EAT Deep Segmentation: Deep Learning-Based Automated Segmentation and Quantification of Epicardial Adipose Tissue from Cardiac CT Images

Abstract

Epicardial adipose tissue (EAT) is the fat depot between the myocardium and the visceral pericardium, and its volume is an established marker of cardiometabolic risk. Measuring it on cardiac computed tomography (CT) requires manual slice-by-slice outlining, which is slow and varies between observers. This paper presents EAT Deep Segmentation, a deep learning framework for automated segmentation and quantification of EAT from cardiac CT images. The framework converts CT volumes and expert labels into paired two-dimensional slices, trains a U-Net encoder-decoder with skip connections to predict a pixel-wise EAT probability map, converts the map into a binary mask, and computes the EAT area from the mask and the pixel spacing. The model was implemented in TensorFlow and evaluated on a public EAT CT dataset, using 96 slices from 12 patients in a first experiment. Performance was measured with the Dice coefficient, intersection over union (IoU), precision and recall. The first training run gave a Dice coefficient of 0.277, an IoU of 0.172, a precision of 0.178 and a recall of 0.779. The high recall and low precision show that the network finds most EAT pixels but over-segments, which is expected with a small cohort and strong class imbalance. The work provides a reproducible baseline pipeline, and the analysis identifies larger patient cohorts, imbalance-aware loss functions and restriction to the pericardial region as the next steps toward clinically useful accuracy.