FractureNet: A Swin Transformer Model for Radiographic Bone Injury Classification

Bone fractures are among the most common musculoskeletal injuries, often requiring timely and accurate diagnosis to prevent long-term disability. Conventional radiographic interpretation is prone to variability and limited by clinician workload, highlighting the need for automated, reliable diagnostic tools. This study proposes FractureNet, a deep learning framework based on the Swin Transformer architecture, designed to classify ten types of bone fractures from radiographic images. The contribution of this research lies in developing a hierarchical and explainable transformer-based model that captures both local and global bone features while maintaining clinical interpretability. The proposed method integrates a structured preprocessing pipeline involving normalization, resizing, and augmentation, followed by Swin Transformer training with hierarchical shifted-window attention. The model's performance was evaluated using accuracy, precision, recall, specificity, and F1-score metrics, along with ROC and confusion matrix analyses. FractureNet achieved an overall accuracy of 90.1%, macro-average AUC of 0.826, and high specificity across all classes. The model performed best on comminuted and fracture-dislocation types, with visualization results from Grad-CAM and Score-CAM confirming that the model accurately focused on fracture regions. These findings validate the interpretability and diagnostic reliability of the proposed framework. In conclusion, FractureNet demonstrates strong performance and transparency, offering a promising foundation for AI-assisted radiographic diagnosis. Future work will enhance dataset diversity and model generalization for broader clinical deployment.