64 - Performance of Artificial Intelligence in Detecting Dental Features in Pediatric Bitewing Radiographs

Background: Artificial intelligence (AI) offers potential for improving detection performance in pediatric dental radiographs, where primary tooth anatomy present unique challenges. This study developed and evaluated the performance of a deep learning model for detecting dental caries, restorations, and appliances in bitewing radiographs of primary teeth.

Methods: This retrospective detection study utilized 441 pediatric bitewing radiographs. An annotation protocol established ground truth labels for 18 classes including enamel caries, dentin caries, restorations, stainless steel crowns, lamina dura status, space maintainers, and tooth identification using Universal numbering System A-T. Images underwent standardized preprocessing including Contrast Limited Adaptive Histogram Equalization enhancement, and were divided into training (70%), validation (15%), and test (15%) sets. You Only Live Once version 8 (YOLOv8-nano) was selected following evaluation of multiple architectures and trained using AdamW optimizer over 150 epochs with conservative data augmentation. Performance was evaluated using mean Average Precision (mAP), precision, and recall.

Results: The model achieved 97.4% mAP@0.5, 95.5% precision, 95.5% recall, and 95.0% F1-score. Tooth identification as a structured labeling task achieved (96.2-100% mAP@0.5), stainless steel crowns (100% mAP@0.5), space maintainers (92.7% mAP@0.5), restorations (100% mAP@0.5), and non-intact lamina dura (mAP@0.5=93.8%, F1-score: 92.0%) demonstrated excellent performance. Dentin caries achieved 91.0% mAP@0.5 (F1-score: 89.2%), while the model primary performance limitation was observed in the detection of enamel caries, which yielded substantially lower metrics (Precision=66.7%, Recall=43.8%, mAP@0.5=53.9%, F1-score: 52.8%).

Conclusion: The YOLOv8-nano model demonstrated high detection performance for pediatric dental feature detection, with performance exceeding previously reported values. Early-stage enamel caries remains challenging for AI model detection, warranting further investigation.