Silhouette images enable estimation of body fat distribution and associated cardiometabolic risk

The study addresses the problem that overall adiposity measures such as BMI and waist circumference inadequately capture inter-individual differences in fat distribution that drive cardiometabolic risk. Prior imaging work shows VAT increases risk, ASAT is neutral, and GFAT may be protective, but imaging is not scalable. The research question is whether simple body silhouettes can be used with deep learning to accurately estimate specific fat depot volumes (VAT, ASAT, GFAT) and VAT/ASAT ratio, and whether these silhouette-derived metrics associate with cardiometabolic diseases independently of BMI and waist circumference. The goal is to bridge the gap between crude anthropometrics and resource-intensive imaging for scalable risk assessment.

Prior studies using MRI, CT, and DEXA linked specific fat depots to cardiometabolic risk, with VAT harmful, ASAT neutral, and GFAT protective. Guidelines recommend waist circumference as a clinical proxy for central adiposity, but it cannot distinguish VAT from ASAT at the individual level. Earlier research using silhouettes, 2D photographs, and 3D optical scans predominantly predicted overall fat or fat-free mass rather than individual depots, did not assess depot ratios, and was limited by small sample sizes (hundreds of participants) restricting validation and generalizability. Anthropometric models can correlate with depot volumes due to dependence on overall body size, but they perform poorly for depot ratios that are relatively independent of BMI and waist circumference. No previous study demonstrated that outline-derived fat distribution metrics stratify cardiometabolic disease risk independent of BMI and waist circumference.

Study population: 40,032 UK Biobank imaging substudy participants with MRI-derived VAT, ASAT, and GFAT volumes previously quantified. Mean age 65 years; 51% female; 97% White. Ethics approvals obtained (UK Biobank application #7089). Silhouette generation: Whole-body MRIs were acquired as six series; resampled to 2.232 × 2.232 × 3.0 mm3, de-duplicated overlaps, merged into 3D volumes. Using fat-phase acquisition, axial images were segmented to delineate the body outline. A surface map of the 3D segmentation was computed and projected to 2D coronal (front-back) and sagittal (side) images. Pixel intensities were binarized (body surface=1, background=0). The coronal and sagittal silhouettes were concatenated and resized to 237×256 pixels as model input. Deep learning model: DenseNet-121 backbone in a hierarchical multi-task configuration jointly predicting VAT, ASAT, GFAT volumes, and VAT/ASAT ratio from the paired silhouettes. Training and evaluation used a nested cross-validation strategy: the cohort was split into five non-overlapping partitions; for each fold, models were trained on three partitions, validated on one, and tested on one; predictions from validation partitions were aggregated to produce unbiased predictions for all participants. Final fold predictions used mean-ensembles of internal cross-validation models. Comparator anthropometric models: Sex-specific linear models using combinations of age, weight, height, BMI, waist circumference, hip circumference, waist-to-hip ratio (WHR), and five bioelectric impedance measurements. The same nested cross-validation and bootstrapping (1000 resamples) were used to estimate performance (R², MAE, 95% CI). Additional comparison included a published multivariable model predicting DEXA-derived VAT mass from 17 anthropometric variables. Association analyses: Sex-stratified logistic regression (prevalent type 2 diabetes, coronary artery disease, hypertension, hypercholesterolemia) adjusted for age and imaging center, with additional adjustment for BMI and waist circumference. Cox proportional-hazards models for incident type 2 diabetes and coronary artery disease over median 2.8 years follow-up with the same covariates. Analyses also examined gradients of disease prevalence across quintiles of silhouette-predicted VAT/ASAT within BMI- and waist-circumference strata. Sensitivity analyses replaced silhouette-predicted VAT/ASAT with MRI-derived VAT/ASAT to compare associations.

• Performance for fat depot volumes using silhouettes: Cross-validated R² for prediction of MRI-derived depots: VAT R²=0.885 (95% CI 0.882–0.887), ASAT R²=0.934 (0.932–0.935), GFAT R²=0.932 (0.930–0.934). Performance consistent across ages; some attenuation across sex and BMI subgroups; VAT prediction attenuated in Black participants (R²=0.784; 95% CI 0.735–0.823). - Silhouettes vs anthropometrics for depot volumes: Compared with BMI-based models, silhouettes improved R² by ΔR²=0.220–0.241 (VAT), 0.114–0.172 (ASAT), 0.248–0.263 (GFAT). Waist-circumference-based models only slightly improved VAT vs BMI (male R²=0.637; female R²=0.659) and were worse for ASAT and GFAT. Combining all anthropometric and bioimpedance measures yielded R²: VAT male 0.724 (0.717–0.732), VAT female 0.731 (0.723–0.739); ASAT male 0.829 (0.823–0.835), ASAT female 0.898 (0.895–0.901); GFAT male 0.793 (0.785–0.801), GFAT female 0.856 (0.852–0.860). Silhouettes still outperformed these by ΔR²=0.101–0.125 (VAT), 0.049–0.065 (ASAT), 0.092–0.098 (GFAT). A published 17-variable anthropometric model for DEXA VAT performed similarly to the combined anthropometric model (male R²=0.719; female R²=0.710), and was outperformed by silhouettes by ΔR²=0.122–0.130. - VAT/ASAT ratio prediction: Anthropometric models predicted poorly (male R²=0.171; female R²=0.262). WHR alone was similar (male R²=0.138; female R²=0.246). Silhouette-based models markedly improved prediction (male R²=0.553; 95% CI 0.542–0.562; female R²=0.504; 95% CI 0.492–0.516). Waist circumference correlated strongly with silhouette-predicted VAT (male R² 0.72; female R² 0.76) and ASAT (male 0.73; female 0.74), but poorly with silhouette-predicted VAT/ASAT (male 0.07; female 0.20). - Disease associations (prevalent): In sex-specific logistic regressions adjusted for age and imaging center, each SD increase in silhouette-predicted VAT/ASAT associated with higher odds of type 2 diabetes (male OR/SD 1.78; 95% CI 1.69–1.87; female OR/SD 1.97; 1.85–2.09). After additional adjustment for BMI and waist circumference: male 1.70 (1.61–1.80), female 1.74 (1.62–1.86). For coronary artery disease adjusted for BMI and waist circumference: male OR/SD 1.22 (1.16–1.29), female 1.21 (1.09–1.33). Trends were consistent using MRI-derived VAT/ASAT. - Disease prevalence gradients: Within BMI- and waist circumference-stratified bins, higher silhouette-predicted VAT/ASAT quintiles showed substantially higher standardized prevalence of type 2 diabetes and coronary artery disease. Example: Overweight men with normal waist circumference in the top VAT/ASAT quintile had type 2 diabetes probability 9.5% (95% CI 8.6–10.4%) vs 3.7% (3.0–4.5%) for overweight men with elevated waist but bottom VAT/ASAT quintile. - Incident outcomes (median 2.8 years): Adjusted Cox models showed increased risk per SD in silhouette-predicted VAT/ASAT for incident type 2 diabetes (male HR/SD 1.33; 95% CI 1.13–1.57; female 1.51; 1.30–1.74) and incident coronary artery disease in males (HR/SD 1.19; 95% CI 1.08–1.30); a similar directional but nonsignificant association in females (HR/SD 1.09; 95% CI 0.94–1.27).

Using only binarized 2D silhouettes derived from body MRI, the deep learning model accurately predicted MRI-quantified fat depots and, critically, VAT/ASAT ratio, addressing the core question of whether simple outline images can recover clinically relevant fat distribution. The results show that silhouettes capture BMI-independent variation in regional adiposity that anthropometric and bioimpedance models miss, particularly for VAT/ASAT ratio, which is relatively independent of overall body size. Strong associations of silhouette-predicted VAT/ASAT with prevalent and incident type 2 diabetes and coronary artery disease, minimally attenuated by adjustment for BMI and waist circumference, demonstrate clinical relevance for risk stratification. These findings suggest that less data-rich, scalable imaging modalities (and potentially consumer-grade imaging) can bridge the gap between detailed imaging biomarkers and routine clinical practice. The approach may also help identify individuals with pathogenic fat distribution phenotypes (e.g., lipodystrophy spectrum) who may benefit from targeted interventions. Overall, the work supports silhouette-derived fat distribution metrics as informative, practical tools for cardiometabolic risk assessment beyond traditional measures.

The study demonstrates that deep learning applied to simple silhouette images can accurately estimate VAT, ASAT, GFAT volumes and VAT/ASAT ratio, outperforming anthropometric and bioimpedance-based models, especially for VAT/ASAT. Silhouette-predicted VAT/ASAT is a strong, largely BMI- and waist circumference-independent predictor of prevalent and incident cardiometabolic disease. This establishes a foundation for scalable population-level assessment of fat distribution and associated risk. Future research should validate across diverse ancestries and age groups, develop and validate models using consumer-grade images (e.g., smartphone-derived silhouettes) accounting for acquisition variability, and assess longitudinal responsiveness to lifestyle and therapeutic interventions, including performance at extreme BMI values.

• Generalizability: The cohort was predominantly White and older (mean age ~65), necessitating validation in more diverse, younger, and geographically varied populations; differences in fat distribution across racial groups may affect performance. - Data modality: Silhouettes were derived from MRI; translation to cheaper, real-world modalities (e.g., smartphone images) will require handling acquisition heterogeneity and independent validation. - Extremes of phenotype: Training on cohorts with higher BMI may improve accuracy at distribution extremes. - Longitudinal change: The study did not evaluate accuracy for tracking changes in depot volumes over time; longitudinal assessment during interventions is needed.