An Explainable ResNet-101 Framework for Elbow Fracture Classification in X-ray Images
Keywords:
Elbow Fracture, Deep Learning, ResNet-101, Grad-CAM++, Medical Imaging, Explainable AI, MURA DatasetAbstract
Fracture detection in elbow radiographs remains challenging due to subtle bone irregularities and anatomical variability. This study proposes an explainable deep learning framework for automated elbow fracture detection using a ResNet-101 architecture. The model was trained and evaluated on the elbow subset of the publicly available MURA dataset, with preprocessing and data balancing techniques applied to mitigate class imbalance and enhance generalization. Bayesian hyperparameter optimization was employed to identify optimal training parameters, including learning rate, dropout rate, and network capacity, thereby improving model stability and performance. The optimized model achieved an accuracy of 77.96% with an area under the curve (AUC) of 0.81, demonstrating reliable classification performance. To enhance interpretability, Grad-CAM++ was integrated to generate heatmaps highlighting regions influencing the model’s predictions, enabling verification of clinically relevant areas. The results indicate that the model effectively identifies fracture regions while maintaining consistent performance across diverse radiographic samples. In addition, the system generates automated PDF-based clinical reports that combine predictions with corresponding heatmaps, improving its practical applicability in clinical settings. Overall, the proposed approach provides a supportive, interpretable, and clinically relevant solution for elbow fracture detection, with strong potential for integration into computer-aided diagnostic systems, while future work will focus on extending the model to multi-centre datasets and incorporating additional clinical features to further enhance generalization and diagnostic reliability.
