Detection of dental restorations on panoramic radiographs using different deep learning methods and comparative evaluation of their performance

To evaluate and compare instance segmentation (YOLOv8, YOLO11, YOLO26, Mask R-CNN) and semantic segmentation (U-Net, Attention U-Net, DeepLabV3+) models for the detection and pixel-level delineation of dental restorations and periapical findings on panoramic radiographs. A retrospectively assembled dataset of 849 anonymized images was annotated into eight categories (caries, composite and amalgam restorations, crowns, root canal treatments, implants, implant-supported crowns, and periapical lesions). Models were trained on standardized 1024 × 1024 images using architecture-specific optimization and augmentation pipelines. Performance was assessed using precision, recall, F1-score, intersection over union (IoU), Dice coefficient, and area under the receiver operating characteristic curve (AUC). Inference latency was also benchmarked on standardized hardware. Single-stage instance segmentation models generally provided favorable accuracy. YOLOv8 achieved a mean IoU of 0.59, a mean Dice coefficient of 0.72, and a macro F1-score of 0.72. In comparison, the two-stage Mask R-CNN reached a mean IoU of 0.54 and a mean Dice of 0.68, while semantic models yielded comparatively lower overlap metrics (mean IoU: 0.36–0.42). Class-wise evaluation indicated that YOLO architectures provided precise delineations for high-contrast structures; for instance, YOLO26 and YOLOv8 achieved IoU scores of 0.91 and 0.87 for implants, respectively, compared to 0.74 for Mask R-CNN. Conversely, Mask R-CNN showed slightly better localization for complex findings like periapical lesions (IoU: 0.29 vs. YOLOv8’s 0.20). Regarding computational efficiency, semantic models were the fastest overall. Among the instance segmenters, YOLOv8 was the most efficient (mean latency: 69.22 ms, p95: 76.39 ms), demonstrating more stable processing times than Mask R-CNN (mean: 72.73 ms, p95: 89.31 ms). Modern single-stage detection frameworks, notably YOLOv8, demonstrate a robust balance of segmentation accuracy and computational efficiency. They present a viable and promising alternative to traditional two-stage and semantic segmentation models for identifying various dental findings. While optimized architectures like YOLOv8 show computational and spatial potential as diagnostic support tools, this is an internally validated, single-center retrospective study. External multicenter validation and prospective clinical testing are required before their actual applicability in routine clinical workflows can be demonstrated.