AI-enabled left atrial volume quantification from CAC CT: comparison with echocardiography.

Abstract

Left atrial (LA) volume is a powerful prognostic factor in cardiovascular disease with emerging AI-driven quantification from non-contrast coronary artery calcium (CAC) CT. The accuracy and agreement of such volume quantification with echocardiography, however, is not well studied. This study compares LA volumes captured from CAC CT and transthoracic two-dimensional echocardiography (2DE). This retrospective cohort included 220 patients undergoing both echocardiography and CAC CT within a 1-year period. CAC CT images were analyzed using a deep learning-based autoencoder model to derive chamber volumes. LA volume measurements were compared using Pearson correlation and Bland-Altman plots. Sub analysis was performed based on CT cardiac phase, body mass index, and sex with LA volume indexed to body-surface area (BSA). Mean LA volume differed significantly between 2DE and CAC CT with volumes of 32.2 ± 13.7 mL/m2 and 45.7 ± 15.9 mL/m2 (P < 0.001), respectively. CAC CT-derived LA volume correlated strongly with 2DE-measured LA volume (r = 0.82, ICC = 0.57, P < 0.001) and substantially with clinical severity (Cohen's κ = 0.70, P < 0.001). 2D-echo systematically underestimated LA volume with CAC-CT, with bias of 13.5 mL/m2 [95% CI, 12.3-14.7]. Agreement remained consistent for patients with (r = 0.83, ICC = 0.66, P < 0.001) and without (r = 0.77, ICC = 0.62, P < 0.001) heart failure. Inter-modality agreement for chamber volumes remained robust across sex, BMI, cardiac phase, and vendor sub-analyses. AI-derived LA volume from CAC CT demonstrated strong correlations with 2DE measurements and severity regardless of cardiac phase, BMI, and sex. Despite systematic differences in absolute volume compared to 2DE, CAC CT revealed relative LA remodeling, supporting the use of CAC CT as a modality for chamber quantification.