Deep learning-based classification of musculoskeletal radiographs: optimizing CNN architectures with model interpretability

Abstract

Medical imaging in the form of X-ray radiography is the foundation of diagnosing musculoskeletal disorders. However, manual interpretation of radiologists is time-consuming and prone to inconsistency, so automated deep learning-based classification is essential. This paper uses convolutional neural networks (CNNs) for classifying musculoskeletal radiographs from the MURA dataset with 5,107 X-ray images. Four deep learning models VGG19, Xception, MobileNetV2, and ConvNeXtBase were trained and tested using benchmark classification performance measures: Recall, Precision, and F1-score. MobileNetV2 proved to be the best baseline with an F1-score of 0.92, which on the application of hyperparameter tuning with increased learning rate, dropout, and L2 regularization rose to 0.94. Preprocessing techniques involving auto-orientation, resizing, and data augmentation (blurring, rotation, shearing, and noise addition) enhanced generalizability in the model. To enhance clinical validity, interpretability methods Local Interpretable Model-Agnostic Explanations (LIME), Saliency Maps, and Gradient-weighted Class Activation Mapping (Grad-CAM) were integrated to achieve visual representation of the model's focus on clinically important anatomical structures. These methods verified the

model's focus on joint and bone regions, in line with radiology expertise. Fine- tuned MobileNetV2 model integrated with interpretability offers high accuracy and computational efficiency, making it clinically suitable for assisting radiologists. This approach enhances diagnostic precision, reduces workload, and advances AI-driven medical image analysis for the detection of musculoskeletal disorders.