Knee Osteoarthritis Analysis Using Deep Learning and XAI on X-Rays

Abstract

Knee osteoarthritis (OA) is a chronic disorder mainly arising from age-related factors affecting the knee joints. Its diagnosis is critically important and is usually done by medical practitioners using X-ray images. Although this process is accurate, it is time-consuming. X-ray images have facilitated the use of deep learning (DL) models for the automation of the diagnosis of knee OA, commonly employing convolutional neural network (CNN) based architectures. However, the lack of models’ interpretability makes the results less trustworthy. This work builds on the existing state-of-the-art (SOTA) pre-trained DL models to understand the model’s behavior in classifying highly complex knee OA cases utilizing a divide-and-conquer approach - from multi-class to a binary class for better results interpretability and explainability using explainable artificial intelligence (XAI). Five SOTA fine-tuned DL models are tested on Kellgren-Lawrence (KL) graded X-ray images. Both multi-class and binary-class (using the multiple subsets derived from the original dataset to examine how the models perform with different data combinations) classification approaches and their interpretability of findings using Gradient-weighted Class Activation Mapping (GradCAM) are undertaken in this study. The GradCAM visualization of EfficientNetb7 demonstrates that when the degree of variance between different classes increases, the model’s efficiency in classifying knee OA also increases. Specifically, it becomes more effective at distinguishing normal and severe cases with 99.13% classification accuracy. However, the model’s efficacy drops to 67% for other cases, indicating that it cannot classify knee OA as effectively as doctors.

Keywords

• Explainable artificial intelligence (XAI)
• deep learning
• knee osteoarthritis
• healthcare
• diagnosis
• classification