Abstract
Epicardial adipose tissue (EAT) significantly contributes to the progression of cardiovascular diseases (CVDs). However, manually quantifying EAT volume is labor-intensive and susceptible to human error. Although there have been some deep learning-based methods for automatic quantification of EAT, they are mostly uninterpretable and fail to harness the complete anatomical characteristics. In this study, we proposed an enhanced deep learning method designed for EAT quantification on coronary computed tomography angiography (CCTA) scan, which integrated both data-driven method and specific morphological information. A total of 108 patients who underwent routine CCTA examinations were included in this study. They were randomly assigned to training set (n = 60), validation set (n = 8), and test set (n = 40). We quantified and calculated the EAT volume based on the CT attenuation values within the predicted pericardium. The automatic method demonstrated strong agreement with expert manual quantification, yielding a median Dice score coefficients (DSC) of 0.916 (Interquartile Range (IQR): 0.846-0.948) for 2D slices. Meanwhile, the median DSC for the 3D volume was 0.896 (IQR: 0.874-0.908) between these two measures, with an excellent correlation of 0.980 (p < 0.001) for EAT volumes. Additionally, our model's Bland-Altman analysis revealed a low bias of -2.39 cm³. The incorporation of pericardial anatomical structures into deep learning methods can effectively enhance the automatic quantification of EAT. The promising results demonstrate its potential for clinical application.