LMSST-GCN: Longitudinal MRI sub-structural texture guided graph convolution network for improved progression prediction of knee osteoarthritis

Background and objectives: Accurate prediction of progression in knee osteoarthritis (KOA) is significant for early personalized intervention. Previous methods commonly focused on quantifying features from a specific sub-structure imaged at baseline and resulted in limited performance. We proposed a longitudinal MRI sub-structural texture-guided graph convolution network (LMSST-GCN) for improved KOA progression prediction.

Methods: 600 KOA participants from the Osteoarthritis Initiative underwent 3 longitudinal MRI scans at baseline, 12 and 24 months. 3D nnU-net was adopted to segment 32 sub-structures of each knee on both IW and DESS sequences at each time point. 105 radiomics features were extracted from each sub-structure, mRMR was used for feature selection, and only the most representative feature was retained to characterize its texture. Each patient was encoded into a 1D vector with 192 features by concatenating all features from 32 sub-structures on the 2 sequences at the 3 time points. Then a population graph was constructed with each vertex representing each patient and edges determining their connection/similarity. The graph was further fed into EdgeGCN to generate the probability of progression. A clinical model and three kinds of machine-learning models including Support Vector Machine (SVM), Random Forest, and Extreme Gradient Boosting (XGBoost) were also constructed for comparison. Interpretability analysis by using GNNExplainer was conducted to explain the association between specific knee sub-structure and KOA progression.

Results: The proposed LMSST-GCN model and its variants (AUC ≥ 0.82) significantly outperformed the clinical model (AUC ≤ 0.72) and machine learning models (AUC ≤ 0.77, p ≤ 0.05 for all). Model performance benefits from the involvement of more sequences and more time points, the highest AUC of 0.85 was achieved by LMSST-GCN model constructed by using all available information. The interpretability analysis demonstrated that the loss of cartilage and sclerosis of subchondral bone at the tibial medial central region, the injury of lateral meniscus, and abnormal changes of the infrapatellar fat pad are more responsible for progression.

Conclusions: The proposed LMSST-GCN model characterized the texture of all knee sub-structures on longitudinal multi-sequence MRI and identified patients prone to progression in the scenario of vertex classification in a population graph, providing a novel strategy for improved prediction of KOA progression. The code was made publicly available at https://github.com/JunyiPeng-SMU/EdgeGCN.