Claudia Iriondo1, Jinhee Lee1, Sharmila Majumdar1, and Valentina Pedoia1
1University of California, San Francisco, San Francisco, CA, United States
1University of California, San Francisco, San Francisco, CA, United States
Cartilage and meniscus point clouds from 40,796 knee MR images are used to train point cloud networks to extract shape features. Features are used as variables in a Cox Proportional Hazard model with existing clinical risk factors. Learned shape features predict incident osteoarthritis.
Figure3. Four example subjects, their processed point clouds and subject description. PAT-FEM-TIB, PAT-FEM, FEM-TIB, and MEN compartment combinations were used to train point cloud networks for OA diagnosis. Each compartment is represented by 8192 randomly sampled points. The first three examples are labelled as having osteoarthritis (Kellgren Lawrence grade >=2), while the last example does not. Output p(OA) is used as the shape biomarker feature for Cox PH models. FEM= femur, TIB=tibia, PAT=patella, MEN= menisci
Figure4. Per compartment test results on pretext OA diagnosis task. (L to R) ROC curve, PR curve, and calibration curve with respective performance metrics. Differences between ROC curves are tested using Delong's method, cartilage models were not significantly different from each other, while menisci was significantly different from all cartilage models (p=1e-14, 1e-10, 1e-15).