RFF-JSS: A radiomic feature fusion and joint space segmentation framework for automated knee osteoarthritis assessment

Knee osteoarthritis (KOA) is a leading cause of pain and functional disability worldwide. Radiographic severity grading using the Kellgren–Lawrence (KL) scale is widely adopted in clinical practice but remains subjective and susceptible to inter- and intra-observer variability, particularly when distinguishing adjacent grades. This study aims to develop an accurate, interpretable, and anatomically guided framework for automated KOA severity grading from plain knee radiographs. A novel Radiomic Feature Fusion and Joint Space Segmentation (RFF-JSS) framework is proposed for automated KOA severity assessment. The pipeline includes (i) standardized preprocessing of knee radiographs, (ii) anatomically informed tibiofemoral joint space segmentation using a U-Net model, (iii) IBSI-compliant multiscale radiomic feature extraction from medial and lateral compartments, (iv) PCA-based radiomic feature fusion for dimensionality reduction and redundancy suppression, and (v) ordinal-aware KOA severity classification using a Random Forest classifier. The framework was evaluated on 5,820 knee radiographs from two large public cohorts—the Osteoarthritis Initiative (OAI) and the Multicenter Osteoarthritis Study (MOST)—with stratified training, validation, and testing splits. The proposed RFF-JSS framework achieved an accuracy of 94.87%, precision of 95.12%, sensitivity of 94.68%, specificity of 96.10%, F1-score of 94.89%, ROC-AUC of 0.972, and a Quadratic Weighted Kappa (QWK) of 0.947. The method consistently outperformed state-of-the-art approaches, including ResNet-SVM, UNet-Radiomics, Deep-KOA, JSN-Net, RF-RFE, and CNN-KLNet. Ablation studies confirmed the critical contributions of joint space segmentation, compartment-wise analysis, feature standardization, and PCA-based radiomic fusion in improving classification robustness and reducing misclassification between adjacent KL grades. The RFF-JSS framework provides a robust, interpretable, and anatomically grounded solution for automated KOA severity grading from plain radiographs. By synergistically combining joint space segmentation with radiomic feature fusion and ordinal-aware classification, the proposed approach bridges the gap between high predictive performance and clinical interpretability, demonstrating strong potential for large-scale screening and clinical decision support in knee osteoarthritis assessment.