Dental Caries Type and Severity Detection: A Comparative Study of YOLOv10 and YOLOv11

Background Accurate diagnosis of dental caries on radiographs is often tricky and varies among clinicians. Advances in deep learning have made automated detection feasible, but few studies have examined detailed classification across multiple stages of the disease. Methods We compared YOLOv10 and YOLOv11 for detecting and classifying caries on bitewing radiographs. A dataset of 1,187 cases was labeled into ten clinically defined categories, covering both severity (sound, mild, moderate, severe) and lesion type (primary, recurrent, crown-related, occlusal). Models were trained with standardized parameters, preprocessing, and augmentation, and performance was evaluated using precision, recall, F1-score, and mean Average Precision at IoU thresholds of 0.5 and 0.5–0.95. Results Both models achieved high overall accuracy. YOLOv10 provided the most consistent performance, reaching a mAP50 of 98.9, mAP50-95 of 98.8%, Precision of 97.6%, Recall of 97% and F1-score of 97.3%. It was especially effective at detecting mild and moderate lesions. YOLOv11, although less stable overall (mAP50–95 of 80.4%), achieved perfect recall for crown-recurrent and occlusal-recurrent lesions, categories that are rare but clinically important. Conclusion YOLOv10 is better suited for routine diagnostic use due to its balanced accuracy and stability, while YOLOv11 maximizes sensitivity for uncommon caries types, although it can also detect common caries well. Compared to earlier YOLO and CNN-based studies, these models achieved higher accuracy and covered a wider range of clinically meaningful categories. Fine-grained, real-time caries classification is therefore achievable and has clear potential to aid decision-making in preventive and restorative dentistry.