Estimating body fat distribution – a driver of cardiometabolic health – from silhouette images

Marcus D. R. Klarqvist, Saaket Agrawal, Nathaniel Diamant, Patrick T. Ellinor, Anthony Philippakis, Kenney Ng, Puneet Batra, Amit V. KheramedRxiv 2022 · DOI

Introduction

Body-mass index (BMI) remains the clinical standard for assessing obesity, yet it is a blunt instrument that fails to account for the significant inter-individual variation in fat distribution. Two individuals with identical BMIs can have vastly different health profiles depending on where their fat is stored. Specifically, visceral adipose tissue (VAT), which surrounds internal organs, is a potent driver of cardiometabolic diseases like type 2 diabetes (T2D) and coronary artery disease (CAD). Conversely, gluteofemoral fat (GFAT) in the hips and thighs may offer protective effects.

While medical imaging such as MRI or CT scans can accurately quantify these fat depots, their cost and complexity make them impractical for routine clinical use or large-scale population screening. Traditional anthropometric measures, such as waist circumference, often fail to distinguish between dangerous internal VAT and less harmful abdominal subcutaneous adipose tissue (ASAT). This study presents a deep learning model trained on 2D body silhouettes to bridge the gap between low-cost anthropometry and high-precision medical imaging.

Methods

The researchers leveraged data from 40,032 participants in the UK Biobank imaging substudy (mean age 65; 51% female). To develop the model, 3D MRI data were converted into 2D coronal (front-facing) and sagittal (side-facing) silhouettes. These silhouettes represent only the external outline of the body, devoid of internal texture or grayscale information.

A DenseNet-121 convolutional neural network (CNN) was employed as a hierarchical multi-task learner. The model was trained to simultaneously predict VAT, ASAT, and GFAT volumes, as well as the VAT/ASAT ratio—a key marker of unhealthy fat distribution. Ground truth labels were derived from high-quality fat volumes previously quantified from MRI scans. The model's performance was validated using a nested cross-validation approach and compared against standard linear models incorporating age, weight, height, BMI, waist and hip circumference, and bioelectric impedance.

Results

The silhouette-based deep learning model demonstrated high predictive accuracy, significantly outperforming traditional clinical metrics.

Predictive Performance

The CNN achieved $R^2$ values of 0.88 for VAT, 0.93 for ASAT, and 0.93 for GFAT. Notably, for the VAT/ASAT ratio—which is largely independent of BMI and difficult to estimate using traditional means—the silhouette model achieved an $R^2$ of 0.50–0.55. In contrast, comparator models combining BMI, waist circumference, and bioimpedance only reached $R^2$ values of 0.17–0.26.

Clinical Risk Stratification

The silhouette-predicted VAT/ASAT ratio was strongly associated with the prevalence of cardiometabolic diseases. Each standard deviation increase in the predicted ratio was linked to an increased odds of type 2 diabetes (OR ~1.7–2.0) and coronary artery disease, even after adjusting for BMI and waist circumference.

The model was particularly effective at identifying "hidden" risk. For example, among men classified as "overweight" by BMI but with a "normal" waist circumference, those in the top quintile of silhouette-predicted VAT/ASAT ratio had a 9.5% probability of diabetes, compared to just 3.7% for those in the bottom quintile.

Discussion

The primary insight of this research is that the external silhouette of the human body contains substantially more information about internal fat distribution than a combination of traditional clinical measurements. By focusing on the VAT/ASAT ratio, the model identifies individuals with the "thin-on-the-outside, fat-on-the-inside" phenotype who might otherwise be missed by standard screenings.

Limitations and Future Directions

While promising, the study has limitations. The UK Biobank cohort is 97% white, and fat distribution patterns are known to vary significantly across different ethnicities. Future validation in more diverse populations is essential. Additionally, the silhouettes used were derived from high-resolution MRIs. Further research is required to determine if the model maintains its accuracy when applied to silhouettes generated from standard smartphone photography, which introduces variables like lighting, clothing, and camera angles.

Conclusion

This "fat shadow" approach provides a scientific foundation for low-cost, population-based assessment of body fat distribution. By enabling more precise risk stratification than BMI or waist circumference alone, this technology could significantly improve the early identification and management of cardiometabolic disease.

Summary updated

Abstract

ABSTRACT Background Inter-individual variation in fat distribution is increasingly recognized as clinically important but is not routinely assessed in clinical practice because quantification requires medical imaging. Objectives We hypothesized that a deep learning model trained on an individual’s body shape outline – or “silhouette” – would enable accurate estimation of specific fat depots, including visceral (VAT), abdominal subcutaneous (ASAT), and gluteofemoral (GFAT) adipose tissue volumes, and VAT/ASAT ratio. We additionally set out to study whether silhouette-estimated VAT/ASAT ratio may stratify risk of cardiometabolic diseases independent of body mass index (BMI) and waist circumference. Methods Two-dimensional coronal and sagittal silhouettes were constructed from whole-body magnetic resonance images in 40,032 participants of the UK Biobank and used to train a convolutional neural network to predict VAT, ASAT, and GFAT volumes, and VAT/ASAT ratio. Logistic and Cox regressions were used to determine the independent association of silhouette-predicted VAT/ASAT ratio with type 2 diabetes and coronary artery disease. Results Mean age of the study participants was 65 years and 51% were female. A deep learning model trained on silhouettes enabled accurate estimation of VAT, ASAT, and GFAT volumes (R 2 : 0.88, 0.93, and 0.93, respectively), outperforming a comparator model combining anthropometric and bioimpedance measures (ΔR 2 = 0.05-0.13). Next, we studied VAT/ASAT ratio, a nearly BMI- and waist circumference-independent marker of unhealthy fat distribution. While the comparator model poorly predicted VAT/ASAT ratio (R 2 : 0.17-0.26), a silhouette-based model enabled significant improvement (R 2 : 0.50-0.55). Silhouette-predicted VAT/ASAT ratio was associated with increased prevalence of type 2 diabetes and coronary artery disease. Conclusions Body silhouette images can estimate important measures of fat distribution, laying the scientific foundation for population-based assessment.
BibTeX
@misc{klarqvist2022estimating,
  author = {Marcus D. R. Klarqvist and Saaket Agrawal and Nathaniel Diamant and Patrick T. Ellinor and Anthony Philippakis and Kenney Ng and Puneet Batra and Amit V. Khera},
  title = {{Estimating body fat distribution – a driver of cardiometabolic health – from silhouette images}},
  year = {2022},
  doi = {10.1101/2022.01.14.22269328},
  howpublished = {medRxiv},
}
Figure