Automated Kellgren–Lawrence grading of knee osteoarthritis using a multi-scale attention-based deep learning framework

Researchers developed a novel deep learning system for automated grading of knee osteoarthritis using the Kellgren–Lawrence scale. The system combines three key components: a Feature Pyramid Network for processing images at different resolutions, a dual attention mechanism to identify important regions, and knowledge distillation to improve accuracy. The model was developed on 5,000 X-ray images (3,500 for training, 1,500 for internal testing) and tested on an additional 2,000 images. On internal validation, the model achieved 72.6% accuracy and F1 score of 0.726, outperforming standard models. On independent test data, it maintained good performance with 68.5% accuracy, showing only a 4.04% decline. Errors occurred only between adjacent severity grades without extreme misclassifications. The system is interpretable and focuses on clinically relevant regions of the X-ray image.