Evaluation of Deep Learning for Caries Detection With Fine-Grained Classification and Postprocessing Improvements
Loading...
Date
Authors
Publisher
Elsevier
Abstract
Objectives:Deeplearningmethodshavebeenproventobeeffectiveindetectingdentalcar-iesinvisiblelightimages.However,existingresearchinvolvesinadequatecategoriesandmainlyfocusesonlocallesionareas.Thisstudyaimstouseadvanceddeeplearningmod-elstoachievecariesdetectionbasedontoothinstances(whereallteethinimagesaredetected)andfine-grainedclassificationaccordingtotheInternationalCariesDetectionandAssessmentSystem(ICDAS).Toaddressthepotentialinstabilityundercomplexsce-narios,wepropose2correctionmethodsthatincorporatebackgroundknowledge.Methods:Atotalof1200selectedhigh-qualityintraoralimageswereexpandedto8,754imagesusingdataaugmentationtechniques,andeachtoothinsidewasannotated.Threeadvancedmodels,YOLO-v8,YOLO-v9,andYOLO-NAS,weretrainedandtestedonthedata-set.Inthestageofpostprocessing,predictedcategorieswerecorrectedwithaweightedaverageofscores,andconfidencescoreswereadaptivelyadjustedbasedonthespatialrelationshipsofteeth.Results:TheproposedmethodsimprovedthemeanAveragePrecision(mAP)scoresby4.7%(p<.01/Mann-Whitney-U-test),2.8%(p<.01),and4.4%(p<.01)acrossthe3models,withthehighestscoreof72.9%onYOLO-v8.Precisionandrecallincreasedby3.8%and5.6%,respectively,whileFPSdecreasedfrom83.1to78.1.Especiallyimprovedthescoresformoderatecariesanddemonstratedgreaterrobustness.Conclusion:Theprimaryobjectiveswereachieved,andthe2proposedcorrectionmethodsbringaneffectiveimprovementtotheexistingalgorithmframework.It’sexpectedtopro-motetheapplicationofartificialintelligenceandinspirefurtherresearch.ClinicalRelevance:Thisresearchisofclinicalvalueduetoitsfunctionalinnovation:thefinerclassificationshouldassistdentistsformulatepersonalizedtreatmentstrategies.Focusingonthedetailedevaluationofeachtoothshouldhelpdeliverbetterandpersonal-isedclinicalcare.
Metadata
License
Except where otherwise noted, this item's license is described as Attribution 4.0 International