Interpretable deep learning for rotator cuff tear diagnosis: A novel convolutional neural network with Grad-CAM visualization on MRI

Accurate diagnosis of rotator cuff tears from magnetic resonance imaging (MRI) is essential for effective clinical management and treatment planning. In this study, we propose a novel convolutional neural network (CNN) architecture specifically designed for classifying rotator cuff tears, integrated with gradient-weighted class activation mapping (Grad-CAM) to provide interpretable insights into the model’s decision-making process. We utilized MRI data from 150 subjects, equally divided between normal and pathological cases, and applied data augmentation techniques, including rotation, scaling, and reflection, to enhance model generalization. The proposed CNN demonstrated superior performance, achieving an average accuracy of 94.5%, sensitivity of 94.6%, precision of 94.1%, and specificity of 93.4%, outperforming established lightweight models such as MobileNetV2 and SqueezeNet. Grad-CAM visualizations confirmed that the model accurately focused on anatomically relevant regions associated with tendon ruptures, thereby enhancing trust in its predictions. These results underscore the potential of our interpretable deep learning framework to deliver reliable, transparent, and clinically actionable diagnostic support for shoulder injuries, paving the way for improved decision-making in orthopedic care. This approach highlights the synergy of advanced CNN design and explainable AI for robust medical imaging applications.