COMPREHENSIVE DETECTION OF MULTI-TYPE WRIST FRACTURES USING IMPROVED YOLOv8 MODEL

Abstract

Bone fractures, particularly those affecting the wrists, shoulders, and arms, are common and significantly impact patient care. This study investigates the utility of YOLOv8, a deep learning model, in detecting ulna and radius fractures which are the crucial components of wrist injuries. Through collaboration with phys- iotherapists and comprehensive data collection, an annotated dataset is curated for precise wrist fracture localization. Advanced data augmentation techniques are used including Mosaic, Mix-up, and copy paste, to enhance dataset diver- sity and model robustness. Along with baseline YOLOv8 and iYOLOv8 + GC

which currently reports high precision (97.2%), we also consider the most opti- mized variant. We also focus on the baseline YOLOv8 results relative to this newer version to demonstrate the focus on high precision vs the balance within recall, F1 score, and precision. The model excelled in detecting various wrist fracture types, advancing fracture detection in medical practice. The improved evaluation metrics, including accuracy, precision, recall, and F1-score, highlighted the robustness of the YOLOv8 model in identifying ulna and radius fractures. YOLOv8 achieved high scores with accuracy and precision score of 0.87 each recall 0.88, and F1-score 0.86, which indicates its proficiency in accurate fracture detection. The comparative analysis has highlighted the balanced performance of YOLOv8, as the best models like iYOLOv8 + GC also record higher values of accuracy (up to 97.2) whereas our strategy is robust in various measures. These findings highlight YOLOv8 as a promising diagnostic tool for expedited diagnosis and improved patient care in wrist injury management.