Body Fat Estimation from Surface Meshes Using Graph Neural Networks

Tamara T. Mueller, Siyu Zhou, Sophie Starck, Friederike Jungmann, Alexander Ziller, Orhun Aksoy, Danylo Movchan, Rickmer Braren, Georgios Kaissis, Daniel Rueckert

Introduction

Standard health metrics such as Body Mass Index (BMI) and waist-to-hip ratios are frequently used in clinical settings but remain "blunt instruments." While they can identify whether an individual is overweight, they fail to distinguish between muscle and fat or identify the specific distribution of adipose tissue. This distinction is clinically vital: Visceral Adipose Tissue (VAT), which surrounds internal organs, is a significantly stronger predictor of type 2 diabetes and cardiovascular disease than Abdominal Subcutaneous Adipose Tissue (ASAT), which resides just beneath the skin.

While MRI and CT scans are the gold standards for measuring these fat volumes, their high cost and time requirements make them impractical for routine population-level screenings. This research presents a novel approach to "look inside" the body by using 3D surface geometry—captured via relatively inexpensive scanners—to accurately predict internal fat volumes using Graph Neural Networks (GNNs).

Methodology

The core innovation of this work lies in treating the human body surface as a graph rather than a 2D image. The researchers converted 3D body segmentations into triangulated surface meshes, where each vertex (node) contains spatial coordinates.

Graph Neural Network Architecture

To process these meshes, the study employed GraphSAGE Convolutions. This GNN variant learns node embeddings by sampling and aggregating features from a node’s local neighborhood, allowing the model to interpret the complex 3D "topography" of the human body. The researchers also performed mesh decimation, systematically reducing the complexity of the meshes from 10,000 faces down to 100 to test the model's robustness and efficiency.

Comparative Baselines

The GNN approach was benchmarked against state-of-the-art Convolutional Neural Networks (CNNs) and DenseNets. These baseline models operated on 2D silhouettes—binary outlines of the body—which represent the current standard for non-invasive shape-based fat estimation.

Dataset and Experimental Setup

The study utilized a massive dataset from the UK Biobank, consisting of 25,298 individuals (12,210 male and 13,088 female). Neck-to-knee MRI scans were algorithmically converted into 3D surface meshes to serve as input data. The ground truth labels for VAT and ASAT volumes were derived directly from the original MRI scans, providing a high-fidelity target for the supervised learning task.

Key Insights

3D Surface Signatures

The results demonstrate that 3D meshes are superior to 2D silhouettes. While silhouettes capture general body shape, they lose critical depth information. The GNN’s ability to utilize the full 3D surface allows it to identify subtle "surface signatures" of internal fat distribution that 2D models cannot detect.

Computational Efficiency

GNNs proved to be significantly more lightweight than deep CNNs. A GNN trained on a 1,000-face mesh reached an optimal trade-off between accuracy and speed, taking only approximately 10 minutes to train, compared to 96 minutes for the DenseNet baseline. This efficiency makes the method viable for resource-constrained environments or integration into portable medical devices.

Robustness to Resolution

Interestingly, the researchers found that high-resolution meshes are not strictly necessary for high accuracy. Even when the mesh was decimated to just 100 faces, the GNN achieved an $R^2$ of 0.858 for VAT, suggesting that global body shape carries the most vital information for internal fat estimation.

Results

The GNN approach outperformed the 2D silhouette baselines in both accuracy and training speed:

  • Accuracy: On a 10,000-face mesh, the GNN achieved an $R^2$ score of 0.893 for VAT and 0.941 for ASAT. In comparison, the DenseNet baseline achieved $R^2$ scores of 0.878 and 0.934, respectively.
  • Sex-Specific Trends: The models exhibited performance variations across sexes; VAT predictions were generally more accurate for men, while ASAT predictions were more accurate for women. This aligns with biological differences in fat storage patterns between sexes.

Conclusion and Future Work

This research demonstrates that GNNs applied to 3D surface meshes provide a highly accurate and efficient means of estimating internal body fat. While the current study utilized meshes derived from MRI data, the ultimate goal is to validate this model using data from low-cost optical 3D surface scanners. Future work may also explore more advanced GNN architectures, such as adaptive graph convolution pooling, to further refine accuracy and reduce the pre-processing overhead required to generate registered 3D meshes.

Summary updated

BibTeX
@inproceedings{mueller2023body,
  author = {Tamara T. Mueller and Siyu Zhou and Sophie Starck and Friederike Jungmann and Alexander Ziller and Orhun Aksoy and Danylo Movchan and Rickmer Braren and Georgios Kaissis and Daniel Rueckert},
  title = {{Body Fat Estimation from Surface Meshes Using Graph Neural Networks}},
  booktitle = {ShapeMI @ MICCAI},
  year = {2023},
  url = {https://openreview.net/forum?id=VXnpw0OQl9},
}
Figure