Deep Learning-Based Quantification of Epicardial Adipose Tissue: A Review

Epicardial adipose tissue (EAT) is a potential imaging biomarker of cardiovascular risk. Manual EAT quantification is time-consuming and subject to observer variability. Deep learning (DL) offers tools for faster and more consistent measurements. This narrative review examines DL approaches to EAT quantification, their anatomical foundations, imaging applications, and clinical validation. We start with the definition of EAT to establish the clear anatomical and embryological differences between EAT and other fat depots (paracardial fat) to ensure accurate AI model construction. We then outline the advancements made in DL architecture development, from traditional U-Net models to advanced ones combining CNN and Transformers with regard to global features acquisition, including new models such as Mamba, which enables higher computation efficiency. Finally, we analyze how these models perform in different imaging techniques, focusing specifically on CT and MRI imaging. Reported segmentation performance varies with the imaging modality, anatomical target, dataset, and evaluation protocol. Although several models show good agreement with expert annotations, external validation and integration across hospital systems remain important challenges. Beyond volume, EAT attenuation and tissue characteristics may provide complementary information for cardiovascular risk assessment. Their clinical use requires reproducible measurements and validation in the intended patient population.