Abstract
BACKGROUND/PURPOSE: Traditional caries detection relies on visual and radiographic analysis. While deep learning has been applied to classify caries extent, no studies classify caries depth using radiomic features in intraoral photographic images. This study evaluated a radiomics-based approach with machine learning (ML) to classify caries extent and depth, traditionally assessed via radiographs, using intraoral photographs. MATERIALS AND METHODS: Standardized intraoral photographs were taken with a Nikon D7500 and Macro Flash MF-R76. Only images of healthy teeth or carious lesions were included. Images were resized, segmented with Labelme, and classified using ICDAS and E-D scales. Data augmentation increased sample size. Radiomic features were extracted for each color channel using Pyradiomics. Feature selection methods (AUC-ROC, ReliefF, LASSO, backward selection) were applied within 5-fold cross-validation to prevent bias. ML classifiers (LDA, k-NN, SVM, NNET) evaluated accuracy, sensitivity, and specificity. Model explainability assessed feature influence via partial dependence plots, residual analysis, and break-down profile. RESULTS: NNET with backward selection achieved high accuracy (87.6%-95.4%). Sensitivity and specificity ranged from 61.5% to 93% and 73%-90%, respectively. Green and red channels significantly impacted predictions, with texture features being critical. The red channel's greater impact reflects its ability to mimic near-infrared light transillumination, enhancing contrast between healthy and decayed tissue. The blue channel had lesser influence, but combined RGB channels yielded the best accuracy. CONCLUSION: Radiomics enables caries depth classification from intraoral photographs, offering a non-invasive, cost-effective alternative to radiographs. This approach could revolutionize diagnostics by reducing reliance on invasive radiological techniques, using accessible and affordable equipment.