Cutting-Edge Deep Learning Approach for Multi-Class Knee Osteoarthritis Diagnosis Support System

Abstract

Knee osteoarthritis commonly referred to as KO is prevalent in both the elderly and the young population and results in limitation of joint movements. Thus, earlier diagnosis and the identification of the disease at an early stage are favorable as the number of interventions is limited at later stages. Historical diagnostic techniques are subjective and depend on the physicians and the possibilities for mistakes are great. This paper develops a new deep-learning model that applies a ConvMixer for feature extraction and the Transformers’ global context modeling capability for feature extraction and classify, so it is a powerful solution for KO classification. The model proposed in this work is designed to subdivide the images into KO-positive and KOnegative ones and to distinguish the major disease severity grades that vary from 0, or “healthy knees,” to 4, or “severe KO” symptoms varying from normal semblance of joint structures to joint injuries that may include new growth of bones, loss of space within a joint or distortion of a joint’s shape. In the first training case, the ConvMixer and the global context modeling capacity of Transformers were incorporated for feature extraction and classification phases, with accuracy standing at 92 percent. In the second experiment, ConvMixer and global context modeling potential of Transformers was used for feature extraction whereas, classification was done with the aid of Adaboost which resulted in an even better accuracy of 97%.