A REFINED YOLOV8 BASED DEEP LEARNING FRAMEWORK FOR AUTOMATED DETECTION AND DIFFERENTIATION OF WRIST FRACTURE TYPES
Abstract
Bone fractures are among the most common musculoskeletal injuries, particularly those affecting the wrist, shoulder, and arm, and their timely identification is essential for effective treatment and patient recovery. This study explores the application of YOLOv8, a deep learning-based object detection model, for identifying fractures in the radius and ulna, the two bones that play a vital role in forearm and wrist function. To support reliable fracture detection, an annotated dataset was developed through systematic data collection in collaboration with physiotherapists, with particular emphasis on accurately locating fracture regions in medical images. Several data augmentation techniques, including Mosaic, MixUp, and Copy-Paste, were incorporated into the training process to increase data variability, address potential limitations in the dataset, and improve the model’s ability to recognize different fracture patterns. These techniques also helped expose the model to a broader range of image variations, potentially improving its generalization to previously unseen cases. The experimental analysis included the baseline YOLOv8 model, the improved iYOLOv8 + GC architecture, and optimized model configurations to examine their effectiveness in detecting different types of wrist fractures. Particular attention was given to comparing precision, recall, and F1-score to understand the strengths and limitations of each approach rather than relying on a single performance indicator. The baseline YOLOv8 model achieved an accuracy of 0.87, a precision of 0.87, a recall of 0.88, and an F1-score of 0.86. These results indicate that the model can identify fracture-related patterns while maintaining a relatively balanced relationship between precision and recall. In comparison, the iYOLOv8 + GC variant reportedly achieved a precision of 97.2%, suggesting a strong ability to correctly identify predicted fracture cases.
However, a higher precision value alone does not necessarily indicate better overall performance, making it important to consider recall, F1-score, and other relevant evaluation measures when comparing the models. The findings therefore demonstrate the importance of evaluating both conventional and enhanced YOLOv8 architectures to determine their suitability for automated fracture detection. Furthermore, the use of deep learning can potentially reduce the time required for preliminary image assessment and help clinicians identify suspicious fracture regions more efficiently. Such automated assistance may be particularly valuable in clinical environments where large numbers of radiographic images require examination. Nevertheless, the model should be considered a potential decision-support tool rather than a replacement for qualified medical professionals. Further evaluation using larger, independent, and clinically diverse datasets is necessary to establish its reliability across different patient populations and imaging conditions. Overall, this research highlights the potential of YOLOv8-based approaches to improve the efficiency and consistency of radius and ulna fracture detection, providing a foundation for future developments in AI-assisted medical image analysis and wrist injury management.
Keywords: Wrist injuries; ulna and radius fractures; YOLOv8; deep learning; fracture detection.
