Automated quantification of pericoronary fat attenuation index using deep learning: a validation study against expert manual measurement

Authors:

Abstract
Background: Pericoronary fat attenuation index (FAI) is a novel biomarker for coronary inflammation and cardiovascular risk. However, manual FAI measurement is time-consuming and prone to observer variability. We aimed to develop and validate a deep learning (DL)-based software for automated FAI quantification. Methods: A DL framework comprising cardiac chamber segmentation, coronary artery segmentation, and vessel trajectory labeling networks was developed using 8,025 coronary computed tomography angiography (CCTA) scans. An external validation cohort of 209 patients (627 vessels) with stable coronary artery disease (CAD) was retrospectively enrolled to evaluate the agreement between the automated DL-FAI and expert manual measurements. Performance was assessed using processing time, Pearson correlation, Bland-Altman analysis, and linear regression. Results: The mean processing time for DL-FAI was significantly shorter than manual measurement (30 seconds vs. 15 minutes, PConclusions: The proposed DL-based method enables rapid automated quantification of coronary FAI. It shows good agreement with manual reference measurement while substantially improving processing efficiency, and it demonstrated preliminary stability in vessels with evident calcification. These findings support its potential use as a standardized method for large-scale FAI analysis.