Carpal Recognition Method for Automated Evaluation of Rheumatoid Arthritis Progression in X-ray Radiograph Using Deep Learning

Abstract

There are more than 600,000 patients with rheumatoid arthritis (RA) in Japan. The modified total sharp score (mTSS) is the standard diagnosis method of RA progression using X-ray radiograph. The mTSS has two evaluation criteria: erosion and joint space narrowing (JSN) where the evaluation points for mTSS in the hand joint are 16 for erosion and 15 for JSN. For the appropriate treatment, it is necessary to measure mTSS periodically and to observe the trend. However, it is hard to diagnose periodically because it takes long diagnostic time due to the large number of evaluation points. Another problem is that the evaluation of mTSS depends on the skill and experience of the physician. Therefore, a computer-aided diagnosis (CAD) system based on X-ray radiograph analysis is needed to reduce the burden on physicians and to enable objective diagnosis. We have previously proposed a CAD system, which can detect finger joint positions of hand and estimate mTSS in hand X-ray radiograph using machine learning. In this study, we propose an automated detection method of mTSS evaluation points in the carpal site from hand X-ray images using deep learning. Experimental results on hand X-ray radiographs of 140 RA patients showed that the mTSS evaluation points at the carpal site could be detected with a root mean squared error of about 7 mm.