Automatic Knee Joint Detection for Kellgren–Lawrence Osteoarthritis Grading in Plain X-Rays

Knee osteoarthritis (OA) is a common degenerative joint disease and is usually assessed on plain radiographs using the Kellgren–Lawrence (KL) grading scale. This work presents a two-stage deep learning workflow for knee joint localization and KL-grade assessment in X-ray images. The dataset was obtained from the Osteoarthritis Initiative (OAI) and included 4007 bilateral radiographs. After contrast enhancement and U-Net-based ROI extraction, the full dataset yielded 8014 single-knee ROIs and 12,520 ROI images for grading after augmentation. For classification, pretrained convolutional neural networks were fine-tuned using an internal 80/20 training-validation split of the ROI dataset. The localization model showed consistent overlap with manually defined knee regions, with all held-out cases achieving a Jaccard Index ≥0.8 and 31/40 achieving ≥0.9. In KL grading, the best-performing configuration achieved a quadratic-weighted Cohen’s kappa (QWK) of 0.87, an accuracy of 0.72, and a weighted F1-score of 0.72. The pattern of disagreement was also informative: errors were concentrated around neighboring grades, mainly KL0–KL1 and KL1–KL2. Grad-CAM visualizations highlighted regions around the joint margins and inter-bone spaces, areas anatomically plausible for KL assessment. Complementary analyses were added using participant-disjoint OAI partitions, a locked test of 1,656 knees from 828 unseen participants, and an independent external cohort of 361 clinically graded knees. Taken together, the results support the proposed workflow as a practical internally validated approach for automated knee ROI localization and KL grading. A broader clinical interpretation will require validation beyond the current OAI-based setting.