Automatic Segmentation of Visible Epicardium Using Deep Learning in CT Image

Abstract

Recently, many studies indicate that epicardial adipose tissue (EAT) volume closely associates with progression of coronary atherosclerosis as well as increased prevalence of coronary artery disease. Segmentation of the epicardium is an important step for estimation of EAT volume. There are two main challenges facing automatic epicardium segmentation. One is that the epicardium exists as a very thin line and the contrast with the surrounding adipose tissue, blood vessels, and the heart is low. The other is that there are quite a few invisible parts. In this study, we propose a fully deep-learning automatic method for visible epicardium segmentation, which is developed using 2D DenseU-Net, demonstrated that it is possible to automatically segment epicardium, and we also find that the segmentation results of visible epicardium are affected by EAT volume.