Beyond dental radiographs, a radiomics-based study for the classification of caries extension and depth

Background/purpose:Traditionalcariesdetectionreliesonvisualandradiographic analysis.Whiledeeplearninghasbeenappliedtoclassifycariesextent,nostudiesclassify cariesdepthusingradiomicfeatures inintraoralphotographicimages.Thisstudyevaluated aradiomics-basedapproachwithmachinelearning(ML)toclassifycariesextentanddepth, traditionallyassessedviaradiographs,usingintraoralphotographs. Materialsandmethods:StandardizedintraoralphotographsweretakenwithaNikonD7500 andMacroFlashMF-R76.Onlyimagesofhealthyteethorcariouslesionswereincluded.Images wereresized, segmentedwithLabelme,andclassifiedusing ICDASandE-Dscales.Data augmentationincreasedsamplesize.Radiomicfeatureswereextractedforeachcolorchannel usingPyradiomics.Featureselectionmethods(AUC-ROC,ReliefF,LASSO,backwardselection) wereappliedwithin5-foldcross-validationtopreventbias.MLclassifiers(LDA,k-NN,SVM, NNET)evaluatedaccuracy,sensitivity,andspecificity.Modelexplainabilityassessedfeature influenceviapartialdependenceplots,residualanalysis,andbreakedownprofile. Results:NNETwithbackwardselectionachievedhighaccuracy(87.6é5.4%).Sensitivityand specificityrangedfrom61.5%to93%and73é0%,respectively.Greenandredchannels