Attention-enhanced deep learning and machine learning framework for knee osteoarthritis severity detection in football players using X-ray images

Abstract

Objective

To develop a highly accurate, interpretable, and scalable framework for automated knee osteoarthritis (OA) classification, integrating attention-enhanced autoencoders for feature extraction and advanced machine learning techniques for robust and clinically reliable severity grading based on X-ray images.

Materials and methods

This study analyzed 5987 knee X-ray images from football athletes (18–45 years) using the Kellgren-Lawrence (KL) grading system. Preprocessing involved resizing, normalization, and quality checks, with data augmentation (rotations, flips, brightness adjustments) to enhance model robustness. A convolutional autoencoder (CAE) with attention mechanisms extracted key features, improving interpretability and accuracy. Machine learning classifiers (SVM, XGBoost, Stacking) processed features from the bottleneck layer, while dimensionality reduction (PCA, LDA, RFE) optimized feature selection, enhancing classification performance.

Results

The study assessed knee OA classification using autoencoders with and without attention mechanisms. Without attention, dimensionality reduction techniques like PCA and RFE performed well, particularly when combined with ensemble classifiers. RFE + Stacking achieved the highest F1-score (82.99 %), while PCA + SVM and PCA + XGBoost delivered high accuracy (86.57 % and 85.52 %, respectively). Incorporating attention mechanisms significantly boosted performance, with RFE + Stacking attaining the best overall results (AUC: 96.5 %, F1-score: 93.5 %). Additionally, PCA + Stacking and PCA + XGBoost demonstrated strong accuracy (92.99 % and 92.12 %, respectively). End-to-end autoencoders with attention outperformed their non-attentive counterparts, reaching an accuracy of 0.94 and an AUC of 0.95. These findings underscore the critical role of attention mechanisms in enhancing model robustness, accuracy, and interpretability, making them highly applicable for clinical decision-making.

Conclusion

This study introduces a highly interpretable AI framework for knee OA classification, integrating attention-enhanced autoencoders to highlight key diagnostic regions in X-ray images. By incorporating attention mechanisms, our model improves transparency and clinical relevance, ensuring that classification decisions are guided by meaningful radiological features rather than arbitrary patterns.