A Comparative Analysis of Deep Learning Approaches for Knee Osteoarthritis Detection Using Indian and Multi-Centric Datasets

Knee osteoarthritis affects 10-15% of the global population over the age of 60, mostly women. Knee osteoarthritis is growing more widespread in India due to changing lifestyle choices, rising urbanisation, and demographics. Furthermore, many serious cases in India go untreated since many believe it is only a concern for the elderly. Additionally, the diagnosis of knee osteoarthritis has been difficult in rural India due to a lack of skilled medical professionals. Additionally, people are ignorant of its severity, as evidenced by KL's grading system, which ranges from normal to severe (Grades 0-4). This study compares various deep learning architectures and their performance metrics in order to analyze the effectiveness of deep learning-based prediction models for knee osteoarthritis detection. This study utilized an Indian dataset because there is limited research conducted on Indian subjects. The OAI and MOST datasets, which comprise data from Western people, were used in the majority of the investigations. This is the first study, in which we compare the effects of knee osteoarthritis using three distinct datasets. In the study, various pre-trained CNN-based models for 5-class and 2-class classification was used. On an Indian dataset, the EfficientNetB1 model achieved testing accuracy of 84% and EfficientNetB7 on a 2-class classification achieved accuracy of 90.2%. Subsequently, 61% and 77.7% of the western dataset's accuracy comes from 5-class and 2-class categorization using VGG-19 model, respectively. Furthermore, the accuracy of the multi-centric dataset using EfficientNetB1, B7 for 5-class classification is 61%, whereas for 2-class classification it is 80.82% respectively.