Accuracy of deep learning in quantifying epicardial adipose tissue: a systematic review and meta-analysis

Epicardial adipose tissue (EAT) is significantly associated with cardiovascular disease and poor prognosis. Deep learning (DL) models have been explored for automatic EAT segmentation, but systematic evidence verifying their effectiveness is currently lacking. We conducted this meta-analysis to evaluate the accuracy of DL models for EAT segmentation and quantification, aiming to support the development of automated tools. We searched the Cochrane Library, Embase, Web of Science, PubMed, and IEEE Xplore through February 20, 2026. Studies reporting DL-based quantitative EAT segmentation were included. Quality was assessed with QUADAS-2. The Dice similarity coefficient (Dice), correlation coefficient (R), and Bland-Altman bias were used to evaluate segmentation efficacy, the correlation between automated segmentation and reference standards, and consistency, respectively. A random-effects model with the Hartung-Knapp-Sidik-Jonkman method pooled these measures. Thirty eligible studies were included, covering five image types. For non-contrast computed tomography (NCCT), the pooled Dice was 0.888 (95% confidence interval [CI]: 0.853-0.924), and the pooled R reached 0.973 (95%CI: 0.958-0.982). The contrast-enhanced computed tomography (CECT) subgroup performed best, yielding a Dice of 0.910 (95% CI: 0.881-0.939) and an R of 0.989 (95% CI: 0.976-0.995). This meta-analysis quantifies the accuracy of DL models for EAT segmentation and volume quantification. The findings indicate that DL achieves robust accuracy on both NCCT and CECT images. However, given that most current studies rely on internal validation, future work should prioritize large-scale, multicenter external validation to support the translation of these promising DL tools into robust clinical diagnostic systems.