KL Grading-Based Knee Abnormality Classification Using Multi-Modal Deep Learning: KOMMCF – A Fusion of X-Ray Imaging and Clinical Biomarkers

Abstract

Osteoarthritis of the knee (KOA) is a progressive musculoskeletal ailment that increasingly distresses the mobility of patients and their quality of life. Although the KL-Kellgren & Lawrence grading system is extensively used to stage osteoarthritis, it is challenging to accurately stage early osteoarthritis because of the minute changes that occur in the joint. This study accentuates a three-grade classification of severity of osteoarthritis of the human knee: Normal (KL:0), Early OA (KL:1–2), and Moderate to Severe OA (KL:3–4). We proposed an efficient multi-modal deep learning framework Knee Osteoarthritis Multi-Modal Classification Framework (KOMMCF). The framework integrates X-ray images and fourteen biomarkers obtained from the Osteoarthritis Initiative (OAI). It encompasses advanced X-ray image processing, bilateral difference features, and normalized biomarkers. An ordinal distance-aware loss function is used to address natural order of KL grades. In addition, patient-level Group K-Fold cross-validation is adopted to ensure robust testing and avoid data leakage. Although mainly trained for three-class classification, the framework was also tested for binary classification (Normal vs. Abnormal OA) and five-class classification (KL0–KL4). For the Kaggle-OAI dataset, mean cross-validation accuracies for the 2-class, 3-class, and 5-class problems were 98.67%, 90.58%, and 80.82%, respectively. Corresponding test accuracies were 98.67%, 90.58%, and 81.34%, with macro-average ROC-AUC values of 0.9993, 0.9683, and 0.94. McNemar’s statistical tests confirmed that KOMMCF significantly outperforms ResNet50, VGG16, and MobileNetV2 (p < 0.001). This work demonstrates power of combining imaging and clinical data in an AI-driven approach to knee abnormality classification, potentially supporting early diagnosis and treatment planning.