96
S.W.Franklin,S.E.Rajan/AppliedSoftComputing22(2014)94–100
extractionoftheimagefeatureandtherecognitionofthevesselstructureareperformedsimultaneously.Analgorithmwhichreliesonmatched lteringapproachwaspresentedinRef.[20],andisinitiatedbystartandendpointsfollowedbyautomaticdetectionofvesselboundariesandthentracksthecenterlineofthevesseland nallyitextractstheusefulparametersofclinicalinterest.TheunsupervisedfuzzyvesseltrackingalgorithmpresentedinRef.[21]isvirtuallyparameter-freeanditperformsautomaticinitial-ization.Italsoovercomestheproblemofvesselpro lemodeling,tracksfundusvesselsautomaticallyandtheclassi cationofvesselandnonvesselregionsalongavesselpro leisdeterminedusingfuzzyC-meansclusteringalgorithm.ThealgorithmpresentedinRef.[22]performsautomaticadaptionwithoutmanualinitializa-tion.Itautomaticallydetectstheseed-pointoftracing,describedaslocalgray-levelminimaanditisapplieddirectlyattheimageintensitylevelwithoutpreprocessing.Consequently,sequencesofexploratorysearchesareinitiatedtodetectthevesselboundariesandthesemodelsareusedintheestimationofnextvesselpoint.InRef.[23]Chutatapeetal.haveproposedatechniquewherethetracingprocessstartsfromthecircumferenceoftheopticdiscandtheuseofKalman lterestimatesthenextpointoflocationinthetrackingprocedure.
Imagedatabase
WehavetestedandevaluatedthetechniqueusedinthismethodbymakinguseoftheretinalcolorimagesavailableonDRIVEdatabase[25].Thisdatabasehasbeenwidelyusedasgroundtruthforperformanceevaluationbyotherresearcherstotesttheirves-selsegmentationmethodologies.TheretinalimagesavailableinthisdatabasewerecapturedusingaCanonCR5nonmydriatic3CCDcamerawitha45 Field-of-View(FOV),indigitalform.TheDRIVEdatabaseconsistsoffortyeye-funduscolorimageswhicharedividedintotrainingandtestset,eachsetcontainingtwentyimagesoutofwhichthetrainingsethasthreeimageswithpathologyandthetestsethasfourimageswithpathology.Theimagesavailableinthetrainingsetweresegmentedonlyonceandtheimagesinthetestsetweresegmentedtwice,resultingintwosets,the rstsetisacceptedasthegroundtruthandusedforalgorithmperfor-manceevaluationinliterature.Thecompleteresultsofthemanualsegmentationarealsoavailableforalltheimagesofthetwosets.Thecolorimagesoftheretinaavailableinthedatabaseareofsize768×584pixelsand8bitspercolorchannel,andwereinitiallysavedinJPEGformat.Moreover,thedatabaseincludesmaskswiththedelimitationofaField-of-View(FOV)ofapproximately540pix-elsindiameterforeachimage,whichisusedfortheperformanceevaluationofourmethod.
Methodology
ThispaperproposesaretinalvesselsegmentationtechniqueusingArti cialNeuralNetworks(ANN).Themainfeaturesoffundusretinalimagesarebloodvessels,whichareusedtodetectthepres-enceofanomaliespresentinhumanretinaautomatically.Theseretinalbloodvesselsareidenti edbyusingamultilayerperceptronneuralnetwork,forwhichtheinputsarederivedfromGaborandmomentinvariants-basedfeatures.Usingneuralnetworks,someofthefeaturessuchashemorrhagesandexudatescanbedetected,whichhelpstoexaminewhetherretinopathyispresentornot.Theimageswithretinopathyalsorequireagradingmeasure,whichisacomplextaskthanidentifyingthepresenceofretinopathy.Itcanbeexaminedbyatrainedperson,becausethecomputerbasedanalysisdetectsonlythechangessuchasneovascularisation,cottonwoolspots,perifovealexudationandvascularchanges[24].Thebasicsteptoachievethisdiagnosisprocedureistolocatetheimpor-tantregionsofthefundus,suchasbloodvessels,automatically.Thedatafromthesesegmentedbloodvesselregionscanfurtherbeprocessedandanalyzedforexaminingthefeaturesofcertaineyediseases.Inthispaper,computerbasedimageanalysisusingneuralnetworkisappliedtodetectandsegmentthebloodvesselsfromretinalimages.
Preprocessing
Preprocessingofimagesisthe rststepcarriedoutbeforepre-sentingtheinputtotheneuralnetwork.TheretinalcolorimagestakenfromtheDRIVEdatabasearepreprocessed,whereeachpixelcontainedthreevalues-red,greenandblue.AfewexamplesgivenasinputimagestothemultilayerperceptronneuralnetworkcanbeseeninFig.1.
TheRGBfeaturesareextractedforthepreprocessedimage,itsmeaniscalculatedandthegrayscaleimageisobtainedforfur-therprocessing.ThetruecolorimageRGBisconvertedtothegrayscaleintensityimagetoevaluateitsadequateabilityforthesegmentationoftheretinalbloodvessels.Theconversionisdonebyeliminatingthehueandsaturationinformation,whileretainingtheluminance[26].Thegiveninputimagesarefurtherprocessedusingmean lteringandafewexamplesareshowninFig.2.
Inthenextstep,theretinalimagesarefurtherprocessedandsmoothenedusingGaussian ltersandafewexamplescanbeseeninFig.3.
ThentheGaborfeaturesareextractedatdifferentorientationsandafewexamplesareshowninFig.4.Gabor lteringisdoneand

Fig.1.Inputimagesgiventomultilayerperceptronneuralnetwork.
ICPR06_A new approach to automated retinal vessel segmentation using ...Gabor filters and scale production, and a threshold probing technique ...
热门文档