Using cortical geometry as a label-free signal for adapting fast surface reconstruction to new MRI scanners

Fast neural networks can now turn a brain MRI into explicit meshes of the cortex in seconds instead of the hours classical software needs. The cortex is a thin, folded sheet. These meshes trace its inner and outer boundaries, and from them one can measure cortical thickness and surface area separately, two quantities that a simple gray-matter volume blurs together. The difficulty is that MRI intensities have no physical calibration, so a network trained on one vendor's scanner can fail quietly on another. Nobody has measured what this does to a reconstructed cortical surface. A broken mesh still yields thickness numbers. They are simply wrong, and summary distance metrics may not show it.

Our proposal starts from one observation. A correct cortical reconstruction has to satisfy geometric facts that do not depend on the scanner. The sheet is a closed surface with no holes or handles. The outer boundary lies outside the inner one. Thickness falls within a narrow biological range. The mesh does not pass through itself. All of these can be checked on any new scan without labels. We propose to use violations of these facts as an unsupervised loss at test time. The loss adapts only the first stage of a two-stage reconstruction pipeline: the 3D network that segments the white-matter region from which the mesh is grown. We will test whether this recovers trustworthy thickness and area measurements on unseen scanners, and whether it does so without erasing real disease effects.

Core Claim

In two-stage reconstructors such as CortexODE, most of the scanner-induced error enters through the upstream white-matter segmentation and not through the mesh-deformation stage. CortexODE first segments white matter with a 3D U-Net, extracts and topology-corrects an initial mesh, and then flows it to the white and pial surfaces with a smooth neural ordinary differential equation, meaning a learned velocity field integrated over time. Its authors note that the quality of the initial surface strongly affects the final result.

If that diagnosis holds, the fix should target the segmenter. A small test-time update of the segmenter, driven only by the geometric validity of the surfaces it produces, should close most of the cross-scanner gap in morphometry. That gap is measured by thickness error, area error, cross-scanner agreement, and retained disease effect size. We claim this approach outperforms intensity augmentation, generic entropy-based test-time adaptation, and post-hoc statistical harmonization. We also claim it approaches fine-tuning on labeled target data.

The contribution has two parts. The first is a measurement study, which the research file confirms has not been done for deep surface reconstructors: where along the pipeline domain shift breaks cortical geometry. The second is a specific adaptation signal grounded in the biology of the cortical sheet.

Research Program

The first component is a failure audit. We train CortexODE, CorticalFlow (a multi-scale chain of 3D U-Nets that flows a template sphere to the cortex in about a second), and SegRecon (a single 3D U-Net that predicts signed distance functions of both surfaces plus parcellation) on a single-vendor source domain. We then measure how geometry degrades on unseen vendors, field strengths, and cohorts.

The key diagnostic is an oracle swap. On target scans we substitute a white-matter segmentation derived from FreeSurfer for the network's own segmentation, keep the flow network frozen, and record how much error disappears. If the swap recovers most of the error, the error is upstream. If it does not, our premise is wrong, and the adaptation target should move to the flow network.

The audit will also count how many topological defects the initial mesh contains before correction. This follows the concern raised by the FeTA 2024 challenge (FeTA 2024), which found that methods with near-identical overlap scores differed by roughly 50% in Euler characteristic difference. The Euler characteristic difference is a count-based measure of spurious handles, holes, and disconnected pieces. That finding suggests distance metrics will understate the damage, and the audit reports both kinds of metric side by side.

The second component is the adaptation loss. It is applied per target site, or per scan, to the segmenter's normalization and late-layer parameters, with the flow network frozen. It sums differentiable penalties computed on the reconstructed surfaces:

$$ \mathcal{L}_{geo} = \lambda_1 \sum_v \mathrm{softplus}(|t_v - c| - h) + \lambda_2 \sum_v \mathrm{relu}(-d_v) + \lambda_3 \, C_{self} + \lambda_4 \, \mathrm{KL}(\hat{p}_{area} \,\|\, p_{area}) $$

The terms are defined as follows.

  • $t_v$ is the thickness at vertex $v$, the distance between corresponding inner and outer points. The first term penalizes thickness outside a band of centre $c$ and half-width $h$, set to the 0–5 mm range that SegRecon uses as a training constraint.
  • $d_v$ is the signed offset of the pial vertex along the white-surface normal. The second term penalizes pial points that fall inside the white surface.
  • $C_{self}$ is a soft collision penalty between nearby faces of the mesh that are not neighbours.
  • $\hat{p}_{area}$ is the scan's distribution of per-vertex area expansion, and $p_{area}$ is the same distribution on source training data. The fourth term is a weak prior that discourages a degenerate solution in which the segmenter shrinks the cortex until it becomes trivially valid.

Gradients reach the segmenter through the differentiable signed-distance representation that seeds the mesh. Only the initial genus-zero topology correction stays non-differentiable, as in both CortexODE and SegRecon.

Volumetric context enters as a deliberate stressor rather than as an ingredient. Clinical scans often have thick slices, and surface reconstructors are tuned to 1 mm isotropic data, as the research file notes for SegRecon. We therefore resample target scans to 5 mm through-plane spacing and interpolate back. This tests whether geometric adaptation still helps when the 3D network has lost through-plane information. It is the regime in which slice-wise processing is known to produce jagged, disconnected structures (Deep convolutional networks for pancreas segmentation in CT imaging).

Baselines

The naive reference is each reconstructor trained on source data and applied unchanged to target data. This is the number every other method must beat. The research file contains in-distribution figures only. CortexODE reports an average symmetric surface distance below 0.2 mm and fewer than 0.4% self-intersecting pial faces across ADNI, HCP, and dHCP. SegRecon reports 88.69% parcellation Dice on MindBoggle-101 with average Hausdorff distance under 0.52 mm for the white surface and under 0.97 mm for the pial surface. There is no reported result on Lacuna for any deep cortical surface reconstructor under cross-scanner shift, so the audit itself establishes the unadapted target-domain numbers.

The strongest unsupervised competitor is heavy intensity augmentation during source training: random bias fields, gamma and contrast curves, noise, and resolution degradation. This is the recipe behind the top FeTA 2024 baseline, 3D nnU-Net with extensive spatial and intensity augmentation. It is also the first thing a reviewer will reach for. We will give it the same architecture, the same number of training steps, and the same hyperparameter-search budget as our method. Our adaptation is applied on top of the augmented model as well as the plain one, so the comparison is "augmentation alone" against "augmentation plus geometric adaptation."

The second competitor is generic test-time adaptation of the same segmenter parameters by minimizing prediction entropy. It uses the same number of update steps and the same learning-rate grid. This baseline isolates whether the cortical-geometry signal matters or whether any unsupervised update would do.

The third is the statistical route: unadapted reconstruction followed by ComBat harmonization of regional thickness and area. This is the approach SCSR relies on for multi-site normative modelling of FreeSurfer thickness maps.

The fourth is a volume-only pipeline: the Computational Anatomy Toolbox with projection-based thickness. Righart et al. report that it correlates with FreeSurfer at r = 0.89 vertex-wise but is less sensitive to aging, with r = −0.21 versus −0.36. It cannot produce vertex-wise area.

The upper reference is supervised fine-tuning on 10–20% of target scans labelled with FreeSurfer pseudo-ground-truth. This is the mix-in fraction that the scanner domain shift study found necessary to restore lost performance, although that study measured classification AUC, not geometry.

The ablations that isolate our key ingredient are three:

  • adapting the flow network instead of the segmenter;
  • dropping each loss term in turn;
  • replacing the hand-specified geometric terms with a learned normative thickness prior in the style of SCSR.

Beating the augmentation baseline means a statistically significant reduction in target-domain distance error and thickness error, a lower count of initial-mesh topological defects, and higher same-subject agreement across scanners, all at matched compute. We consider the method practically successful if it recovers at least half of the gap between the naive model and the supervised fine-tuning reference.

Why not just…?

Why not just train with aggressive intensity and contrast randomization? This is the real threat, and we treat it as one. The research file shows that weak augmentation fails. Gaussian noise did not rescue cross-vendor MRI performance (scanner domain shift study), and extensively augmented nnU-Nets in FeTA 2024 still lost about 0.075 Dice across sites and about 0.10 across image-quality levels. However, the file contains no evidence on strong contrast randomization applied to surface reconstructors. We therefore stack our method on the strongly augmented model. If the residual gain is negligible, the honest conclusion is that augmentation suffices, and the paper becomes the audit plus that negative result.

Why not just fine-tune on a few target scans processed with FreeSurfer? Where FreeSurfer works on the target scanner, this is cheap and probably strong, which is why it is our upper reference rather than a straw man. Its weakness is the regime that matters most clinically. On thick-slice and low-quality scans, FreeSurfer's own output degrades, and the fine-tuned network would inherit those errors, the supervision bias the file flags for all deep surface models. The anisotropic-resampling experiment directly tests whether pseudo-label fine-tuning or label-free geometric adaptation degrades more gracefully.

Why not just harmonize the output measures with ComBat? ComBat corrects additive and multiplicative site offsets in tabular values. It cannot repair a mesh with a handle bridging a sulcus or a truncated gyrus, and it removes site effects using covariates that may themselves be confounded with diagnosis. The same-subject cross-scanner experiment distinguishes the two approaches. If ComBat matches our agreement scores while retaining the Alzheimer's effect size, then geometric adaptation is unnecessary for regional statistics, though it may still matter for vertex-wise maps.

Why not just compute thickness in the volume with CAT and skip meshes? Volume-based projection thickness gives no vertex-wise surface area and no folding measures. According to Joint Analysis of Cortical Area and Thickness and the depression sex-differences study, separating area from thickness is exactly what yields statistical sensitivity. CAT is also slower, at about an hour per scan, and its own robustness across scanners is not reported in the file. We include it on the thickness metrics only.

Proposed Experiments

The source domain is a single vendor at a single field strength within ADNI. The target domains are three: the remaining ADNI vendors and field strengths, including the subset of subjects with scans at both 1.5T and 3T where available; OASIS and the multi-site ABIDE-I; and MindBoggle-101, whose manual labels give an evaluation not tied to FreeSurfer.

Metrics are reported per target domain and fall into four groups:

  • surface distance (average symmetric surface distance and Hausdorff distance, both against FreeSurfer);
  • topological validity (self-intersection rate, and handles plus holes in the initial mesh before correction);
  • morphometric fidelity (vertex-wise thickness and area error, and intraclass correlation of regional thickness and area for the same subject on two scanners);
  • clinical fidelity (the Alzheimer's-versus-control effect size for entorhinal and temporal thinning, compared with FreeSurfer's effect size on the same subjects).

Three outcomes decide the project. The first is the oracle swap. If replacing the white-matter segmentation removes most target-domain error, the upstream-error premise holds. If not, the adaptation moves to the flow network, and the ablation reports this. The second is the comparison with the strongly augmented baseline at matched compute, judged on cross-scanner agreement of thickness and area, not only on distance metrics. The third is the disease-retention test. Adaptation must not shrink the Alzheimer's effect size relative to unadapted reconstruction. This test matters most for the learned-normative-prior ablation, because a prior trained on healthy brains could pull atrophic cortex toward normal.

The anisotropic stress test then shows whether any gain survives once volumetric context is degraded. The method has to work on routine clinical scans as well as on research-grade ones, so this result determines whether it is useful in practice.