Rheumatoid arthritis classification and prediction by consistency-based deep learning using extremity MRI scans
Yanli Li, Tahereh Hassanzadeh, Denis P. Shamonin, M. Reijnierse, Annette H M van der Helm–van Mil, Berend C. Stoel
Abstract
Predicting the development of rheumatoid arthritis (RA) in an early stage through magnetic resonance imaging (MRI) can initiate timely treatment and improve long-term patient outcomes. Although manual prediction is time-consuming and requires expert knowledge, automatic RA prediction has not been fully investigated. While standard models fail to achieve acceptable performance, we present a consistency-based deep learning framework to classify and predict RA automatically and precisely, including an output-standardized model, customized self-supervised pretraining and a loss function that is based on label consistency between original and augmented inputs. For training and evaluation, we used a database, containing 5945 MRI scans of carpal, metacarpophalangeal (MCP), and metatarsophalangeal (MTP) joints, from 2151 subjects obtained over a period of ten years. Four (three classification- and one prediction-) tasks were defined to distinguish two patient groups (with recent-onset arthritis and clinically suspect arthralgia) from healthy controls and RA from other arthritis patients within the recent-onset arthritis group, and predict RA development in a period of two years within the clinically suspect arthralgia group. The proposed method was evaluated with the area under the receiver operating curve (AUROC) on a separate test set, achieving mean AUROCs of 83.6%, 83.3%, and 69.7% in the three classification tasks, and 67.8% in the prediction task. This proves the existence of early signs of RA in MRI and the potential of a consistency-based deep learning model to detect these early signs and predict RA.