Whole-Body Shape Attention for First-Trimester Visceral Fat Screening

Early pregnancy risk models for gestational diabetes still take body size as pre-pregnancy body mass index (BMI), a number that cannot tell a woman carrying inflammatory fat around her organs from one carrying muscle or harmless fat under the skin. The fat around the organs, called visceral fat, is the metabolically dangerous kind, but measuring it takes CT or MRI. Neither is used for routine first-trimester screening: CT involves radiation, and MRI is costly. Recent work shows that the outer shape of the body, captured by a 3D surface scan, predicts internal fat depots surprisingly well. That work, however, used only non-pregnant adults. We propose to test whether a cheap body-surface scan taken at 8–14 weeks of gestation can estimate visceral fat accurately enough to improve early gestational diabetes prediction beyond BMI and routine blood markers.

The modeling bet is specific. Visceral fat is not visible at any one spot on the body surface. It shows up as a relationship between distant regions: a waist that is large relative to the hips, thighs and limbs. Clinicians already encode this by hand in ratios such as the weight-adjusted waist index and A Body Shape Index. We argue that attention, a network layer that lets every body region compare itself directly with every other region, is the natural learned version of these ratios. We also argue that an earlier negative result for attention came from an input that contained no distant regions to compare. The simulation direction plays a supporting role. Early pregnancy slightly deforms the lower abdomen, and we use a fast learned stand-in for a slow soft-tissue simulation to generate pregnancy-like deformations for training. We treat this as an optional component and include a test that decides whether it is needed at all.

Core Claim

Our claim is that attention over whole-body anatomical regions beats both local convolutional networks and the best existing nonlinear shape regressor for estimating visceral fat, but only when the input spans the whole body. On abdomen-only input it should confer no advantage, and the gain should be largest in women. Two findings in the file motivate this and seem to contradict it.

The first is AbdCTBench. It built 23,506 CT-derived abdominal surface meshes rendered as 2D depth maps and found that compact ResNets matched or beat ViT and Swin transformers at predicting 16 CT biomarkers. We read this as a statement about the input rather than about attention. A depth map of the abdomen alone contains no hips, thighs or arms, so there is nothing distant for attention to relate.

The second is 3D convolutional deep learning for nonlinear estimation of body composition from whole body morphology. It encodes 6,890-vertex whole-body meshes with a graph-convolution autoencoder and regresses DXA body composition with a Gaussian process, reaching R² above 0.86 on all variables in the Shape Up! Adults dataset. In females, however, its learned features did not outperform a Gaussian process on raw coordinates. Female fat distribution is dominated by the hip and thigh, so separating visceral from gluteofemoral fat plausibly requires comparing the abdomen with regions far from it. Local graph convolutions do this poorly and attention does it directly.

The application claim follows from the second part of the program. We claim that a scan-derived visceral fat estimate, added to the first-trimester model of Chen et al., improves discrimination for later gestational diabetes. That model combines family history, pre-pregnancy BMI, progesterone, AST, APTT and the triglyceride-to-HDL ratio, and its AUC falls from 0.85 in training to 0.73 in validation. The improvement should come chiefly from normal-BMI women with high visceral fat, the group that the tissue-type literature (Mo et al., Sia et al.) says BMI hides.

Research Program

The first stage is architecture, and it runs on existing data. We partition the standardized 6,890-vertex mesh used in CAESAR and Shape Up! Adults into anatomically consistent patches: abdomen bands, flanks, hips, thighs, upper arms, chest and neck. Each patch is encoded by a small local geometric network so that local curvature detail is kept. A transformer then attends across the patch tokens. Keeping a local encoder under a global mixer answers the inductive-bias worry raised by AbdCTBench, because fine curvature is not handed wholesale to attention. With about 30 to 60 tokens, the quadratic cost of attention is trivial, so we need none of the token-compression machinery of the turbulence surrogate of Nafi et al.. We borrow that paper's argument, not its tokenizer. Their surrogate used attention because pressure at a point depends on the whole domain. We are making the analogous claim that visceral fat is a whole-body quantity.

We pretrain on DFAUST pose variation, following the recipe of Tian et al., in which pretraining cut reconstruction error from 22.7 mm to 2.0 mm. We then fine-tune on Shape Up! Adults, with Shape Up! Kids included as that paper did, against DXA visceral fat, fat-free mass and the other targets. AbdCTBench serves as a second, abdomen-only testbed with CT-derived depot labels.

The second stage handles pregnancy deformation. Starting from CT body segmentations of the kind produced by Comp2Comp, whose Dice scores are about 0.96 to 0.97 for visceral fat, subcutaneous fat and muscle, we build a finite-element soft-tissue model. It enlarges the uterus and adds modest subcutaneous and fluid expansion across gestational weeks 8 to 16, then records the resulting skin-surface displacement. Each such simulation is slow. Rather than run thousands of them, we compress displacement fields onto a low-dimensional basis. We start with the linear projection used in reduced-order modeling (Constantine et al.) and move to a nonlinear autoencoder manifold of the kind used by Chen et al. for material point simulations only if the linear basis fails. We then fit a regressor from gestational week and uterine size to latent coordinates, which produces deformations in milliseconds for training-time augmentation.

AirfRANS (Bonnet et al.) supplies a constraint here. Its surrogates reproduced bulk fields well but smoothed away wall gradients, and their drag correlations became poor or negative. Our target is surface curvature, which is itself a wall quantity. We therefore evaluate the surrogate on vertex-wise curvature error against held-out finite-element runs, not only on displacement error.

The third stage is a prospective cohort, because no pregnant surface-scan data exist in the file. We will enrol roughly 1,000 singleton pregnancies at 8–14 weeks. Each participant receives a smartphone depth or LiDAR scan and the standard first-trimester labs used by Chen et al. A subset of about 200 also receives non-contrast abdominal MRI, which is acceptable in pregnancy, with visceral and subcutaneous fat segmented as the depot reference. The outcome is a 24–28 week oral glucose tolerance test. Chen et al. report a prevalence of about 25% in their cohort; at that rate, 1,000 participants yield roughly 250 cases, enough to detect AUC differences on the order of 0.05 in a paired comparison. Preeclampsia is recorded as a secondary outcome. The leptin/ceramide work (Huang et al., AUC 0.92 in only 40 women) is too small to anchor a primary endpoint.

Baselines

The naive reference point is pre-pregnancy BMI together with the hand-built shape ratios computed from the same scan: weight-adjusted waist index, A Body Shape Index, waist-to-height ratio and body roundness index. For depot estimation these enter a linear regression. For gestational diabetes they enter logistic regression alongside the laboratory panel. These baselines use the same participants and the same splits. If the learned model cannot beat hand-built long-range ratios computed from the very same scan, the attention claim fails.

The strongest existing method for depot estimation is the graph autoencoder plus Gaussian process regressor of Tian et al.. It reports R² above 0.86 for all predicted variables on Shape Up! Adults, up to 20% lower error than linear regression, and visceral fat precision within 1–2 times that of DXA. We rerun it ourselves from the same pretraining data (DFAUST, CAESAR) with the same fine-tuning splits, and give it a hyperparameter search budget equal to ours. We also include its raw-coordinate Gaussian process variant, which was the better model in females. Beating it means a lower visceral fat error on the held-out female subgroup, with a bootstrap confidence interval that excludes zero, and non-inferiority in males. For the abdomen-only setting, the comparison is the best CNN on AbdCTBench, retrained under our splits. For context, that paper reports AUROC 0.801 for type 2 diabetes and 0.839 for mortality, but the file does not state a visceral-fat-specific error for it.

For gestational diabetes, the strongest existing model is that of Chen et al., with a reported validation AUC of 0.73 (95% CI 0.62–0.83). We refit it on our cohort using its exact variables. Beating it means that adding the scan-derived visceral fat estimate raises out-of-sample AUC with a DeLong interval above zero and improves net reclassification within the normal-BMI stratum.

The ablation that isolates the key ingredient keeps the same local patch encoder and replaces cross-patch attention with pooling followed by a small multilayer perceptron. It is run on whole-body and on abdomen-cropped input. A second ablation removes the simulated pregnancy augmentation.

Why not just…?

Why not just use waist-to-height ratio and weight-adjusted waist index from a tape measure? These are the strongest cheap alternative, and they already capture the idea that visceral fat is a ratio between distant body regions. Tian et al. found that nonlinear shape models reduced error by up to 20% relative to linear anthropometric regression, but that comparison was in non-pregnant adults. Our naive baseline computes exactly these ratios from the same scan, so the cohort measures directly how much learning adds. If the gain in gestational diabetes AUC over the ratios is negligible, the honest conclusion is to recommend the tape measure, and we will report it that way.

Why not just use the Gaussian process on shape coordinates, or a CNN on depth maps, and skip attention? This is a real threat. AbdCTBench is direct evidence that local models are sufficient on abdominal input. The two-by-two design of architecture (attention versus pooling) crossed with input extent (whole body versus abdomen crop) is built to settle it. The attention claim survives only if attention helps on whole-body input, does not help on abdomen crops, and helps most in females. Any other pattern means global context is not the active ingredient.

Why not just apply a non-pregnant model directly and skip the simulation? At 8–14 weeks the uterus is still largely within the pelvis, so the deformation may be too small to matter. We test this directly. We compare MRI-referenced visceral fat error in the cohort for models trained with and without simulated deformation, stratified by gestational week. If the difference is null before 12 weeks, the simulation stage is dropped from the screening pipeline and kept only for extending the window later into pregnancy.

Why not just add a first-trimester lipid ratio and ignore body shape? The triglyceride-to-HDL ratio is already in Chen et al.'s model and carries a 5.51-fold odds per unit. The question is whether shape adds information beyond it, which is why every comparison is incremental over the full laboratory panel rather than over BMI alone.

Proposed Experiments

The first decisive check runs on existing data before any cohort money is spent. It is the architecture-by-input-extent experiment on Shape Up! Adults, reporting visceral fat mean absolute error and R² separately by sex, with AbdCTBench as the abdomen-only replicate. The prediction is a positive interaction: attention helps on whole-body input and not on crops, with the largest gain in females. A null interaction falsifies the core claim.

The second check tests the surrogate. Held-out finite-element deformations are compared against the reduced surrogate on displacement error and on vertex-wise curvature error in the lower abdomen. We also report wall-clock time per sample for both, which is the latency gain that makes augmentation feasible.

The third check runs in the cohort's MRI subset. It measures agreement between scan-estimated and MRI visceral fat through Bland-Altman limits and R², and asks whether error grows with gestational week.

The final and decisive check is the increment in gestational diabetes prediction over the refit Chen et al. model, judged by AUC, calibration slope and net reclassification in normal-BMI women. It is evaluated on a temporally held-out validation block to avoid the optimism that shrank their AUC from 0.85 to 0.73.