Deep learning-based quantification of epicardial adipose tissue volume from non-contrast computed tomography images:...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Epicardial adipose tissue (EAT) is a key biomarker for coronary artery disease (CAD) progression.
• Accurate quantification of EAT volume is crucial for risk stratification.
• Current methods for EAT volume measurement can be time-consuming.

Purpose of the Study:

• To develop and validate a deep learning system for automated EAT volume quantification.
• To assess the system's performance using non-contrast computed tomography (NCCT) scans.
• To evaluate the system's generalizability across diverse ethnic populations.

Main Methods:

• A 3D UNet++ deep learning model was trained on 1243 NCCT scans for pericardium segmentation.
• Intensity thresholding was applied to derive EAT volume.
• The model was validated on an external cohort of 160 patients, including non-Asian individuals.

Main Results:

• The AI system quantified EAT volume in approximately 30 seconds per scan.
• Excellent agreement was observed between AI-predicted and expert-annotated EAT volumes (r = 0.975).
• AI-derived EAT volume independently predicted obstructive CAD and improved risk prediction when added to coronary calcium scoring.

Conclusions:

• A validated deep learning system enables automated EAT volume quantification from NCCT scans.
• The system demonstrates high accuracy and generalizability across diverse ethnicities.
• This tool has potential for routine EAT assessment and enhanced CAD risk stratification.