An MRI-based radiomics framework for early identification and progression stratification in knee osteoarthritis: data from the osteoarthritis initiative.

OBJECTIVES: To develop a cascaded machine learning model based on MRI radiomics features from cartilage and subchondral bone to predict the incidence and progression of knee osteoarthritis (KOA), thereby addressing the need for early intervention in pre-radiographic stages. MATERIALS AND METHODS: The study analyzed 456 participants without radiographic OA (Kellgren-Lawrence [KL] 0-1) at baseline were selected from the Osteoarthritis Initiative (OAI) and randomly divided into training and testing cohorts (7:3). Participants underwent 3D DESS MRI of the right knee and were stratified into incident KOA and non-KOA groups based on 4-year radiographic outcomes, using 1:2 propensity score matching (PSM) to adjust for baseline confounders. Early and late progressors were further classified based on the timing of radiographic progression. Radiomic features of cartilage and subchondral bone were extracted from nnU-Net-based segmentation. Optimal features were selected through Least Absolute Shrinkage and Selection Operator (LASSO) regression and principal component analysis (PCA). A two-stage logistic regression (LR) classification framework was developed to predict KOA incidence and progression, with a cascaded LR model implemented for multi-class classification. Model performance was primarily evaluated using the area under curve (AUC), with SHapley Additive exPlanations (SHAP) for interpretability. RESULTS: SHAP analysis identified square_glrlm_ShortRunLowGrayLevelEmphasis and wavelet-HHL_firstorder_Minimum from subchondral bone as primary predictors. The combined cartilage and subchondral bone radiomics models demonstrated superior performance in predicting incidence (AUC: 0.985, 95% CI: 0.969-1.000) and progression (AUC: 0.738, 95% CI: 0.565-0.911) of KOA. The cascaded model achieved AUCs > 0.800 across categories, with an overall accuracy of 0.791. CONCLUSIONS: The MRI radiomics framework integrating cartilage and subchondral bone features, effectively predicts KOA incidence and progression, enhances individualized risk stratification, and facilitates timely clinical decisions. TRIAL REGISTRATION: Clinical trial number not applicable.