Development and Internal Evaluation of a Deep Learning Model for Fine-Grained Detection of Dental Trauma on CBCT-Dominant Image Records.

Objectives

To develop and internally evaluate a YOLO-based model for fine-grained localization and classification of dental trauma subtypes.

Methods

This retrospective single-center study used a CBCT-dominant dataset of 1,256 annotated trauma instances (1,065 instances derived from sagittal CBCT images and 191 instances derived from periapical radiographs) to develop and internally evaluate a YOLO26x object detection model. Performance was assessed using average precision (AP), mAP@0.5, mAP@0.5:0.95, precision, recall, and inference speed.

Results

AP was 90.4% for complicated crown-root fracture, 88.2% for uncomplicated crown fracture, and 83.9% for root fracture, but 27.3% for alveolar fracture and 48.9% for uncomplicated crown-root fracture. Overall mAP@0.5, mAP@0.5:0.95, precision, and recall were 69.9%, 42.5%, 0.675, and 0.704, respectively; mean inference time was 3.3 ms per image.

Conclusions

The YOLO26x model showed subtype-specific performance in fine-grained detection of dental trauma, with better performance for injuries with clearer radiographic boundaries and lower performance for alveolar fracture and uncomplicated crown-root fracture. These internal findings support preliminary feasibility and warrant multicenter validation with patient-level separation. CLINICAL

Significance

If validated in multicenter studies with patient-level separation, AI-assisted image analysis could support lesion localization and subtype recognition when clinically indicated imaging is available; it should complement, rather than replace, comprehensive clinical assessment.