Feasibility of multi-class dental caries detection using deep learning–based smartphone images: a pilot prospective study

Aim: This pilot study evaluated the feasibility of multi-class dental caries detection using deep learning-based smartphone images. Method: Seventy adults provided six self-taken intraoral smartphone images in a non-clinical setting. Four dentists annotated each image based on the visual assessment of caries using binary (none vs. caries), four-class (none, initial, moderate, or advanced), and five-class (none, initial-1, initial-2, moderate, or advanced), according to two internationally recognized criteria (ICDAS-II and ADA CCS). A YOLOv6-L6 deep learning model was trained on an expert-labeled dataset, with model performance evaluated using sensitivity, specificity, precision, accuracy, F1-score, and mean average precision (mAP). Results: Intra-examiner reliability was substantial to almost perfect (κ = 0.765-0.952), whereas inter-examiner agreement was low to moderate (κ = 0.358-0.406). Across all classification schemes, the models achieved a sensitivity ≥86%, a specificity >97%, a precision >90%, and an accuracy >93%. F1-scores remained high (>88%), declining as the classes increased. The binary classification model demonstrated the highest mAP (65.2%), whereas the four- and five-class models showed progressively lower mAP values. Notably, the five-class model achieved the highest specificity (98.45%). Conclusion: This pilot study demonstrates the feasibility of smartphone-based deep learning as a supportive aid for caries detection and its potential role in patient-driven screening.