Deep learning-based identification of periodontal infrabony defects with regenerative potential: A multicenter retrospective study.

BACKGROUND

Current radiographic evaluation of periodontal infrabony defects is subjective and prone to variability, limiting evidence-based regenerative surgery planning. This study aimed to develop a dual-model deep learning framework integrating YOLO V8 and nnU-Net V2 to automate defect identification, quantification, and surgical guidance on parallel intraoral radiographs (PTRs).

METHODS

A multicenter retrospective study utilized 580 PTRs from three institutions. The model combined YOLO V8 (tooth instance segmentation) and nnU-Net V2 (tissue semantic segmentation) to localize anatomical landmarks, such as the cemento-enamel junction (CEJ), lowest point of the defect (LP), alveolar crest (ALC), and root apex (APEX), and to calculate defect parameters (depth, width, angle). An internal dataset (n = 387) derived from Shanghai Stomatological Hospital was used for model training and cross-validation, while an external dataset (n = 93) from two independent institutions (Shanghai Ninth People's Hospital and Tongji University) served for generalizability testing. The datasets were evaluated via 5-fold cross-validation. Performance metrics included Dice coefficient, mean absolute error (MAE), sensitivity, specificity, and F1-score.

RESULTS

The model achieved high precision in landmark localization (MAE: 0.20-0.54 mm) and strong generalizability, with external specificity of 0.96 and accuracy of 0.89. It outperformed specialists in diagnostic speed (1.5 vs. 2.1 min/patient) and accuracy (96% vs. 92%). Postoperative analysis demonstrated significant defect resolution. Defects were categorized via a color-coded system (e.g., depth ≥3 mm recommended for regeneration), aligning with clinical guidelines.

CONCLUSIONS

This study pioneers a dual-model AI framework for infrabony defect analysis, offering precise, guideline-aligned surgical recommendations via a color-coded system. By standardizing defect assessment and enhancing diagnostic efficiency, the model bridges radiographic interpretation and clinical decision-making, advancing personalized periodontal care.

PLAIN LANGUAGE SUMMARY: A dual-model AI framework combining YOLO V8 and nnU-Net V2 accurately identifies and quantifies periodontal infrabony defects on radiographs, outperforming specialists in speed and diagnostic accuracy.