From detection to grading: A hybrid KOA-YOLOv5-RF model for knee osteoarthritis diagnosis

This study presents a novel computer-aided diagnostic (CAD) system for detecting and grading the severity of knee osteoarthritis(KOA) from X-ray images, utilizing a hybrid deep learning and machine learning framework. The system combines YOLOv5 for precise knee joint localization and segmentation with a Random Forest classifier for ordinal Kellgren-Lawrence (KL) grading. Trained on a curated and augmented dataset of 1535 X-ray images, the model achieves an overall KL grading accuracy of 87 %. Evaluation includes ROC-AUC curves, Cohen's kappa scores, and grade-wise sensitivity and specificity metrics. This hybrid approach offers a scalable, interpretable, and clinically relevant tool for supporting radiologists in early KOA diagnosis, especially in resource-constrained settings.•Combines the powerful feature extraction capabilities of the YOLOv5 deep learning architecture with the classification strength of the Random Forest model.•YOLOv5 is used for knee joint segmentation to reduce background noise and improve classifier accuracy by focusing on the region of interest.•Achieves 87 % overall accuracy in KL grading, with enhanced sensitivity to subtle changes in early-stage KOA (Grades 1-2).