. Cline,andF. Jolesz,“RoutineQuantitativeAnalysisofBrain andCerebrospinalFluid Spaceswith MR Imaging,” Journal of MagneticResonanceImaging, vol. 2, pp.619–629,1992.

13. S.K. Warfield,M. Kaus,F. A. Jolesz,andR. Kikinis, “Adaptive,TemplateModerated,Spa-tially VaryingStatisticalClassification,” Medical Image Analysis, vol. 4, no. 1, pp. 43–55,2000.

14. I. Ragnemalm,“The Euclideandistancetransformin arbitrarydimensions,” PatternRecog-nition Letters, vol. 14,pp.883–888,1993.

15. W. M. Wells,P. Viola, H. Atsumi,S.Nakajima,andR. Kikinis, “Multi-modal volumeregis-trationby maximizationof mutualinformation,” MedicalImage Analysis, vol. 1, pp.35–51,March1996.

16. R. O. DudaandP. E. Hart, PatternClassificationandSceneAnalysis. JohnWiley & Sons,Inc.,1973.

17. J. DenglerandM. Schmidt,“The DynamicPyramid– A Model for Motion AnalysiswithControlledContinuity,” InternationalJournal of Pattern Recognition and Artificial Intelli-gence, vol. 2, no.2, pp.275–286,1988.

18. S.K. Warfield,A. Robatino,J.Dengler, F. A. Jolesz,andR.Kikinis, “NonlinearRegistrationandTemplateDrivenSegmentation,” in Brain Warping(A. W. Toga,ed.),ch.4, pp.67–84,AcademicPress,SanDiego,USA, 1999.

19. D. V. Iosifescu,M. E.Shenton,S.K. Warfield,R. Kikinis, J.Dengler, F. A. Jolesz,andR. W.McCarley, “An AutomatedRegistrationAlgorithm for MeasuringMRI SubcorticalBrainStructures,” NeuroImage, vol. 6, pp.12–25,1997.

20. S. K. Warfield, F. A. Jolesz,andR. Kikinis, “Real-TimeImageSegmentationfor Image-GuidedSurgery,” in Supercomputing1998, pp.1114:1–14,November1998.

21. S.G. Silverman,K. Tuncali,D. F. Adams,E. vanSonnenberg, K. H. Zou,D. F. Kacher, P. R.Morrison,andF. A. Jolesz,“MRI-Guided percutaneouscryotherapy of liver tumors:initialexperience,” Radiology. In press.