Development and validation of an enhanced dental caries detection algorithm based on multi-scale feature mixing and adaptive pyramid network: a technical development study

Background: Dental caries affects 75% of the global population, yet traditional visual examination suffers from high subjectivity and difficulty in early detection. This study developed an enhanced deep learning algorithm to improve automated caries detection accuracy in oral photographs. Methods: We designed MSFAPN\_CDA incorporating multi-scale feature fusion and adaptive pyramid networks. The algorithm was trained and validated on 2,000 oral photographs from two institutions. Performance was evaluated using precision, recall, mAP@0.5, and compared against state-of-the-art detection algorithms through statistical analysis. Results: MSFAPN_CDA achieved 94.7% mAP@0.5, 87.0% precision, and 88.7% recall, representing a 7.9 percentage point improvement over baseline YOLO11. The algorithm significantly outperformed YOLOv5 (89.2%), YOLOv8 (88.7%), and RT-DETR (82.2%) with 9.86ms inference time. Ablation studies confirmed each component's contribution (p<0.001). Conclusions: MSFAPN_CDA demonstrates superior performance for automated caries detection with real-time capability, providing valuable technical support for early diagnosis and oral health management in clinical practice.