Comparison of Machine Learning Methods for Predicting Modified Total Shape Score in X-ray Radiography

It is predicted that there are 600,000 to 1,000,000 patients with rheumatoid arthritis (RA) in Japan. To quantitatively diagnose RA using X-ray images, the modified total sharp score (mTSS) has been used, although, the evaluation method depends on the experience of the physicians. It desires computer-aided diagnosis (CAD) systems to improve the quality of diagnosis for RA patients. In this study, we compare a method using ridge regression (RR) and methods using 3 models (VGG16, DenseNet201 and Xception) of convolutional neural network (CNN) on mTSS prediction methods for hand rheumatoid arthritis of the hand. To compare the 4 method, we conducted an experiment on 90 RA patients using X-ray images of their hands. The experimental results showed that the mTSS prediction of erosion was best with RR, and the mTSS prediction of JSN was best with VGG16.