Disentanglement of Latent Factors of Radiographic Knee Images for Knee Osteoarthritis Severity Classification

Abstract

Knee osteoarthritis is the most common form of the musculoskeletal disorder that usually happens to the elderly. The diagnosis and treatment for knee osteoarthritis are causing a huge economic burden to society. Traditionally, the diagnosis for knee osteoarthritis was done by analyzing MRI (Magnetic Resonance Imaging). However, this method has problems in the way that it is costly and has limited access since they are only available in specialized medical institutions. There have been many researches that attempt to use x-ray images that are safe, cost-efficient, and commonly available. Their method often fails and shows poor results due to a lack of training dataset. In this paper, I proposed a novel autoencoder-based knee osteoarthritis classification system. Unlike the previous deep learning-based research, the proposed method disentangles the osteoarthritis-related latent factors from knee x-ray images. These latent factors are then trained to predict the severity of osteoarthritis. The proposed method achieves accuracy of 78.9% on the publicly available dataset. Our method produces accurate results without having a large dataset while successfully avoiding the curse of dimensionality. Throughout the comprehensive experiment, I have shown that the proposed method outperforms the existing state-of-the-art methods by a great accuracy margin.