Several recent systems rebuild the folded surface of the human brain from a 3D MRI scan by smoothly morphing a sphere. They deform a template mesh, a fixed set of vertices and triangles, until it matches the cortex. Because the template is morphed rather than rebuilt from scratch, every output shares the same vertices. Vertex 10,000 in one brain corresponds to vertex 10,000 in another, and the surface never tears. The networks that decide how each vertex moves are local, however. A vertex's motion depends only on image features within a convolutional neighbourhood. The best-documented symptom is that retraining the same model with a different random seed sends vertices in hard regions, notably the lateral parietal cortex and lingual gyrus, to noticeably different places (Template-Based Cortical Surface Reconstruction with Minimal Energy Deformation).
We propose giving every vertex access to the whole surface without giving up the no-tearing guarantee. The idea rests on one observation. The fixed template supplies a permanent coordinate system, latitude and longitude on the original sphere, that travels with the vertices as they move. That coordinate system lets us borrow the "criss-cross" trick from 2D segmentation. Each vertex attends to its own latitude ring and longitude meridian, and two such passes connect any vertex to any other at roughly $N^{1.5}$ cost instead of $N^2$. The resulting per-vertex messages are then spread into a smooth 3D velocity field before the mesh is moved. That step preserves the mathematical guarantee that the deformation is a diffeomorphism, a smooth invertible map that cannot fold the surface through itself.
Core Claim
Existing work splits along a clean line. T-Pixel2Mesh runs full self-attention across all template vertices to capture global symmetry and occluded parts. It can afford this only at the coarsest stage, and it moves vertices with unconstrained offsets that cannot rule out self-intersection. CorticalFlow takes the opposite route. It integrates a velocity field predicted by a 3D UNet (a convolutional encoder-decoder network) as an ordinary differential equation, which makes the deformation diffeomorphic in principle. Each vertex, however, only reads local features at its current position. The architecture contains no vertex-to-vertex attention despite the paper's title.
Our claim is that these two properties are not in tension once two choices are made. First, global communication is routed along paths defined on the template sphere rather than on the moving geometry. Second, its output enters the dynamics only through a smooth ambient field. We further claim, as a falsifiable hypothesis, that a meaningful part of the seed-to-seed variance in high-curvature regions comes from missing long-range context. The alternative explanation is irreducible anatomical ambiguity, and the experiments are designed to tell the two apart.
Mechanism
Let the template have vertices $i = 1, \dots, N$ with fixed spherical coordinates $(\theta_i, \phi_i)$. We bin the vertices into about $\sqrt{N}$ latitude rings and $\sqrt{N}$ meridians. Pole crowding is handled by running a second ring system in a frame rotated by 90 degrees and averaging the two.
At integration step $t$, vertex $i$ holds a feature $h_i$. This feature combines image features sampled trilinearly at its current position $x_i(t)$, as CorticalFlow and Vox2Cortex do, with its current position. Two recurrent criss-cross passes over rings and meridians, following the recurrent design of CCNet, produce a globally informed feature $g_i$. Latitude rings are closed loops on the sphere, and meridians run pole to pole without wrap-around. There is therefore no arbitrary seam of the kind EccoMamba reports when flattening grids into 1D sequences for state-space scans.
The per-vertex messages are lifted to an ambient velocity:
$$v_t(x) = u_t(x) + \sum_{j} K_\sigma\big(x - x_j(t)\big)\, W g_j ,$$
Here $u_t$ is the usual local UNet velocity, $K_\sigma$ is a Gaussian kernel of width $\sigma$, and $W$ is a learned linear map. In practice the sum is computed by splatting onto a coarse grid and interpolating. The vertices then follow
$$\frac{d x_i}{dt} = v_t\big(x_i(t)\big).$$
Because $v_t$ is a sum of smooth terms whose Lipschitz constant is bounded by $\|W g\|$ and $1/\sigma$, the continuous flow is a diffeomorphism by the same argument CorticalFlow uses. We constrain step size relative to that bound to narrow the gap between theory and the discrete solver, which CorticalFlow identifies as its remaining source of self-intersections.
We keep the Minimal Energy Deformation (MED) loss from Madlindl et al., which penalizes total vertex path length, so that variance reduction from regularization and from context can be separated.
Research Program
The roles of the cited work are as follows. The reproducibility findings of Madlindl et al. and the regional variance they report motivate the project. CorticalFlow and V2C-Flow supply the flow backbone and the training protocol against FreeSurfer surfaces. CCNet supplies the factorized attention, and the Compact Generalized Non-local Network gives an alternative linear-cost global block that we test as a drop-in replacement for the ring attention. ViT-V-Net constrains the design. It shows that putting global attention in a volumetric bottleneck costs heavy GPU memory and coarse patches, which is why we place attention on the mesh rather than the voxels.
We will build the model on the V2C-Flow codebase setup. We will train on both hemispheres jointly, which Madlindl et al. did not do, so that inter-hemispheric context is available. We will then analyse where attention mass goes for vertices in the parietal and lingual regions.
Baselines
The naive reference point is V2C-Flow without MED, trained with five seeds. It establishes the raw inter-seed vertex root-mean-square deviation (RMSD) and the share of self-intersecting faces.
The strongest existing method on reproducibility is V2C-Flow with MED. Its paper reports surface distance to FreeSurfer of approximately 0.2 to 0.3 mm average symmetric surface distance (ASSD) on ADNI and a 25% to 50% reproducibility improvement in RMSD across seeds. No absolute regional RMSD value was found on Lacuna for this method, so we will reproduce it.
The strongest method on accuracy and regularity is CorticalFlow. It reports a 17.38% Chamfer reduction over DeepCSR and at least 32.58% fewer self-intersecting faces than Voxel2Mesh or Neural Mesh Flow.
All baselines will use the ADNI split of Madlindl et al. (1,154 train, 169 validation, 323 test scans). They will share the same template resolution, the same FreeSurfer v7.2 targets, matched parameter count within 10%, and an equal hyperparameter search budget of twenty trials each, all with MED enabled.
Two ablations isolate the key ingredient. The first keeps the identical attention module but assigns vertices to rings by random index rather than by template coordinates. This tests whether template anchoring matters or whether any sparse global mixing suffices. The second applies $g_i$ as direct vertex offsets instead of through the smooth lift. This tests whether the lift is what preserves regularity.
Beating the baseline means three things together:
- A statistically significant reduction in regional inter-seed RMSD in the parietal and lingual regions, beyond MED alone.
- ASSD no worse than MED's.
- A self-intersection rate no worse than CorticalFlow's.
Why not just…?
Why not just use a deeper or dilated UNet? Larger convolutional receptive fields are the cheapest fix. ViT-V-Net's comparison against VoxelMorph-2 on brain registration (Dice 0.726 vs 0.711) suggests convolutions alone leave some long-range alignment on the table. That evidence comes from a different task. We will therefore include a parameter-matched UNet with one more level and dilated bottleneck convolutions. If it closes the regional variance gap, our structured routing is unnecessary and we will say so.
Why not just broadcast a single pooled global feature to every vertex? This is the real threat. A squeeze-and-excite style global code is nearly free, and it may carry all the "which brain is this" context the velocity field needs. Pairwise routing matters only if specific distant regions inform specific local decisions, for example homologous regions across hemispheres. The pooled-code variant is a mandatory baseline. The proposal's central claim stands or falls on beating it in the high-variance regions, not on average ASSD.
Why not just ensemble seeds or regularize harder? Averaging several runs reduces variance by construction, and stronger MED weights shorten trajectories further. Neither changes what the network can see, and averaging vertex positions from independent flows forfeits the per-model diffeomorphism. We will report a five-seed ensemble as an upper reference for variance and an MED weight sweep. This lets readers see whether context buys something that regularization cannot.
Why not put global attention in the voxel encoder, ViT-V-Net style? That is the natural transfer from registration. It pays quadratic cost over 3D patches and loses boundary detail to patching. We run it at the largest patch resolution that fits the shared GPU budget and report memory alongside accuracy.
Proposed Experiments
The primary benchmark is ADNI with the split above. It is evaluated with ASSD and Hausdorff distance to FreeSurfer, percentage of self-intersecting faces, and per-region inter-seed RMSD over five seeds. Generalization is checked on OASIS-1, where Vox2Cortex reports ASSD of roughly 0.3 mm. Test-retest consistency uses the 40-scan, three-subject set of Madlindl et al.
The decisive check addresses the ambiguity explanation. Our method may lower inter-seed variance in parietal and lingual regions while test-retest variance there stays high. FreeSurfer's own test-retest disagreement in those regions may also already match the residual variance. In that case the remaining spread is label or anatomical ambiguity, and the receptive-field hypothesis explains only the part we removed. We report that split explicitly.
A second check addresses whether the routing is used at long range. We use attention maps and a counterfactual that masks the contralateral hemisphere during inference. The test is whether left-hemisphere vertices actually rely on right-hemisphere information.
Finally, we measure peak memory and runtime at 30k and 435k vertices against CorticalFlow's reported 1.1 seconds for four surfaces. Template-anchored routing is only worthwhile if it scales to the resolutions where local models currently have no alternative.
References
- Template-Based Cortical Surface Reconstruction with Minimal Energy Deformation
- T-Pixel2Mesh: Combining Global and Local Transformer for 3D Mesh Generation from a Single Image
- CorticalFlow: A Diffeomorphic Mesh Transformer Network for Cortical Surface Reconstruction
- Vox2Cortex: Fast Explicit Reconstruction of Cortical Surfaces from 3D MRI Scans with Geometric Deep Neural Networks
- CCNet: Criss-Cross Attention for Semantic Segmentation
- EccoMamba: Enhanced Cross-hierarchical Continuity Orthogonal Mamba for Medical Image Segmentation
- Compact Generalized Non-local Network
- ViT-V-Net: Vision Transformer for Unsupervised Volumetric Medical Image Registration