Brain imaging genetics increasingly relies on neural networks to turn MRI scans into cortical surfaces. These surfaces are the folded inner and outer boundaries of the brain's gray matter, and quantities read off them, such as local cortical thickness, become the traits that genetic studies analyze. The newer reconstructors are neural implicit surface models. Instead of storing a surface as a mesh, they train a network that takes any 3D point and returns its signed distance to the surface, and the surface is wherever that distance is zero. These models are fast and precise, but they are trained almost entirely on cohorts of European ancestry, and their topology-repair steps pull each reconstruction toward a standard template. Nobody has checked whether their errors grow for people whose brains differ systematically from that training distribution.
This matters because of what happens downstream. If reconstruction error depends on ancestry, it acts as a hidden, ancestry-correlated measurement error in exactly the traits that cross-ancestry genetic studies compare. It could create spurious population differences in anatomy. It could also make polygenic prediction of brain traits look less portable to non-European groups than it really is. We propose to measure this error against known ground truth, trace where it enters the pipeline, and test whether a targeted fix removes it without also erasing real anatomical variation.
Core Claim
Our hypothesis is that learned signed-distance cortical reconstructors carry an ancestry-dependent error component that is invisible in standard benchmarks. We expect it to be concentrated in deep, narrow folds and in the template-based topology-correction step. We further expect it to measurably reduce the heritability and cross-ancestry polygenic prediction of vertex-level cortical thickness in non-European participants.
The contribution is an audit that uses known ground truth, a causal decomposition of where the error comes from, and a test of whether that error distorts genetic conclusions. We do not propose a new "ancestry-disentangled" shape model. Such a model would be the obvious combination of these fields, but it would remove genuine population differences in anatomy along with the bias.
Relation to Prior Work
The method-supplying work is the implicit cortical reconstruction line. DeepCSR predicts signed distance fields from MRI using features sampled at arbitrary coordinates, then applies spherical level-set topology correction. Hybrid-CSR passes explicit mesh deformation through a differentiable Poisson implicit field and restores sphere-like topology by registration. "Recon-all-clinical" trains a U-Net on synthetically randomized images to predict signed distance fields from heterogeneous clinical scans. All three report accuracy on ADNI, OASIS, HCP or test-retest data without stratifying by ancestry. We reuse their architectures unchanged as the objects under audit.
Recon-all-clinical also supplies a key tool: its synthetic image generator, which lets us create scans whose true surface is exactly known.
The motivating work comes from genetics. DisPred shows that ancestry-correlated structure degrades genetic risk prediction in AFR, admixed and SAS groups, and removes it by splitting a SNP autoencoder's hidden representation into ancestry and disease parts. We borrow the concern but not the remedy. Applied to anatomy, their split would remove true population differences, so we target measurement error instead.
GIANT shows that the way the brain is carved into regions changes genetic discovery power. Its heritability-informed regions yield 61.72 significant SNPs per region versus 36.59 for MUSE, and polygenic scores explain 3.74% versus 2.64% of volume. That result tells us measurement choices upstream of genetic analysis have first-order effects, which is the premise we test for surface reconstruction. The genetic architecture of multimodal human brain age constrains our design. It found linear LASSO beating deep models on regional summaries, and it worked in a predominantly European cohort. We therefore keep the downstream genetic analyses deliberately standard and linear, so any effect is attributable to the surface measurement rather than to modeling choices.
Finally, Neural ADMIXTURE supplies continuous ancestry proportions in a single forward pass. This lets us treat admixed individuals as points on a continuum rather than forcing them into discrete groups.
Research Program
Ground-truth audit. The first component is an audit in which the true surface is known.
We take multi-ancestry UK Biobank imaging participants. Ancestry is assigned by Neural ADMIXTURE proportions, and groups are matched on age, sex, intracranial volume and imaging site. The matching matters because head size and age plausibly drive reconstruction error by themselves.
For each participant we take the reference surfaces and generate synthetic T1 scans from them with the Recon-all-clinical generator, at both isotropic and thick-slice settings. Because each image is rendered from a known surface, reconstruction error on it is measured exactly, with no dependence on FreeSurfer's own template assumptions. This separates "the network is worse on this anatomy" from "the labels are worse on this anatomy."
We then run DeepCSR, Hybrid-CSR, Recon-all-clinical and FreeSurfer. For each, we regress per-vertex error on continuous ancestry proportions, controlling for the matched covariates.
Localizing the source. The second component traces where the error enters, using ablations on the same data.
We compare error before and after topology correction to isolate the template's contribution. We measure the displacement that the correction step induces, which is a real-data quantity requiring no ground truth, along with the number of topological defects before correction.
We retrain DeepCSR on training sets with controlled ancestry composition at a fixed total size. This tells us whether the gap tracks the composition of the training data.
We also test the hypothesis that local fold geometry mediates the error, rather than ancestry as such. To do so, we check whether sulcal depth and curvature account for the ancestry effect.
Downstream consequences. The third component asks whether the error changes genetic conclusions.
Vertex-wise thickness from each pipeline is used in ancestry-stratified SNP heritability estimates. We also train polygenic scores in Europeans and evaluate them in other groups, and compare the resulting portability across pipelines. If reconstruction error is ancestry-dependent, pipelines with larger error gaps should show lower apparent heritability and portability in non-European groups.
Baselines
Naive reference. The do-nothing reference is the off-the-shelf DeepCSR and Recon-all-clinical models exactly as released, trained on their original ADNI and HCP-dominated data. This is the configuration current imaging-genetics users would actually deploy, so any measured gap is a gap in practice.
Strongest existing method. On accuracy, the strongest competitor is Hybrid-CSR. It reports about 0.20 mm pial average symmetric surface distance versus about 0.26 mm for competing methods on ADNI. Recon-all-clinical reports errors generally below 0.4 mm even at 6 mm slices, and parcellation Dice above 0.90. We found no reported result on Lacuna for ancestry-stratified cortical reconstruction error for any method, so these aggregate figures are the only anchors.
FreeSurfer is included as the classical template-based reference. It is not neural, but it shares the template-driven topology repair we suspect.
Fair comparison. All methods see identical synthetic and real inputs. Retrained variants use the same architecture, iteration budget and hyperparameter search. A fix counts as beating the baselines only under two conditions. First, it must shrink the ancestry-conditional error gap to within its bootstrap confidence interval of zero. Second, it must not raise overall error above the released model's.
Ablation. The key ablation isolates topology correction. We use the same trained network with and without the template-based correction, and with a template-free correction that enforces sphere-like topology using only a defect-local repair step.
Why not just…?
Why not just add ancestry principal components as covariates in the genetic analysis? This removes ancestry-linked shifts in the mean of the measured trait. It does not remove ancestry-linked differences in measurement noise, which weaken genotype-trait associations in the noisier group. Adding the components also cannot tell a reconstruction artifact from real anatomy.
This alternative is a genuine threat if the error turns out to be a pure mean offset. Our audit settles it directly. We decompose the ancestry effect on error into a mean shift and a variance change. If only the shift is present and principal-component adjustment restores heritability parity, we will report that the simple adjustment suffices.
Why not just keep using FreeSurfer, which guarantees correct topology? FreeSurfer's topology repair is itself template-guided, so it is a suspect rather than a control. It is also far slower: DeepCSR reports at least a 13× speedup. We audit FreeSurfer with the same ground-truth protocol, so the proposal answers this question rather than assuming the answer.
Why not just retrain on an ancestry-balanced dataset? This may well be sufficient, and it is one of our experimental arms rather than a competitor we argue against. The composition-controlled retraining tells us whether the gap scales with training share.
If balancing closes the gap, the practical recommendation is simple: rebalance. If a residual gap persists and it disappears when template-based correction is removed, the template is the culprit, and rebalancing alone cannot fix it.
Why not just disentangle ancestry out of the shape representation, as DisPred does for SNPs? Population differences in anatomy are signal, not noise, for cross-ancestry genetics. A reconstructor trained to be ancestry-invariant would systematically erase them. The synthetic audit would expose this failure: an ancestry-invariant model should show increased error on anatomies that genuinely differ from the training average.
Proposed Experiments and Decisive Checks
Primary test. The primary test is the per-vertex regression of synthetic-ground-truth error on continuous ancestry proportions, for each pipeline, with matched covariates. We report the effect by region and by fold depth.
Placebo check. We shuffle ancestry labels within matched strata. The resulting gap must vanish; if it does not, confounding remains and we stop before interpreting further.
Mediation check. We ask whether local curvature and sulcal depth absorb the ancestry effect. If they do, the bias concerns fold geometry and should be fixed by oversampling hard geometry, not by ancestry labels.
Downstream check. This is the check that determines whether the work matters for genetics. We compare ancestry-stratified heritability and European-to-other polygenic-score portability of vertex-wise thickness, derived from the released model versus the best-debiased variant, on the same individuals.
If the debiased surfaces raise non-European heritability or portability beyond sampling error, reconstruction bias is shown to be a real contributor to the apparent non-portability of imaging traits. If no pipeline shows an ancestry-conditional gap, that negative result is also useful. It would license the use of learned reconstructors in multi-ancestry biobanks and would redirect portability concerns toward linkage and effect-size differences in the genetics itself.
References
- DeepCSR: A 3D Deep Learning Approach for Cortical Surface Reconstruction
- Hybrid-CSR: Coupling Explicit and Implicit Shape Representation for Cortical Surface Reconstruction
- "Recon-all-clinical": Cortical surface reconstruction and analysis of heterogeneous clinical brain MRI
- Improving genetic risk prediction across diverse population by disentangling ancestry representations
- A genetically informed brain atlas for enhancing brain imaging genomics
- The genetic architecture of multimodal human brain age
- Neural ADMIXTURE: rapid population clustering with autoencoders