Hybrid Deep Learning Framework for the Analysis of Ganglion Cell Complex and Optic Nerve Head from SD-OCT Scans

Abstract

Glaucoma isaneyeconditionthatoccursduetoincreaseinintra-ocularpressure whichdamagestheopticnerve.Itisaprogressivedegenerativeopticneuropathyand optic nervedamageisirreversible.Glaucomaisworld'sleadingcauseofirreversible blindness, butcanbepreventedwithproperandtimelytreatment.It'sprevention and treatmenthasbeenamajorfocusofinternationaldirectivesanddi erentlatest imaging toolandtechniqueshavebeendevelopedforearlydetectionandmonitoring of glaucoma.OpticalCoherenceTomography(OCT)isoneofthemostadvanced imaging techniquefordetectionandanalysisofvariousretinaldiseases.OCTisnow widely beingusedfordetection,analysisandmonitoringofglaucomabutinmostof the scenariosadetailedmanuale ortisrequired.Di erentresearchersandleading researchgroupsacrosstheglobeareworkinginthisareatoassistophthalmologists for reliableandearlydetectionofthisdisease.However,detailedautomatedanal-ysis ofOCTscanstoprovidemoreusefulandindepthclinicalinsightsisstillan area toexplore.Inthisresearch,wepresentarobustframeworkfordetailedanal-ysis ofOCTscanswhichprovidesinsightsrelatedtomajorretinallayersdirectly associatedwithglaucomatohelpophthalmologistsinglaucomascreening,grading and progressmonitoring.Theproposeframeworkisdividedintotwomainmodules where the rstmodulefocusesonextractionofcuptodiscratio(CDR)asaclinical indicator andsecondmoduledealswithglaucomamonitoringthroughanalysisof ganglion cellcomplex(GCC)andlaminacribrosa(LC)regions.The rstmodule classi es theinputOCTscanintohealthandglaucomatousimagebasedonCDR estimation. TheCDRvalueiscalculatedusingtheinnerlimitingmembrane(ILM) and retinalpigmentedepithelium(RPE)layersofretina.Theproposedframework uses structuretensorstoextractcandidatelayerpixels,andapatchacrosseachcan- didate layerpixelisextracted,whicharefurtherclassi edasILMandRPEusing convolutionalneuralnetwork(CNN).Theselayersarefurtherre nedusinggraph searchbyaddressingmissingandnoisypoints.Theclinicallyde nedgeometryof ILM andRPElayersisusedtocomputeCDRvalueanddiagnoseglaucoma.The second moduleperformsclassi cationintohealthyandglaucomaimage,andgrad-ing intoearlyandadvancecasesbasedonGCCanalysis.Theproposedframework encompasses ahybridconvolutionalnetworkthatextractstheretinalnerve ber layer(RNFL),retinalganglioncell(RGC)withtheinnerplexiformlayer(IPL)and GCC regions.Thethicknesspro lesoftheseextractedregionsarecomputedand their meanvaluesarepassedasafeaturevectortothesupervisedsupportvector machinesforgradingthescreenedglaucomatousscanaseitherearlysuspectora severecase.BothmodulesofproposedframeworkarevalidatedusingalocalOCT dataset from196patients,asubpartofwhichhasbeenmadepubliclyavailableto other researchers.The rstmoduleisabletoextractILMandRPEwithabso- lute meanerrorof6.01and5.44pixels,respectively,andit ndsCDRvaluewithin averagerangeof 0:09 ascomparedwithglaucomaexpert.Thesecondmodule of proposedframeworkachievestheF1scoreof0.9577fordiagnosingglaucoma,a mean dicecoe cientscoreof0.8697forextractingtheRGCregionsandanaccuracy of 0.9117forgradingglaucomatousprogression.Furthermore,theperformanceof the proposedframeworkisclinicallyveri edwiththemarkingsoffourexpertoph-i thalmologists, achievingastatisticallysigni cantPearsoncorrelationcoe cientof 0.9236. Theproposedframeworkcontributesbyprovidingquantitativeassessment of structuralabnormalitieslikeCDR,GCCandLCtodetect,analyzeandgrade glaucoma usingOCTscans.