Explainable Deep Learning for Shoulder Fracture Detection: EfficientNet-Based Evaluation on the MURA Dataset

Accurate shoulder bone fracture diagnosis allows timely planning for treatment. Although some research explored deep learning with musculoskeletal lesions, their reliability lacks adequate explanation methods to be practically applied to a clinical environment. In our research, a system based on deep learning for automated fractures of shoulders, using the MURA dataset, was introduced. We compared a number of models of CNN, including DenseNet-201, ResNet-101, VGG-19, Inception-V3, and EfficientNet-B0. The highest performance, 88.45% accuracy, 91.01% precision, 86.10% recall, 88.43% F1-score, and 0.77 Cohen's kappa, was achieved with EfficientNet-B0. In order to improve model interpretability, explanation methods through AI, such as Grad-CAM, LIME, and Occlusion Sensitivity, were included, which expressed decision-critical regions and global/local intuitions. Feature visualization with convolutional layers further explained hierarchical learning. Overall, our model compared to three radiologists' Cohen's kappa value ranges between 0.79 to 0.86 (mean 0.84), resulted in similar or better outcome. These results highlight the potential of EfficientNet-B0, supported by robust XAI tools, as a reliable and interpretable decision-support system for musculoskeletal radiology.