Enhancing Pediatric Wrist Fracture Detection: A Two-Stage YOLOv8 Approach with Class-Specific Data Augmentation

Detection of pediatric wrist fractures is equally plagued by class imbalance and intricacy of some fracture types. We propose a method to counter this challenge through using class-specific data augmentation and a focused training strategy. From a subset of the original dataset, we selectively augmented five underrepresented and complex classes—Bone Anomaly, Bone Lesion, Pronator Sign, Soft Tissue, and Periosteal Reaction—using techniques such as rotation, zoom-out, and brightness adjustments. This selective augmentation allowed the class to be balance and enhanced feature clarity, particularly for classes with limited or complex data. Moreover, by isolating each class during augmentation, the model could focus on learning intricate features more effectively. Our two-stage training process first trained the model on these five challenging classes, then fine-tuned it on the full eight-class dataset. Trained for both YOLOv8 architectures (YOLOv8n and YOLOv8m), our approach achieved a state-of-the-art performance, with the YOLOv8m model attaining an mAP50 of 0.928 and an mAP50-95 of 0.823 on the original validation set. Notably, underrepresented classes as Bone Lesion reached precision of 0.946 and recall of 0.933 while with mAP 50 of 0.923 for Periosteal Reaction. This work highlights class-specific augmentation and focused training's effectiveness in addressing dataset imbalances and enhancing performance. It provides a reliable and complete solution for detecting wrist fractures in a pediatric setting, ideally suited for any real-world application in medicine.