Utilizing Cascade Deep Metric Learning for the Kellgren-Lawrence Grading of Knee Osteoarthritis Classification from X-Ray Images

Accurate automated grading of knee osteoarthritis from X-ray images remains challenging due to inter-radiologist label noise and the inherent ordinal nature of the Kellgren-Lawrence (KL) grading scale. Most existing deep learning approaches address label noise handling, feature embedding, and ordinal modeling as separate components, which limits robustness and consistency across adjacent grades. This study presents a Cascade Deep Metric Learning (CDML) framework that integrates adaptive label reliability updating and ordinal-aware metric learning within a unified cascade design. By iteratively refining feature embeddings under ordinal constraints, the proposed framework explicitly addresses both noisy annotations and ordinal dependencies in KL grading. Experiments conducted on the Osteoarthritis Initiative (OAI) dataset demonstrated that the proposed method achieved an accuracy of 82.12%, a Mean Absolute Error (MAE) of 0.179, and a Quadratic Weighted Kappa (QWK) of 0.935, outperforming baseline deep metric learning methods. Ablation studies further confirmed the effectiveness of the cascade design in improving robustness and preserving clinically consistent ordinal predictions.