Comparative deep learning analysis of pulp stone detection in restored and healthy teeth on panoramic radiographs

Abstract The aim of this study was to comparatively evaluate the performance of deep learning–based artificial intelligence (AI) models trained on panoramic radiographs in detecting pulp stone (PS) in restored and healthy teeth. In this study, 1,119 panoramic radiographs containing PS, acquired between January 2022 and June 2025, were used. PS on the panoramic radiographs was annotated using the LabelMe software. A total of 4,230 PS labels were included in the dataset. These labels were classified into two groups according to the tooth condition: healthy teeth ( n = 3345) and restored teeth ( n = 885). For baseline performance comparison, YOLOv8 and YOLO11 deep learning architectures were used. In addition, a YOLO–GNN hybrid model was developed to reduce the performance loss observed in restored teeth. Model performance was evaluated on the test dataset using the mAP₅₀ and mAP₅₀–₉₅ metrics. According to the results, PS detection performance was consistently higher in healthy teeth compared to restored teeth across all models. Across all evaluated YOLO11 variants, baseline mAP₅₀ values for the healthy class ranged from 79.1% to 81.5%, while restored-class performance ranged from 67.6% to 73.1%. Following YOLO-GNN hybrid training, restored-class mAP₅₀ improved across all three architectures. The largest gain was observed in YOLO11l, increasing from 67.6% to 72.0% (+ 4.4 pp), whereas YOLO11m achieved the highest absolute restored-class mAP₅₀ value of 76.3%, increasing from 73.1% (+ 3.2 pp). Relative to the original baseline, healthy-class performance was maintained or improved across the final YOLO-GNN configurations. These gains reflect the cumulative effect of increased resolution, fine-tuning strategy, and GNN integration. Ablation studies confirmed that the GNN module’s isolated contribution ranged from + 1.1 to + 1.8 pp in restored-class mAP₅₀, with statistically significant lesion-level differences observed for YOLO11m ( p = 0.001) and YOLO11l ( p < 0.001). This study demonstrated that deep learning–based AI models are capable and promising in detecting PS on panoramic radiographs. Detection performance was higher in healthy teeth than in restored teeth. YOLO11 outperformed YOLOv8, and the GNN-based hybrid model improved the detection performance in restored teeth. Further multicenter studies with larger datasets are required to enhance clinical generalizability.