Pediatric Wrist Fracture Detection in Radiographs Using the YOLOv11n Object Detection Model

Abstract

Fracture detection in pediatric wrist radiographs is challenging due to incomplete skeletal ossification, small bone structures, and subtle hairline (non-displaced) fractures that can be difficult to identify, while growth-plate (physeal) radiolucency often mimics fracture appearance. This study evaluates YOLOv11n, a lightweight one-stage object detector that incorporates multi-scale feature extraction components (e.g., Spatial Pyramid Pooling–Fast, SPPF), for automated pediatric wrist fracture detection and localization. The model was trained and evaluated on the GRAZPEDWRI-DX benchmark dataset comprising 20,327 pediatric wrist radiographs (14,269 training, 4,048 validation, 2,010 test images) using transfer learning with the Ultralytics training pipeline and default online augmentation strategies. YOLOv11n achieved mAP@50 = 0.936 on the validation set and mAP@50 = 0.940 on the test set, with precision = 0.923 and recall = 0.850 on validation and precision = 0.926 and recall = 0.870 on test. Runtime profiling on an NVIDIA Tesla T4 GPU indicated end-to-end per-image latency below 5 ms, supporting near-real-time clinical decision-support workflows. The mAP@50–95 values (0.564 on validation and 0.552 on test) indicate reduced localization tightness under stricter IoU criteria, consistent with the greater difficulty of precisely localizing subtle fracture regions. Overall, YOLOv11n provides a favorable balance between detection performance and computational efficiency for pediatric wrist fracture detection; however, external multi- institutional validation and targeted strategies (e.g., multi-view fusion and pediatric anatomy- aware modeling) are recommended before clinical deployment to improve sensitivity to subtle fractures and reduce growth-plate-related false positives.