Deep learning segmentation and quantification method for assessing epicardial adipose tissue in CT calcium score...

Area of Science:

• Radiology
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Epicardial adipose tissue (EAT) volume is a biomarker for coronary artery disease (CAD) and major adverse cardiac events.
• Manual EAT quantification is labor-intensive, requires expertise, and is susceptible to errors.

Purpose of the Study:

• To develop and validate DeepFat, a deep learning algorithm for automated EAT volume assessment.
• To evaluate the accuracy and efficiency of DeepFat compared to manual segmentation on non-contrast low-dose CT calcium score images.

Main Methods:

• DeepFat utilizes a HU-attention-window and a novel slab-of-slices with bisection (bisect) preprocessing technique.
• Segmentation is performed on axial CT slices, followed by EAT volume calculation using a fat window threshold.
• The method segments tissue within the pericardial sac on CT images.

Main Results:

• DeepFat achieved excellent performance with a volume Dice score of 88.52% ± 3.3 and slice Dice of 87.70% ± 7.5.
• The algorithm demonstrated a low EAT error of 0.5% ± 8.1 and high correlation (R=98.52%) with manual segmentation.
• The HU-attention-window and bisect methods individually improved Dice volume scores.

Conclusions:

• DeepFat provides an accurate, automated method for EAT volume quantification from CT calcium score images.
• The algorithm's performance is comparable to inter-observer variability, offering a reliable alternative to manual assessment.
• DeepFat has the potential to improve the efficiency and consistency of cardiovascular risk stratification.