Quantification of Epicardial Adipose Tissue Volume and Attenuation for Cardiac CT Scans Using Deep Learning in a...

Area of Science:

• Cardiology
• Radiology
• Artificial Intelligence

Background:

• Epicardial adipose tissue (EAT) is an independent prognostic marker for atrial fibrillation (AF) and influences myocardial function.
• EAT volume (EATv) and density (EATd) are key CT-derived parameters for quantifying EAT.
• Increased EATv correlates with AF prevalence and recurrence, while higher EATd indicates inflammation and plaque presence.

Purpose of the Study:

• To develop a fully automated deep learning (DL) framework for quantifying EAT volume (EATv) and density (EATd).
• To reduce variability and time associated with manual EAT quantification.
• To enhance reproducibility in studies involving EAT measurements.

Main Methods:

• A DL framework combining image classification and segmentation models was developed.
• The framework first selects relevant CT slices containing EAT and then segments EAT.
• EATv and EATd were estimated using segmentation masks on a 300-patient dataset (300-patient dataset split into training and evaluation sets).

Main Results:

• The classification model achieved 98% accuracy in precision, recall, and F1 scores.
• The segmentation model achieved a median Dice similarity coefficient of 0.84.
• High correlation coefficients were observed between labeled and predicted EATv (0.971) and EATd (0.972) on the evaluation set.

Conclusions:

• The proposed DL framework offers a fast, robust, and accurate method for EAT segmentation and quantification.
• This automated approach facilitates reproducible EAT measurements for clinical practice and large-scale research.
• The framework supports patient-level and high-throughput EAT quantification projects.