<![CDATA[Pediatric wrist fractures are common in emergency care and require specialized expertise for accurate radiographic interpretation. This study developed and evaluated a fracture-only detector for pediatric wrist radiographs using YOLOv8 and assessed its external performance. The public GRAZPEDWRI-DX dataset, comprising 20,327 images from 6,091 patients, was used, with 18,090 fracture bounding boxes retained and non-fracture images included as background. Images were normalized, enhanced using contrast limited adaptive histogram equalization (CLAHE), and resized to 640×640 pixels. Data were split by patient identity into training, validation, and test sets at a 70/20/10 ratio to prevent patient overlap. A pretrained YOLOv8s model was trained with early stopping. On the held-out test set of 1,918 images from 610 patients, the model achieved a precision of 0.9236, recall of 0.8607, mAP50 of 0.9275, and mAP50-95 of 0.5374, with no evidence of validation overfitting. However, zero-shot evaluation on the FracAtlas dataset showed very low performance, including an mAP50 of 0.0343 overall and 0.0336 on the hand/wrist subset. CLAHE matching and lower IoU thresholds produced minimal improvement. These findings indicate strong internal performance but poor external generalization. The study is limited by a single run, single architecture, single-institution dataset, and absence of comparison models or clinical validation.]]>
Copyrights © 2026
