Learned knee cartilage and meniscus shape features are associated with osteoarthritis incidence

We derive cartilage and meniscus point clouds from 40,796 high resolution knee MR images and train point cloud networks to extract osteoarthritis shape features. We demonstrate the utility of these learned features by assessing their relative contributions in a Cox Proportional Hazard Regression model with existing clinical risk factors predicting incident radiographic osteoarthritis. Shape biomarkers for tibiofemoral joint cartilage and menisci had significantly increased hazard ratios. The best performing shape biomarker– tibial and femoral cartilage shape– combined with clinical risk factors achieved a concordance index of 0.759. Our findings suggest point cloud learned shape features are promising OA biomarkers.