Improving robustness in X-ray image classification through attention mechanisms in convolutional neural networks

Aim Musculoskeletal radiographs (X-rays) are commonly used to diagnose bone and joint abnormalities because they are affordable and accessible. Convolutional neural networks (CNN) have demonstrated strong performance in medical image classification. However, interpreting these images reliably is challenging due to a lack of labelled data, inherent image noise, and the lack of explainable artificial intelligence (AI). This research aims to improve the robustness against noise, accuracy, and interpretability of musculoskeletal radiograph classification by addressing these key challenges. Method This study proposes an attention-enhanced deep learning framework that combines in-domain self-supervised pretraining, region-specific attention-based convolutional neural networks, and hierarchical feature fusion. To reduce redundancy and improve generalisation, descriptor-based feature selection is applied before classification. The final classification is performed through an ensemble of classifiers, and visual explanation techniques (such as Grad-CAM) are used to enhance model transparency. Finding The proposed model demonstrated high classification accuracy, achieving 95.48% on the MURA dataset and 99.80% on the Chest X-ray dataset. The model worked reliably even when images were noisy and was able to detect essential signs in unclear scans. The visualisations showed that the attention layers focused on meaningful areas in the X-rays, making the results easier to understand. Implication Integrating attention mechanisms, feature selection, and ensemble learning significantly mitigated overfitting and improved generalisation and explainability. The model demonstrated consistent performance across datasets and provided interpretable outputs, supporting its potential for clinical deployment. Conclusion This study presents a robust and interpretable deep learning framework for musculoskeletal X-ray analysis. By addressing core limitations in existing models, namely, data scarcity, noise sensitivity, and a lack of explainability, the proposed method enhances the reliability of AI-assisted diagnosis in medical imaging.