Deep learning–derived pericardial adipose tissue from cardiovascular magnetic resonance imaging in 62,000 UK Biobank participants

Background

Pericardial adipose tissue (PAT) is a metabolically active cardiac fat depot implicated in inflammation, atrial remodelling, and coronary atherosclerosis. Although linked to cardiovascular disease independent of body mass index (BMI), translation has been limited by lack of automated, reproducible quantification at scale. We developed and evaluated a deep learning framework for PAT segmentation from cardiovascular magnetic resonance (CMR) images in the UK Biobank and assessed its value as a scalable biomarker for cardiovascular risk.

Methods

A manually annotated reference dataset of 3,462 participants undergoing mid-ventricular T1-mapping CMR was created, including one T1 map and seven T1-weighted images per participant. An nnU-Net–based framework was adapted for automated PAT segmentation and quantification. Performance was evaluated using five-fold cross-validation against manual measurements.

The validated model was deployed in 62,868 participants. Associations between PAT and cardiometabolic traits and CMR cardiac measures were examined using multivariable linear regression. Prospective associations with incident coronary heart disease (CHD) and atrial fibrillation (AF) were assessed in 61,336 participants using Cox models adjusted for age, sex, BMI, blood pressure, diabetes, and lipids. Incremental prognostic value beyond conventional risk factors was evaluated using change in discrimination (ΔC-statistic). Sex-specific equations based on body size and composition were derived for non-imaging estimation of PAT.

Results

Automated segmentation achieved high accuracy (Dice=0.84) with close agreement to manual PAT areas (mean difference 0.26 cm²; r=0.91), exceeding inter-observer variability (0.53 cm²). PAT was quantified in 62,611 participants, generating the largest standardized CMR-based PAT dataset. Mean PAT area was 13.48 cm² (men 15.30, women 11.89). PAT correlated strongly with visceral adipose tissue (r=0.68) and moderately with anthropometric measures.

During follow-up, 1-SD higher PAT was associated with 17% higher CHD risk (N=1121; HR 1.17, 95% CI 1.10-1.25) and 20% higher AF risk (N=1032; HR 1.20, 95% CI 1.12-1.28). After adjustment for BMI, associations attenuated to ~10% but remained significant. PAT improved risk discrimination beyond conventional factors (ΔC-statistic 0.016). Sex-specific equations identified strongest predictors: in males, age, body lean mass, and fat (33% variance); in females, age, lean mass, and weight (35% variance). Predicted PAT correlated with imaging-derived values (r=0.65) and retained associations with CHD and AF (HR per SD predicted PAT: CHD 1.17, AF 1.19).

Conclusions

We present a fully automated, reproducible deep learning framework for population-scale PAT quantification from CMR. PAT independently predicts CHD and AF, adds prognostic value beyond conventional risk factors, and can be estimated using validated non-imaging equations, enhancing translational potential.