A Convolutional Neural Network Approach for 2D3D Medical Image Registration

《A Convolutional Neural Network Approach for 2D3D Medical Image Registration》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、ACONVOLUTIONALNEURALNETWORKAPPROACHFOR2D/3DMEDICALIMAGEREGISTRATIONShunMiaoy?Z.JaneWangyRuiLiao?yElectricalandComputerEngineering,UniversityofBritishColumbia,Canada?MedicalImagingTechnology,SiemensMedicalSolution,USAABSTRACTverysmallcapturerange,andtheirrobustnessdegradessig-nificantlyastheerro

2、rintheinitialtransformationparametersWepresentaConvolutionalNeuralNetwork(CNN)regres-increases.Inaddition,intensity-basedmethodsarecomputa-sionbasedframeworkfor2-D/3-Dmedicalimageregistra-tionallyexpensive,becausenumericaloptimizationoftenre-tion,whichdirectlyestimatesthetransformationparamete

3、rsquiresalargenumberofevaluationsoftheobjectivefunction,fromimagefeaturesextractedfromtheDRRandtheX-rayeachevolvingarenderingoftheDRRandacomputationofimagesusinglearnedhierarchicalregressors.Ourframeworkthesimilaritymeasure,bothofwhicharecomputationallyex-consistsoflearningandapplicationstages

4、.Inthelearningpensive.Therefore,intensity-basedmethodsaretypicallynotstage,CNNregressorsaretrainedusingsupervisedmachinesuitablefortime-criticalorreal-timeapplications.learningtorevealthecorrelationbetweenthetransformationFeature-basedmethodsusesimilaritymeasuresthatareparametersandtheimagefea

5、tures.Intheapplicationstage,computedbasedonthegeometricfeaturesextractedfromCNNregressorsareappliedonextractedimagefeaturesinaboth2-Dand3-Dimages,e.g.,corners,linesandsegmenta-hierarchicalmannertoestimatethetransformationparame-tions[6][7][8].Byutilizinggeometricfeatures,feature-basedters.Oure

6、xperimentresultsdemonstratethattheproposedmethodstypicallyhaveahighercomputationalefficiencyandmethodcanachievereal-time2-D/3-Dregistrationwithveryalargercapturerangecomparingtointensity-basedmethods.high(i.e.,sub-milliliter)accuracy.Thedownsideoffeature-basedmethodsliesinthefactthatIndexTerms—

7、2-D/3-DRegistration,ImageGuidedIn-registrationpurelyreliesontheextractedgeometricfeatures.tervention,ConvolutionalNeuralNetwork,DeepLearningTherefore,theerrorinfeaturedetectionstepisinevitablypropagatedintotheregistrationresult,andinsom