Extraction of visual features with eye tracking for saliency driven 2D3D registration
ImageandVisionComputing23(2005)999–1008
doc.guandang.net
Extractionofvisualfeatureswitheyetrackingforsaliencydriven
2D/3Dregistration
AdrianJ.Chung*,FaniDeligianni,Xiao-PengHu,Guang-ZhongYang
RoyalSociety/WolfsonFoundationMedicalImageComputingLaboratory,DepartmentofComputing,ImperialCollege,
180Queen’sGate,SW72BZLondon,UK
Received30January2004;receivedinrevisedform19May2005;accepted1July2005
Abstract
Thispaperpresentsanewtechniqueforderivinginformationonvisualsaliencywithexperimentaleye-trackingdata.Thestrengthandpotentialpitfallsofthemethodaredemonstratedwithfeaturecorrespondencefor2Dto3Dimageregistration.Withthisapplication,aneye-trackingsystemisemployedtodeterminewhichfeaturesinendoscopyvideoimagesareconsideredtobesalientfromagroupofhumanobservers.Byusingthisinformation,abiologicallyinspiredsaliencymapisderivedbytransformingeachobservedvideoimageintoafeaturespacerepresentation.Featuresrelatedtovisualattentionaredeterminedbyusingafeaturenormalisationprocessbasedontherelativeabundanceofimagefeatureswithinthebackgroundimageandthosedwelledonvisualsearchscanpaths.Thesefeaturesarethenback-projectedtotheimagedomaintodeterminespatialareaofinterestforeachunseenendoscopyvideoimage.Thederivedsaliencymapisemployedtoprovideanimagesimilaritymeasurethatformstheheartofanew2D/3Dregistrationmethodwithmuchreducedrenderingoverheadbyonlyprocessing-selectiveregionsofinterestasdeterminedbythesaliencymap.Signi cantimprovementsinposeestimationef ciencyareachievedwithoutapparentreductioninregistrationaccuracywhencomparedtothatofusinganintensity-basedsimilaritymeasure.
q2005ElsevierB.V.Allrightsreserved.
Keywords:2D/3Dregistration;Saliency;Imagecorrelation;Eyetracking
1.Introduction
Newtechniquesforminimallyinvasivesurgery[1,2]havebroughtsigni cantbene tstopatientcare,includingreducedtrauma,shortenedhospitalisation,andimproveddiagnosticaccuracyandtherapeuticoutcome.Endoscopyisthemostcommonprocedureinminimalaccesssurgery,butrequiresahighdegreeofmanualdexterityfromtheoperatorowingtothecomplexityoftheinstrumentcontrols,restrictedvisionandlackoftactileperception.Trainingmethodsfortheseskillsinvolvetheuseofcomputersimulation,whichneedstobeasrealisticaspossibleforspecialisttrainingtobeeffective.Oneimportantfactoris
*Correspondingauthor.Tel.:C442075948318;fax:C442075818024.
E-mailaddress:a.chung@imperial.ac.uk(A.J.Chung).
0262-8856/$-seefrontmatterq2005ElsevierB.V.Allrightsreserved.doi:10.1016/j.imavis.2005.07.003
thelevelofvisualrealismpresentedtothetrainee,whichismadepossiblebyrecentadvancesinimage-basedmodellingandrendering[3].Genuinesurfacetextureinformationcanbeextractedfrommatchedrealendoscopicvideoswithgeometryderivedfromtomographicimagesofthesamepatient.Thisprocessreliesontheaccurateregistrationof2Dvideoimagesto3Dtomographicdatasets.Themethodof2D/3Dregistrationisamuchresearchedtopic,andmanyapplicationsarisenarefrommedicalimagingandsurgicalplanningapplications[4–10].
Ingeneral,methodsofmedicalimageregistrationcanbeclassi edintolandmark,segmentation,andvoxel-basedtechniques[11,12].Voxel-basedtechniques[13]haveshowntobeeffectivefortheregistrationofbothunimodalormultimodalimagesdependingonthesimilaritymeasuresandoptimisationstrategyused.Oneofthetypical2D/3Dregistrationproblemsaddressedistheregistrationof uoroscopeimagesagainst3DCTorMRIdatasets,forwhichavoxelsimilaritymeasureistypicallyused.Registering2Dbronchoscopyvideoto3D



tomographic
erfromtheeye-trackingdatasettolearndifferentfeatures’parameters forsaliencymapconstruction.Fig.1(d)showsthatitachievesbetterresultsthantheinformation-drivenmethod[14...
erfromtheeye-trackingdatasettolearndifferentfeatures’parameters forsaliencymapconstruction.Fig.1(d)showsthatitachievesbetterresultsthantheinformation-drivenmethod[14...
热门文档