XKneeNet: An Explainable Lightweight Deep Learning Framework for Multi-Class Knee Osteoarthritis Grading from X-Ray Images

Abstract

Knee osteoarthritis (KOA) is a progressive musculoskeletal disorder that poses significant diagnostic challenges, particularly in early-stage identification. To address the limitations of computationally intensive deep learning models with limited explainability, this study proposes XKneeNet-a lightweight and interpretable deep learning framework for automatic KOA severity classification using radiographic images. The framework leverages a fine-tuned ShuffleNet for efficient deep feature extraction, followed by feature dimensionality reduction via the Minimum Redundancy Maximum Relevance (MRMR) technique. The reduced features are classified using a decision tree model optimized through Bayesian hyperparameter tuning. Additionally, Grad-CAM++ is employed to generate classspecific visual explanations, enhancing model transparency and clinical trust. Evaluated on a five-class benchmark knee X-ray dataset, XKneeNet achieved a high classification accuracy of 96.06%, with an average AUC of 0.9610. The results demonstrate that XKneeNet effectively balances diagnostic performance, computational efficiency, and interpretability, making it suitable for real-time clinical applications.

Ähnliche Arbeiten