Quantification of Epicardial Adipose Tissue in Low-Dose Computed Tomography Images

Abstract

The total volume of Epicardial Adipose Tissue (EAT) is a well-known independent early marker of coronary heart disease. Though several deep learning methods were proposed for CT-based EAT volume estimation with promising results recently, automatic EAT quantification on screening Low-Dose CT (LDCT) has not been studied. We first systematically investigate a deep-learning-based approach for EAT quantification on challenging noisy LDCT images using a large dataset consisting of 493 LDCT and 154 CT studies from 569 subjects. Our results demonstrate that (1) 3D U-net precisely segment the pericardium interior region (Dice score 0.95 ± 0.00 ); (2) postprocessing based on narrow 1-mm Gaussian filter does not require adjustments of EAT Hounsfield interval and leads to accurate estimation of EAT volume (Pearson’s R 0.96±0.01 ) comparing to CT-based manual EAT assessment for the same subjects.