Diffusion model-enhanced patella shape analysis predicts knee osteoarthritis progression and surgical replacement risk

Abstract

Patellofemoral osteoarthritis (OA) is a significant contributor to knee OA symptoms and functional impairments, with early morphological changes often originating in the patellofemoral joint. To facilitate early interventions in knee OA management, there is a pressing need for prognostic artificial intelligence (AI) models that can accurately forecast and explain the risk of disease deterioration. Recent studies have demonstrated the benefits of incorporating image generative models as an intermediate step in predicting knee OA progression from radiographs. Building on the remarkable advancements in image generation quality of diffusion models, this study presents a novel approach to knee OA prognosis using synthetic follow-up patella shapes for patella shape analysis. Our pipeline commences with a two-step patella segmentation model, achieving high accuracy (Dice score: 0.973, IoU: 0.947) in extracting precise patella masks from lateral knee radiographs. We then employ a diffusion model to predict patella shape trajectories over a 60-month period, conditioned on baseline patella masks. This model generates synthetic follow-up patella shapes that capture future morphological changes associated with OA progression. To predict knee OA outcomes, we developed a specialized 2-channel 1D-CNN with circular padding, incorporating both baseline and synthetic follow-up patella shapes to predict key outcomes of disease onset and end-stage. Our results demonstrate the significant prognostic value of patella shapes derived from lateral view radiographs. SynPatNet exhibited excellent performance in predicting patellofemoral OA onset (AUC: 0.909, vs. 0.830 for baseline model) and knee replacement risk (AUC: 0.823, vs. 0.773 for baseline), outperforming models using only baseline shapes. Notably, our model's knee replacement risk scores showed significant correlations with MRI-based gradings (Spearman's rho up to 0.51 for osteophytes, 0.33 for cartilage morphology, and 0.31 for bone attrition; p < 0.001), supporting the biological plausibility of the learned shape signal. The significant boost in prognostic performance achieved by providing synthesized follow-up images to the model highlights the importance of capturing OA pathological information in patella shapes. Furthermore, the synthesized future radiographs offer interpretable explanations for the model's predictions, addressing the limitations of black box decision-making. This study demonstrates the potential of incorporating diffusion models into deep learning-based prognostic applications for OA. By leveraging readily available lateral knee radiographs, our approach offers a cost-effective means of enhancing OA risk stratification and underscores the importance of considering the patellofemoral joint in comprehensive knee OA assessment.