Detection and Classification of Acute Lymphoblastic Leukemia Utilizing Deep Transfer Learning

A mutationintheDNAofasinglecellthatcompromisesitsfunctioninitiatesleukemia.Thisleadstothe overproductionofimmaturewhitebloodcells,whichencroachuponthespacerequiredforthegeneration of healthybloodcells.Leukemiaistreatableifidentifiedinitsinitialstages.Nonetheless,itsdiagnosisis both arduousandtime-consuming.Inthisstudy,anovelapproachfordiagnosingleukemiaacrossfour stages—Benign,Early,Pre,andPro—utilizingdeeplearningtechniques.Firstly,wehademployedtwo ConvolutionalNeuralNetwork(CNN)models:MobileNetV2withanalteredheadandabespokemodel. The custommodelhasmultipleconvolutionallayers,eachpairedwithcorrespondingmaxpoolinglayers. Nowwehaveemployedtwomorepretrainedmodel:InceptionV2andVGG16withalteredheads.We utilized MobileNetV2,InceptionV2andVGG16withImageNetweights,andtheheadwasadjustedto integratethefinalresults.Finally,wehaveutilizedensemblewithsoftvotingtechniquetocomprehendthe results frommultipleneuralnetworks.Theutilizeddatasetisapubliclyavailablecollectionofbloodcell smear imagestitled“AcuteLymphoblasticLeukemia(ALL)imagedataset”,andthenusedtheSynthetic MinorityOversamplingTechnique(SMOTE)toaugmentandbalancethetrainingdataset.Whichattained an accuracyof96.34%withthecustommodel,whileMobileNetV2andInceptionV2achievedasuperior accuracy of99.39%.TheVGG16achievedanaccuracyof99.08%.Finally,theensembletechnique ensured morepromisingresultwith99.70%accuracy.Thepre-trainedmodelexhibitedencouraging results andanincreasedlikelihoodofreal-worldapplication.