Graph neural networks can now estimate internal fat depots from the outer shape of the body. A graph neural network is a model that passes information between neighboring points on a mesh. Body Fat Estimation from Surface Meshes using Graph Neural Networks converts UK Biobank neck-to-knee MRI segmentations into triangulated surface meshes and predicts visceral fat (VAT, the fat around the organs) and abdominal subcutaneous fat (ASAT, the fat under the skin). It reports that VAT is predicted more accurately in men and ASAT more accurately in women. The usual reading is physiological: men store fat viscerally and women store it subcutaneously and in the hips and thighs. There is a less flattering possibility. These meshes are open tubes, cut at the neck and at the knees, and they are simplified uniformly everywhere. Points on a cut edge have about half the neighbors of interior points. Spatial-network theory has shown repeatedly that edge points like these dominate error in finite systems. Female metabolically relevant fat sits in the gluteofemoral region (hips and thighs), which runs into the knee cut. If boundary distortion falls unevenly on male and female anatomy, part of the reported sex difference measures how the scan was cropped, not how bodies differ.
This note proposes a controlled measurement study, followed by a minimal correction. The goal is to split the sex-specific predictability of fat depots from surface shape into a part that survives boundary-matched, density-matched conditions and a part that does not. The idea is simple to state. We move the cut planes, vary how the mesh is simplified, and give the model explicit knowledge of where the boundaries are. Then we watch whether the male-female accuracy gap moves. If it moves, current claims about sexual dimorphism from shape-based models need revisiting, and screening tools built on optical body scanners need boundary-aware design. If it does not move, we have the first evidence that the dimorphism in these models is physiological rather than geometric.
Core Claim
The sex-specific accuracy pattern of surface-mesh fat predictors depends on two arbitrary modeling choices: where the mesh is truncated, and how uniformly it is simplified. Both choices are predictably asymmetric across sexes because of where each sex stores fat. We claim this dependence is measurable with simple perturbations. We also claim that a cheap boundary-aware model removes most of the sensitivity: one that knows each vertex's distance to the cut and normalizes neighbor aggregation by lost neighborhood. If that model still shows a sex gap, the gap is the honest one.
Why the Spatial-Network Literature Bears on This
The motivating evidence comes from finite wireless networks. Results on Finite Wireless Sensor Networks: Connectivity and Coverage splits a bounded region into interior, edge and corner zones, and shows that edge nodes, which lose roughly half their coverage, drive most coverage failure. At 100 nodes, its asymptotic predictions miss disconnection probability by ten orders of magnitude. Connectivity in Secure Wireless Sensor Networks under Transmission Constraints similarly shows that ignoring edges underestimates the transmission range that boundary nodes need. Together these say that edge effects are first-order in finite graphs. A 100- to 1,000-face body mesh is a finite graph.
One paper constrains the argument. Cube Versus Torus Models and the Euclidean Minimum Spanning Tree Constant proves that boundary effects vanish as point count grows, because surface grows more slowly than volume. That suggests dense meshes should be immune. We treat this as a testable prediction: boundary sensitivity should shrink from 100 to 10,000 faces. The finite-network results warn that the shrinkage may be slow. The correction method draws on Parametric estimation for Gaussian fields indexed by graphs, which removes edge bias in finite-subgraph estimation with a boundary correction matrix that reweights terms near the edge. Its proofs require a homogeneous graph, which a body mesh is not. We therefore borrow only the idea of reweighting by lost neighborhood, implemented as degree-aware normalization in message passing, and we test it empirically rather than claiming its guarantees transfer.
The second lens is non-uniform density. A Novel Range-free Localization Algorithm for Anisotropic Networks shows that when points are unevenly spread or blocked by obstacles, graph hop counts stop tracking physical distance. Uniform mesh simplification creates exactly this mismatch. It spends faces evenly, so high-curvature transitions get a few coarse triangles, such as the waist, the gluteal fold and the hip-thigh junction. Deep learning-enabled MRI phenotyping uncovers regional body composition heterogeneity and disease associations in two European population cohorts shows that these regions carry the opposing signals: VAT raises type 2 diabetes odds (OR 1.19), and gluteofemoral fat (GFAT) lowers them (OR 0.69). The finite-network papers assume uniform points in convex regions, and the body-composition papers ignore boundaries entirely. This project sits where the two meet.
Research Program
All work uses UK Biobank neck-to-knee MRI with existing depot labels, following the mesh construction of Mueller et al. We first build a boundary-perturbation suite. For each subject, we regenerate meshes with the lower cut plane moved upward in fixed steps, removing thigh tissue that never enters the target VAT or ASAT volumes. We also regenerate them with the upper cut lowered at the neck. Against these truncation conditions we set a capped condition, in which the open ends are closed with a flat disc so no vertex is a boundary vertex. This is the analogue of the torus "no-edge" control from network theory. Second, at matched face counts of 100, 1,000 and 10,000, we compare uniform simplification against curvature-adaptive simplification, which spends more faces where the surface bends. The error-driven graph augmentation in mise_gnn offers one implementation route; no reported result was found on Lacuna for it on anatomical meshes. Third, we define the boundary-aware model. It is GraphSAGE with two additions per vertex: geodesic distance to the nearest cut, and aggregation normalized by the vertex's degree relative to the interior median.
The main quantity is per-sex sensitivity. For each target depot (VAT, ASAT, GFAT, and the VAT/ASAT ratio), we compute the slope of held-out $R^2$ against cut displacement, separately for men and women. We also compute what fraction of each sex's prediction error is attributed, by gradient attribution, to vertices within a fixed geodesic band of a cut. Our hypothesis is specific. Female ASAT and GFAT predictions should be more sensitive to the lower cut than male VAT predictions. Uniform simplification at 100 faces should hurt female predictions more than male ones, and the gap should narrow when the boundary-aware model is used.
Baselines
The naive reference is a sex-stratified linear model on BMI and waist circumference. Silhouette images enable estimation of body fat distribution and associated cardiometabolic risk reports BMI-based models at $R^2$ 0.608 to 0.833 for depot volumes. Anthropometric models reach only 0.171 (men) and 0.262 (women) for the VAT/ASAT ratio. Every mesh model should clear these easily. Their role is to anchor how much of the sex gap exists even without any geometry.
The strongest shape-based competitor is the unmodified GraphSAGE model of Mueller et al., reproduced with their mesh pipeline. It reports $R^2$ 0.893 for VAT and 0.941 for ASAT at 10,000 faces, and 0.858 for VAT at 100 faces. These two figures compare depots, not sexes. The paper reports a sex difference qualitatively, and no reported result was found on Lacuna for its per-sex $R^2$ values. We will therefore measure per-sex numbers ourselves under the original protocol before anything else. The second strong competitor is the silhouette DenseNet-121 from the silhouette paper. It reports $R^2$ 0.885 (VAT), 0.934 (ASAT) and 0.932 (GFAT), and 0.553 (men) and 0.504 (women) for the VAT/ASAT ratio. We render silhouettes from the same MRI segmentations so that both models see identical subjects. All models share the same train/test split, stratified by sex and age, and the same hyperparameter search budget. Each model is tuned separately for each truncation condition.
The key ablation removes the distance-to-cut feature and the degree normalization one at a time, and pairs the capped mesh with plain GraphSAGE. This isolates whether boundary knowledge or simply the removal of boundaries does the work. Beating the competitors here is not mainly about headline $R^2$. The success criterion is that the boundary-aware model's per-sex $R^2$ is flat across cut displacements, within bootstrap error, where plain GraphSAGE is not. The secondary criterion is that it exceeds 0.553 and 0.504 on the VAT/ASAT ratio in the respective sexes.
Why not just…?
Why not just train separate models for men and women, or feed sex as an input? This addresses calibration, not cause. A sex-stratified model would still absorb cut-plane artifacts into its sex-specific weights, and the reported gap would remain uninterpretable. We include sex-stratified versions of every model. If stratification alone makes per-sex sensitivity to cut displacement disappear, the boundary-aware machinery is unnecessary, and we will report that.
Why not just augment training with random crop jitter? This is the most serious threat. Training on randomly shifted cut planes may teach plain GraphSAGE to ignore edges as well as our explicit features do. We will run crop-jitter augmentation as a full baseline under the same budget. If it matches the boundary-aware model's flat sensitivity curves, it becomes the recommended fix. The diagnostic finding about the sex gap stands either way.
Why not just use denser meshes, since Jaillet says edges wash out? The 10,000-face condition tests this directly. Finite-network results such as Eslami et al. suggest the convergence is far slower than asymptotics imply. Mueller et al.'s finding that 100 faces still gives VAT $R^2$ 0.858 shows global shape carries most of the signal. That observation concerns pooled accuracy, though, and says nothing about whether the residual error sits at the edges and differs by sex.
Why not just use the 2D silhouette model? A silhouette also has a bottom edge at the knee crop, so it shares the truncation problem without a way to represent it. We will apply the same lower-cut perturbation to silhouettes. If silhouettes prove insensitive while meshes do not, that is itself evidence that the mesh boundary matters.
Decisive Checks
The central test asks whether the male-female difference in held-out $R^2$, per depot, shrinks by a meaningful amount, judged against bootstrap confidence intervals, when we move from the original open, uniformly simplified mesh to the capped or boundary-aware condition. If the difference is unchanged across every perturbation, the physiological reading is supported, and that is a publishable negative result.
A second check uses NAKO, the German National Cohort, as an external cohort. We first confirm whether its imaging field of view ends at different anatomical levels than UK Biobank's. If it does, a boundary-aware model trained on UK Biobank should lose less accuracy on NAKO than plain GraphSAGE does, with the larger benefit expected in women.
The third check is downstream. Predicted VAT/ASAT ratios from each model enter sex-specific logistic regressions for type 2 diabetes. We compare the resulting per-SD odds ratios against the silhouette paper's 1.70 (men) and 1.74 (women), to test whether removing boundary artifacts strengthens, weakens or merely relabels the sex-specific risk signal.
References
- Body Fat Estimation from Surface Meshes Using Graph Neural Networks
- Results on finite wireless sensor networks: Connectivity and coverage
- Connectivity in secure wireless sensor networks under transmission constraints
- Cube Versus Torus Models and the Euclidean Minimum Spanning Tree Constant
- Parametric estimation for Gaussian fields indexed by graphs
- A Novel Range-free Localization Algorithm for Anisotropic Networks
- Deep learning-enabled MRI phenotyping uncovers regional body composition heterogeneity and disease associations in two European population cohorts
- Silhouette images enable estimation of body fat distribution and associated cardiometabolic risk