MediVision: An Explainable and Robust Deep Learning Framework for Knee Osteoarthritis Grading

Abstract

Knee osteoarthritis (KOA) is a major cause of disability in the elderly. Typically, diagnosis relies on X-rays and the Kellgren-Lawrence (KL) grading scale. However, interpretation of X-ray images can vary and be inconsistent among clinicians. This study presents MediVision, a deep-learning tool aimed at automating the classification and grading of KOA. Three neural network models, InceptionV3, Xception, and DenseNet201, were tested on a dataset of 3,300 knee X-rays annotated by experts. DenseNet201 outperformed the other networks, achieving a test accuracy of 93.87 %. To clarify the model's decision-making process, techniques such as Grad-CAM and occlusion sensitivity were used, showing that the model focused on key features, such as joint space narrowing. The robustness of the model was further evaluated under noisy conditions, with DenseNet201 maintaining an accuracy of above 97.8 %. A web application was developed to demonstrate the potential of MediVision in a clinical setting. Despite its promise, this study faced challenges, including data imbalance and reliance on a single public data source. Nonetheless, these findings suggest that combining accuracy, robustness, and interpretability can enhance the use of artificial intelligence in the clinical assessment of KOA.

Ähnliche Arbeiten