Distilling clinical sensitivity: Explainable and lightweight knee osteoarthritis diagnosis via decoupled knowledge distillation

Knee Osteoarthritis (KOA) is a leading cause of global disability, necessitating early and accurate diagnosis to prevent irreversible joint degeneration. While deep learning has automated radiographic assessment, current state-of-the-art models incur massive computational costs, hindering deployment in resource-constrained clinical settings. To bridge this gap, we propose a high-efficiency framework utilizing Decoupled Knowledge Distillation (DKD). We employ a Swin Transformer teacher to guide a lightweight GhostNetV2 student, decoupling the distillation loss into Target Class (TCKD) and Non-Target Class (NCKD) components. The method achieves a binary classification accuracy of 87.02% (95% CI: 0.870 ± 0.016) on the Osteoarthritis Initiative (OAI) dataset. While overall accuracy is statistically comparable to a non-distilled baseline (P > 0.05),the DKD framework yields a statistically significant improvement in diagnostic sensitivity (0.825 vs. 0.767; non-overlapping 95% CI), critically reducing false negatives. The model matches the heavy teacher’s performance while being approximately 6 times smaller and 5 times faster. Validated via LayerCAM, the model demonstrates precise localization of osteophytes consistent with radiological interpretations, presenting a scalable, clinically safe solution for automated KOA screening.