Fine-grained epicardial adipose tissue segmentation in cardiac CT images with position priors and edge enhancement.

Cardiovascular disease stands as the leading cause of death globally, and substantial research has revealed its close correlation with the distribution of epicardial adipose tissue (EAT). Moreover, existing studies have demonstrated that EAT exhibits significant differences in distribution patterns and pathophysiological roles across various anatomical regions of the heart. Therefore, the quantitative analysis of EAT at different cardiac locations is crucial, and fine-grained segmentation of EAT via cardiac CT is an efficient method for obtaining the corresponding metrics. The existing computer-aided segmentation approaches typically treat EAT as a unified whole, which fails to meet the demands of nuanced diagnostics, and faces challenges such as class imbalance, thin structures, and anatomical variation, leading to low segmentation accuracy, limiting its application in cardiovascular disease risk assessment. To address these issues, we extend the existing segmentation strategy to the fine-grained segmentation of the left ventricle- (LV-), right ventricle- (RV-), and peri-atrium- (PA-) EAT, and propose the PRAEE framework based on position priors and edge enhancement. The core innovations of the proposed method are as follows: (1)Position-Prior Regularization: Considering the spatial distribution patterns of EAT in different anatomical regions, we design a regularization module that incorporates prior knowledge of typical spatial locations of various types of EAT as auxiliary constraints. This mechanism effectively guides the model to more accurately localize and differentiate EAT across anatomical regions, enabling an initial segmentation.(2)Adaptive Edge Enhancement: To further improve segmentation accuracy, we develop an edge enhancement module that explicitly extracts critical edge information through morphological operations and integrates it into the network architecture, significantly refining segmentation along boundary regions. Our approach has been validated on both a self-collected EAT dataset and the publicly available ACDC and MM-WHS datasets, demonstrating its effectiveness in enhancing fine-grained discrimination and edge detail performance.