Predicting Knee Osteoarthritis Radiographic Progression: A Multimodal Diffusion Approach

This study aimed to develop KOA-Diff, a multimodal diffusion model predicting future knee radiographs for accurate, visually interpretable forecasting of knee osteoarthritis (KOA) progression.

KOA-Diff utilizes a latent diffusion backbone and a multimodal fusion network to integrate baseline X-rays, MRIs, and clinical characteristics. A feature-targeting attention mechanism guides image generation toward critical anatomical structures. The model was trained and internally validated on the Osteoarthritis Initiative (OAI) dataset and externally validated on a private cohort (PUTH-KOA, n=80). We evaluated the synthesized future radiographs using joint space width error (JSWE, error relative to the actual width) and binary Kellgren-Lawrence (KL) grade accuracy (early [KL ≤2] vs. advanced [KL ≥3]). Additionally, blinded experts graded structural features. Finally, clinical utility was tested by using model outputs to assist clinicians in identifying high-risk patients exhibiting a longitudinal KL grade increase ≥2 levels over 8 years.

In the OAI dataset, KOA-Diff achieved a JSWE of 0.71%-1.21% and binary KL-grade accuracy of 84.0%-91.8% over 96 months. External validation demonstrated promising performance with 90.0% binary KL-grade accuracy and 2.06% JSWE. In the blinded evaluation, experts rated the accuracy of synthesized features on a 5-point scale. For the OAI cohort, the model scored 4.162 (joint space narrowing), 3.895 (osteophyte formation), and 3.943 (subchondral bone sclerosis). PUTH-KOA scores were consistently high (4.050, 3.925, 3.775, respectively). With model assistance, clinicians' accuracy in identifying high-risk patients improved from 58.3% to 77.5%.

KOA-Diff synthesizes future radiographs reflecting accurate structural progression, providing an interpretable tool for KOA prediction.