Deep-Learning Segmentation of Epicardial Adipose Tissue Using Four-Chamber Cardiac Magnetic Resonance Imaging

Area of Science:

• Cardiovascular Imaging
• Medical Artificial Intelligence
• Radiology

Background:

• Epicardial adipose tissue (EAT) overload is a significant cardiovascular risk factor.
• Manual quantification of EAT in cardiac MRI is time-consuming and prone to variability.
• Accurate EAT assessment is crucial for risk stratification in conditions like obesity and diabetes.

Purpose of the Study:

• To develop and validate an automated method for quantifying EAT area in cardiac MRI using deep learning.
• To compare the performance of a multi-frame U-Net model against a fully convolutional network (FCN) and human observers.
• To assess the clinical utility of automated EAT quantification for identifying patients at risk of EAT overload.

Main Methods:

• Deep learning segmentation using multi-frame fully convolutional networks (FCN), including an optimized U-Net, was applied to 3T cardiac cine MRI data from 100 subjects.
• Networks were trained on three consecutive cine frames for central frame segmentation using dice loss and evaluated via 4-fold cross-validation and an independent dataset.
• Segmentation performance was quantified using Dice Similarity Coefficient (DSC) and Relative Surface Error (RSE), compared against inter-observer variability.

Main Results:

• The multi-frame U-Net model achieved segmentation performance comparable to inter-observer variability (DSC = 0.77).
• U-Net significantly outperformed the FCN baseline model (p < 0.0001) across all evaluated metrics.
• Automated quantification demonstrated a 14.2% precision for the clinically relevant EAT range, achieving 70% accuracy in scaling patient risk.

Conclusions:

• Multi-frame U-Net provides an accurate and automated approach for EAT area quantification in standard cardiac cine MRI.
• The developed method offers a reliable tool for assessing EAT overload and patient cardiovascular risk.
• The FSLeyes plugin makes this advanced automated EAT quantification method accessible to the research community.