A deep-learning pipeline using nnU-Net architecture trained on 165 cardiac MRI cases achieved automated segmentation of epicardial adipose tissue (EAT) in under one minute per case, versus 58.4 ± 7.9 minutes for manual segmentation. The model reached a Dice similarity coefficient of 0.896, volumetric correlation of r = 0.984, and intraclass correlation of 0.988 against expert ground truth — critically outperforming inter-observer agreement (bias of −25.24 mL) and matching intra-observer reproducibility (bias of 2.72 mL). Only 6.7% of cases required minor manual correction.
EAT has emerged as a clinically significant biomarker precisely because it is metabolically active, secreting pro-inflammatory adipokines directly onto the myocardium and coronary vessels. Elevated EAT volume has been independently associated with atrial fibrillation risk, coronary artery disease progression, and adverse cardiac remodeling — making scalable, reproducible quantification a genuine clinical priority. Previous CT-based EAT methods exist but CMR's 3D Dixon technique offers superior soft-tissue contrast without ionizing radiation, making it preferable for longitudinal research and younger patients. The 58-fold speed improvement this pipeline delivers could realistically unlock large-scale population studies previously blocked by annotation bottlenecks. Limitations include a modest validation cohort of 30 cases and training data predominantly from a single institutional dataset, raising questions about generalizability across scanner vendors and acquisition protocols. As a preprint posted on medRxiv, these results have not yet undergone peer review and should be interpreted cautiously. If externally validated, this tool could be considered paradigm-shifting for EAT research workflows.
