A Case Study on Bone Fracture Detection Using YOLOv۸–YOLOv۱۲ Models

Background and Aims: Bone fracture diagnosis and localization are crucial steps in the treatment process. Specialists rely on medical imaging to assess the affected area and determine the type of fracture. However, human limitations such as fatigue, inattention, and time constraints can compromise diagnostic accuracy, leading to potential misdiagnoses and incorrect treatment decisions. To address these challenges, this study explores the use of deep learning-based object detection models, particularly YOLO architectures, for automated bone fracture detection and localization. Methods: This study evaluates several YOLO-based models from version ۸ to ۱۲ on the FracAtlas dataset, a specialized dataset for bone fracture detection. To enhance model generalization, data augmentation techniques are applied. The models' performance is assessed using precision, recall, and mean Average Precision at IoU ۰.۵۰ (mAP@۵۰). Furthermore, the best-performing model is tested in an end-to-end detection and localization setting to evaluate its real-world applicability. Results: Experimental results indicate that YOLOv۸-S with augmentation achieves the best balance between precision (۰.۸۱۹۷), recall (۰.۶۷۱۶), and mAP@۵۰ (۰.۷۸۳۷), making it the most effective model for fracture localization. However, when applied in an end-to-end detection and localization scenario, performance declines, with precision, recall, and mAP@۵۰ dropping to ۰.۸۹۱۰, ۰.۴۱۷۹, and ۰.۶۳۳۸, respectively. This reduction is attributed to the increased complexity of distinguishing subtle fracture patterns from the non-fractured regions. Conclusion: The findings highlight the effectiveness of YOLO-based architectures in fracture detection while underscoring the challenges posed by real-world medical imaging conditions. Future work should focus on expanding dataset diversity, improving augmentation strategies, and optimizing model architectures to enhance detection accuracy, particularly for subtle fractures.

کلیدواژه ها:

Artificial_Intelligence ، Medical_Imaging ، Bone_fracture_detection ، Augmentation ، YOLOV۸-S