Harnessing Deep Learning to Predict and Classify Knee Osteoarthritis from X-Ray imagery

Abstract

Knee osteoarthritis (OA) represents the most prevalent subtype of arthritis—a chronic and progressive musculoskeletal disorder—characterized radiographically by joint space narrowing, osteophyte formation, subchondral sclerosis, and bony deformities. Radiographic imaging (X-ray) remains the clinical gold standard for OA diagnosis due to its cost-effectiveness and immediate accessibility. Disease severity is traditionally assessed using the Kellgren and Lawrence (KL) grading system, which stratifies osteoarthritis progression from mild to severe stages. Timely identification of OA facilitates early intervention and can significantly decelerate degenerative changes. However, many contemporary diagnostic models simplify or exclude intermediate KL grades to optimize predictive performance, which compromises diagnostic granularity. This study introduces a deep learning–based ordinal classification framework for the automated detection and grading of knee OA from a single posteroanterior weight-bearing X-ray image, adhering strictly to the KL grading scale.

The study utilizes X-ray images sourced from the Osteoarthritis Initiative (OAI) dataset, divided into training, validation, and testing cohorts in a 70:20:10 ratio. Leveraging transfer learning, we fine-tuned multiple state-of-the-art CNN architectures—ResNet-34, VGG-19, DenseNet-121, and DenseNet-161—and employed an ensemble strategy to enhance overall model robustness and performance. Our approach demonstrated exceptional outcomes, achieving an overall accuracy of 98%, a Quadratic Weighted Kappa of 0.99, and a 95% confidence interval, with significant improvements across all KL grading levels. Furthermore, the proposed method surpasses existing state-of-the-art automated diagnostic systems.