Automatic Deep Learning Segmentation and Quantification of Epicardial Adipose Tissue in Non-Contrast Cardiac CT scans.

Abstract

An Automatic deep learning semantic segmentation (ADLS) using DeepLab-v3-plus technique is proposed for a full and accurate whole heart Epicardial adipose tissue (EAT) segmentation from non-contrast cardiac CT scan. The ADLS algorithm was trained on manual segmented scans of the enclosed region of the pericardium (sac), which represents the internal heart tissues where the EAT is located. A level of 40 Hounsfield unit (HU) and a window of 350 HU was applied to every axial slice for contrast enhancement. Each slice was associated with two additional consecutive slices, representing the three-channel single input image of the deep network. The detected output mask region, as a post-step, was thresholded between [-190, -30] HU to detect the EAT region. A median filter with kernel size 3mm was applied to remove the noise. Using 70 CT scans (50 training/20 testing), the ADLS showed excellent results compared to manual segmentation (ground truth). The total average Dice score was (89.31%±1.96) with a high correlation of (R=97.15%, p-value <0.001), while the average error of EAT volume was (0.79±9.21).Clinical Relevance- Epicardial adipose tissue (EAT) volume aids in predicting atherosclerosis development and is linked to major adverse cardiac events. However, accurate manual segmentation is considered tedious work and requires skilled expertise.