Diffusion model-empowered patella shape analysis predicts knee osteoarthritis outcomes

We developed and validated an artificial intelligence pipeline that leverages diffusion models to enhance prognostic assessment of knee osteoarthritis (OA) by analyzing longitudinal changes in patella shape on lateral knee radiographs.

In this retrospective study of 2,913 participants from the Multicenter Osteoarthritis Study, left-knee weight-bearing lateral radiographs obtained at baseline and 60 months were analyzed. Our pipeline commences with an automatic segmentation for patella shapes, followed by a diffusion model to predict patella shape trajectories over 60 months. We developed the Synthetic Patella Shape Incorporated Convolutional Neural Network (SynPatNet), a specialized 2-channel 1-dimensional convolutional neural network (CNN), to incorporate both baseline and synthetic follow-up patella shapes for predicting key outcomes of disease onset and end-stage.

The diffusion model generates plausible synthetic patella shapes that predict deformations and osteophyte developments at the 60-month follow-up. Incorporating synthetic follow-up shapes with baseline patella shapes significantly improved OA outcome prediction: for patellofemoral OA onset, SynPatNet achieved an area under receiver operating characteristic curve (AUC) of 0.909 (vs. 0.830 for baseline model); for knee replacement, an AUC of 0.823 (vs. 0.773 for baseline). Augmenting Kellgren-Lawrence (KL) grade with SynPatNet further improved knee replacement prediction (AUC 0.838) over KL grade alone (AUC 0.785). Noteworthily, our knee replacement risk prediction score showed significant correlations with MRI-based (osteophytes/cartilage morphology/bone attrition) gradings, with Spearman's rho up to (0.51/0.33/0.31, p ​< ​0.001).

Generative diffusion modelling of patellar morphology on lateral knee radiographs provides complementary information to conventional radiographic and clinical metrics that substantially improves prognostication of knee OA.