Automated pericardium segmentation and epicardial adipose tissue quantification from computed tomography images

Background and Objective

Epicardial Adipose Tissue (EAT) is regarded as an independent risk factor for cardiovascular disease, and an increase in its volume is closely associated with disorders such as coronary artery atherosclerosis. Traditional manual and semi-automatic methods for EAT segmentation rely on subjective judgment, resulting in uncertainty and unreliability, which limits their application in clinical practice. Therefore, this study aims to develop a fully automatic segmentation and quantification method to improve the accuracy of EAT assessment.

Methods

A Boundary-Enhanced Multi-scale U-Net network with a Convolutional Transformer (BMT-UNet) is developed to segment the pericardium. The BMT-UNet comprises Boundary-Enhanced (BE) modules, Multi-Scale (MS) modules, and a Convolutional Transformer (ConvT) module. The MS and BE modules in the encoding part are designed to capture detailed boundary features and accurately delineate the pericardium boundary by combining multi-scale features with morphological operations, leveraging their complementarity. The ConvT module integrates global contextual information, thereby enhancing overall segmentation accuracy and addressing the issue of internal holes in the segmented pericardial images. The volume of EAT is automatically quantified using standard fat thresholds with a range of −190 to −30 HU.

Results

For a Coronary Computed Tomography Angiography (CCTA) dataset which contained 50 patients, the Dice coefficient and Hausdorff distance for the proposed method of pericardial and EAT segmentation are 98.3% ± 0.2%, 5.7±0.8 mm, and 93.9% ± 1.7%, 2.1 ± 0.3 mm, respectively. The linear regression coefficient between the EAT volume segmented and the actual volume is 0.982, and the Pearson correlation coefficient is 0.99. Bland-Altman analysis further confirmed the high consistency between the automated and manual methods. These results demonstrate a significant improvement over existing methods, particularly in terms of segmentation precision and reliability, which are critical for clinical application.

Conclusions

This work develops an automated method for quantifying EAT in Computed Tomography (CT) images, and the results agreed closely with expert evaluations. Code is available at: https://github.com/wy-9903/BMT-UNet.