Deep learning-based classification of dental caries from intraoral photographs using ICDAS criteria

Dental caries remains a major global health burden, yet current artificial intelligence tools are often constrained by limited classification systems or rely on specialized imaging equipment unsuitable for self-examination. This study aims to develop and validate a deep learning-based approach using intraoral photographs to detect and classify dental caries using the full International Caries Detection and Assessment System (ICDAS) alongside modified ICDAS, merged ICDAS, and binary classification schemes. ICDAS scores were assigned through direct clinical examination. Two annotators independently examined each patient and compared their results. Disagreements were first resolved by joint re-examination. If consensus could not be reached, a third senior clinician examined the patient and made the final determination. A dataset of 1606 intraoral photographs was collected from volunteers aged five years and older across multiple clinics in China. The dataset was split into training, validation, and test sets at an 8:1:1 ratio at the patient level. The YOLOv12x and YOLOv8x architectures were trained and tested, with performance evaluated in terms of precision, recall, and mean average precision at 50% IoU (mAP50) with 95% confidence intervals. Gaussian blur was systematically applied to the test set at seven σ levels (0 to 11) to evaluate robustness. YOLOv12x significantly outperformed YOLOv8x in full ICDAS (mAP50 0.762 [0.755–0.769] vs. 0.736 [0.733–0.740]; F1 0.715 ± 0.013 vs. 0.693 ± 0.007; p < 0.0001), modified ICDAS (mAP50 0.785 [0.780–0.790] vs. 0.764 [0.759–0.770]; F1 0.731 ± 0.011 vs. 0.718 ± 0.011; p < 0.0001), and merged ICDAS (mAP50 0.897 [0.893–0.901] vs. 0.888 [0.885–0.891]; F1 0.835 ± 0.007 vs. 0.821 ± 0.007; p = 0.0012) schemes, but not in the binary task (mAP50 0.938 [0.935–0.941] vs. 0.937 [0.935–0.940]; F1 0.885 ± 0.003 vs. 0.888 ± 0.007; p = 0.7807). Codes 0 and 6 achieved the highest accuracy, whereas codes 2–5 showed frequent adjacent-stage confusion, particularly between codes 3 and 4. Under Gaussian blur, mAP50 declined from 0.776 to 0.498 at σ = 11, with extreme classes proving most resilient. Metallic restorations and implants were often misclassified as advanced caries. The attention mechanism in YOLOv12x improves fine-grained ICDAS classification but does not help in binary detection, where both models approach ceiling performance. Discriminating adjacent early-stage lesions is substantially harder than detecting caries. The model is better suited to screening and triage than to standalone diagnosis. Future work should incorporate data from additional regions and imaging equipment, and validate performance in uncurated clinical settings.