Associating Knee Osteoarthritis Progression with Temporal-Regional Graph Convolutional Network Analysis on MR Images

Background: Artificial intelligence shows promise in assessing knee osteoarthritis (OA) progression on MR images, but faces challenges in accuracy and interpretability. Purpose: To introduce a temporal-regional graph convolutional network (TRGCN) on MR images to study the association between knee OA progression status and network outcome. Study Type: Retrospective. Population: 194 OA progressors (mean age, 62 +/- 9 years) and 406 controls (mean age, 61 +/- 9 years) from the OA Initiative were randomly divided into training (80%) and testing (20%) cohorts. Field Strength/Sequence: Sagittal 2D IW-TSE-FS (IW) and 3D-DESS-WE (DESS) at 3T. Assessment: Anatomical subregions of cartilage, subchondral bone, meniscus, and the infrapatellar fat pad at baseline, 12-month, and 24-month were automatically segmented and served as inputs to form compartment-based graphs for a TRGCN model, which containing both regional and temporal information. The performance of models based on (i) clinical variables alone, (ii) radiologist score alone, (iii) combined features (containing i and ii), (iv) composite TRGCN (combining TRGCN, i and ii), (v) radiomics features, (vi) convolutional neural network based on Densenet-169 were compared. Statistical Tests: DeLong test was performed to compare the areas under the ROC curve (AUC) of all models. Additionally, interpretability analysis was done to evaluate the contributions of individual regions. A P value <0.05 was considered significant. Results: The composite TRGCN outperformed all other models with AUCs of 0.841 (DESS) and 0.856 (IW) in the testing cohort (all P < 0.05). Interpretability analysis highlighted cartilage's importance over other structures (42%-45%), tibiofemoral joint's (TFJ) dominance over patellofemoral joint (PFJ) (58%-67% vs. 12%-37%), and importance scores changes in compartments over time (TFJ vs. PFJ: baseline: 44% vs. 43%, 12-month: 52% vs. 39%, 24-month: 31% vs. 48%). Data Conclusion: The composite TRGCN, capturing temporal and regional information, demonstrated superior discriminative ability compared with other methods, providing interpretable insights for identifying knee OA progression. Level of Evidence: 4. Technical Efficacy: Stage 2.