Interpretable Deep Learning for Musculoskeletal Radiograph Classification Using ResNet, DenseNet, and Explainable AI Methods

Abstract

Musculoskeletal (MSK) disorders are a leading cause of disability worldwide. While radiography remains the primary diagnostic tool, increasing demand and a limited radiologist workforce often result in delayed or inconsistent MSK diagnoses. Deep learning methods, particularly Convolutional Neural Network (CNN), have shown potential in automating image classification; however, their black-box nature limits clinical adoption and trust. This study presents an explainable deep learning framework for binary classification of MSK radiographs using the MURA v1.1 dataset, employing ResNet-50 and DenseNet-121 architectures augmented with a prominent explainability technique i.e., Grad-CAM for coarse spatial localization and Integrated Gradients for fine-grained attributions. The pipeline integrates preprocessing, augmentation, and transfer learning, with evaluation across standard metrics. DenseNet-121 achieved the strongest validation performance with 0.8017 accuracy, 0.7875 precision, 0.8020 recall, 0.7947 F1-score, and 0.8773 AUC-ROC, outperforming ResNet-50, which yielded 0.7804 accuracy and 0.7296 F1-score. These results demonstrate that dense connectivity better balances precision and recall, reducing the likelihood of missed abnormalities while maintaining interpretability. Our findings underscore the clinical potential of combining predictive accuracy with transparent explainability, supporting radiologist trust, auditability, and safe AI deployment in MSK imaging.

Ähnliche Arbeiten