AI-Driven Caries Scoring In Rats Via Modified Keyes System

Despite the widespread use of Keyes scoring in rat caries research, its manual execution remains subjective and time-intensive (Keyes PH. 1958. Dental Caries in the Molar Teeth of Rats: II. a Method for Diagnosing and Scoring Several Types of Lesions Simultaneously. J Dent Res. 37(6):1088-1099). This study investigates the feasibility of integrating convolutional neural networks (CNNs) with a refined Keyes protocol to automate caries assessment. Following ethical approval, caries were induced in 24 female Sprague-Dawley rats. Mandibular molars were sectioned sagittally, stained with ammonium urethane, and imaged via stereo microscopy. A modified Keyes scoring system was applied to 286 high-quality images, categorizing lesions into no/enamel caries (nE), mild (Es), moderate (Em), and severe dentin caries (Ex). A customized ResNet-18 model was implemented in PyTorch, utilizing multi-label classification to independently score three molar regions. Training spanned 300 epochs with Adam optimization (learning rate 0.001), and performance was validated against manual annotations. The model achieved an overall accuracy of 94%, with the third molars demonstrating the highest precision (96.6%) compared to first (93.3%) and second molars (95.9%). Moderate dentin caries (Em) showed exceptional recognition rates (97.5%), while severe caries (Ex) exhibited reduced accuracy (87.4–93.0%). AUC analysis revealed superior diagnostic consistency for first molars (88%) over second (83%) and third molars (80%). The AI-enhanced framework demonstrated excellent diagnostic accuracy while reducing evaluator variability, offering scalable rat caries analysis.