Deep neural networks improve diagnostic accuracy of rheumatoid arthritis using diffuse optical tomography

Rheumatoid arthritis (RA) is an auto-inflammatory disease that causes pain, swelling and stiffness in joints. Diffuse optical tomography (DOT) has shown promise as a non-invasive, diagnostic imaging tool for RA. However high inter-subject variability of derived optical parameters infer that at an early stage, small pathophysiological changes resulting from inflammation may be difficult to detect. A set of deep neural network models for RA classification is proposed together with a numerical model of the finger to generate data to overcome the inherent problem of insufficient clinical DOT images available. These proposed deep neural network models have been applied to automatically classify DOT images of inflamed and non-inflamed joints. The results demonstrate that three proposed deep neural network methods improve the diagnostic accuracy as compared to the widely applied support vector machine, especially for high inter-subject variability cases. Residual network achieved the highest accuracy (>99%) on the generated database, and highway and convolutional neural networks reached 99% and 90%, respectively. The results show that deep neural network methods are highly suitable for RA classification from DOT data and highlight the potential for deep neural network methods to be used as a computer aided tool in DOT diagnostic systems.