Radiomics and Machine Learning for Automated Grading of Knee Osteoarthritis

Knee osteoarthritis (KOA) is a progressive degenerative joint disorder characterized by cartilage loss, subchondral bone remodeling, and inflammation, ultimately leading to pain and impaired mobility. Early and accurate assessment of KOA severity is essential for personalized clinical management. Although the Kellgren–Lawrence (KL) grading system remains the radiographic gold standard, it is limited by subjectivity and inter-observer variability. This study evaluates a radiomics-based machine learning framework using Teachable Machine for automated KOA grading from knee X-ray images. Radiographs were categorized into KL Grades 0–4, and radiomic features capturing intensity and texture patterns were used to train models under varying epoch and batch-size configurations. The results demonstrate that model performance varied significantly across grades, with consistently higher accuracy for extreme grades (G0 and G4) and lower performance for early and intermediate grades (G1–G3), reflecting the subtle nature of mild KOA changes. Optimal performance was observed at moderate training durations (80–90 epochs) and larger batch sizes, while extended training (100 epochs) led to overfitting and reduced generalizability. Despite these challenges, the radiomics-based approach shows potential for objective and reproducible KOA severity assessment. This work highlights the feasibility of accessible ML platforms for supporting clinical decision-making and improving KOA diagnostic workflows.