SlimDy-YOLO: A Lightweight Dynamic Detector for Pediatric Wrist Fracture Detection in X-ray Images

In pediatric orthopedics, timely and accurate detection of wrist fractures is essential for effective treatment. However, manual interpretation of X-ray radiographs remains difficult due to growth plates that resemble fracture lines, occult fractures with subtle signs, and the need for fast inference in resource-limited clinical settings. To address these challenges, we propose SlimDy-YOLO, a lightweight dynamic YOLOv11-based detection framework. The framework introduces D-HGNet, a dynamic backbone that combines an efficient PPHGNetV2-inspired design with dynamic convolution for improved fine-grained feature extraction. This improves discrimination between actual fractures and growth plates and increases sensitivity to occult fractures with subtle signs. To further improve efficiency, SlimDy-YOLO adopts Slim-Neck to reduce model complexity during multi-scale feature fusion, and incorporates DySample to better preserve subtle fracture cues during upsampling. On the GRAZPEDWRI-DX dataset, SlimDy-YOLO achieves 68.47\% mAP$_{50}$ with only 40.5 GFLOPs and 16.83M parameters, while achieving 265.96 FPS for real-time inference. These results demonstrate that SlimDy-YOLO outperforms representative real-time baselines, achieving a superior balance between accuracy and efficiency, thus offering a promising solution for automated screening in resource-constrained clinical settings.