Diffusion MRI tractography reconstructs the brain's white-matter wiring by following, step by step, the main direction in which water diffuses inside each small cube of tissue (a voxel). It fails most often where several bundles squeeze through the same voxels and cross, a region called a bottleneck. There, a tracker reaching the far side of the crossing has to decide which outgoing path continues which incoming one, and the local signal does not contain the answer. This note treats that decision as a vertex-correspondence problem. Fiber populations entering a bottleneck form one point set and those leaving it form another. The task is to find the matching between the two sets that conserves fiber mass, keeps directions smooth, and, most importantly, keeps neighboring axons neighbors on the way through. That last property can be written as a Gromov-Wasserstein optimal transport problem, a method that matches two point sets by comparing the distances within each set rather than absolute positions.
Why this matters: the largest controlled evaluation of tractography, 96 pipelines from 20 groups on a digital phantom, found four invalid bundles for every valid one. Half of those invalid bundles appeared systematically across groups, and supplying noise-free data did not remove them (The challenge of mapping the human connectome based on diffusion tractography). That is direct evidence that local orientation estimates cannot settle global routing. The proposal supplies a global constraint applied only where it is needed. The constraint comes from the graph-matching literature, but it is applied inside a single brain rather than across graphs.
Core Claim
Many systematic false bundles come from streamlines that switch bundles inside a crossing. Such switches break two things that real axons respect. They break mass balance: the fiber density entering a bottleneck along one bundle should leave along that bundle. They also break topographic order: axons that sit next to each other on entry stay roughly next to each other on exit.
Concretely, for each detected bottleneck region we build two vertex sets. The entry set contains the fixels on the region's boundary that point inward. A fixel is one fiber population within a voxel, carrying a position, a direction, and a density from constrained spherical deconvolution (a method that sharpens the diffusion signal into distinct fiber peaks). The exit set contains the boundary fixels that point outward. We then solve for a soft correspondence $T$, where $T_{ij}$ is the fraction of fiber mass routed from entry fixel $i$ to exit fixel $j$:
$$\min_{T \ge 0} \; \sum_{i,j} C_{ij} T_{ij} \;+\; \alpha \sum_{i,j,k,l} \big(D^{\text{in}}_{ik} - D^{\text{out}}_{jl}\big)^2 T_{ij} T_{kl} \;+\; \lambda\, \mathrm{KL}(T\mathbf{1}\,\|\,a) + \lambda\, \mathrm{KL}(T^\top\mathbf{1}\,\|\,b)$$
The symbols are as follows. $C_{ij}$ is a continuity cost: the curvature of the smoothest path joining entry fixel $i$ to exit fixel $j$, plus angular mismatch. $D^{\text{in}}$ and $D^{\text{out}}$ are within-set geodesic distances along the boundary surface. $a$ and $b$ are fixel fiber densities. $\alpha$ weights the topography term. $\lambda$ sets how strictly mass must balance.
The two KL terms make the problem unbalanced, so fanning (one entry, many exits) and partial terminations are allowed rather than forced into a one-to-one matching. The quadratic term only compares distances. A bundle that rotates or flips as a whole inside the crossing therefore pays nothing, because rotations and flips preserve distances. Only streamlines that scramble or jump between bundles are penalized. Outside bottlenecks, ordinary local tracking runs unchanged. Inside them, streamlines are routed by sampling from the rows of $T$.
The claim to be tested is this: on the ISMRM 2015 phantom, adding the topography term reduces invalid bundles relative to the same pipeline using only continuity and mass terms, without losing valid bundles.
Relation to Prior Work
Three groups of papers play different roles here.
The first group motivates the problem. The phantom study above establishes the bottleneck failure and the 36% average volume recovery (The challenge of mapping the human connectome based on diffusion tractography). Signal-level advances do not close the gap. Multi-tissue deconvolution finds multiple fibers in about 61% of voxels versus 23% for older methods (Ground-Truth Effects in Learning-Based Fiber Orientation Distribution Estimation in Neonatal Brains). Spatially regularized super-resolved deconvolution sharpens peaks (Spatially regularized super-resolved constrained spherical deconvolution (SR 2 -CSD) of diffusion MRI data). Equivariant and spherical networks speed up or improve fixel estimation (Equivariant spatio-hemispherical networks for diffusion MRI deconvolution, DeepFixel: Crossing white matter fiber identification through spherical convolutional neural networks). All of these improve the vertices of our matching problem. None of them decides the matching itself. We reuse their outputs as inputs.
The second group supplies the method. Gromov-Wasserstein transport aligns sets by comparing internal geometry (COPT: Coordinated Optimal Transport on Graphs, Fused Gromov-Wasserstein Graph Mixup for Graph-level Classifications). The latter's fused feature-plus-structure cost is the template for our objective. Functional-map alignment enforces near-isometric correspondences as a smoothing prior (Graph Alignment via Dual-Pass Spectral Encoding and Latent Space Communication), which is the same intuition as our topography term. Those papers align two whole graphs. We align the two boundaries of one tissue region, where the "isometry" is a physical claim about axon packing rather than a modeling convenience.
The third group is the closest connecting work, and it operates downstream. WL-align matches atlas-parcel connectomes across subjects with Weisfeiler-Lehman signatures and Hungarian matching (Network alignment and similarity reveal atlas-based topological differences in structural connectomes). Optimal-transport and block-model methods compare connectomes across modalities or subjects (Evolvable Graph Diffusion Optimal Transport with Pattern-Specific Alignment for Brain Connectome Modeling, From Your Block to Our Block: How to Find Shared Structure Between Stochastic Block Models over Multiple Graphs). All of these take the tractography-derived adjacency matrix as given. We move correspondence upstream, into the construction of the streamlines that produce that matrix.
The SSBM paper also warns that strict one-to-one block matching misses splitting and branching. That warning is why our formulation is unbalanced rather than bijective.
A possible tension comes from anatomy-to-tract mapping, which generates bundles from T1-weighted anatomy alone (Anatomy-to-tract mapping infers white matter pathways without diffusion streamline propagation). That result suggests global priors could replace local signal entirely. We take a middle position: diffusion data define the vertices, and a geometric prior resolves only the routing that the data cannot.
Research Program
The work has four parts.
Bottleneck detection. Voxels with two or more fixels separated by less than a chosen angle, or where a probabilistic tracker's next-step distribution is multimodal, are grown into connected regions. On the phantom, region definitions are validated against ground-truth crossing locations.
Boundary graphs. Inward and outward fixels on each region boundary become vertices. Densities come from deconvolution lobe integrals. Distances are geodesics on the boundary surface.
The solver. We use entropic unbalanced fused Gromov-Wasserstein with several random initializations, because the quadratic term is nonconvex. Regions hold tens to hundreds of boundary fixels, so per-region cost is small. The total cost scales with the number of bottlenecks, not with brain size. This is the main computational contrast with whole-brain quadratic assignment, which Graph Alignment Kernels using Weisfeiler and Leman Hierarchies notes is NP-hard.
Routing. Streamlines from any base tracker that reach an entry fixel continue through the region along a smooth curve to an exit fixel sampled from the corresponding row of $T$. Mass marginals also tell us where to reseed, which directly targets volume underestimation.
Controlled phantoms. With the Fiberfox simulator (The challenge of mapping the human connectome based on diffusion tractography) we build crossings of three kinds:
- bundles that keep their internal order, where the prior is correct;
- bundles that rotate as a whole, where the prior is distance-preserving and should still hold;
- bundles whose fibers genuinely interleave, where the prior is wrong.
The third kind maps where the method should break.
Baselines
The naive reference point is the base tracker with no bottleneck handling: standard deterministic and probabilistic tracking on deconvolution peaks, run through the Tractometer pipeline on the ISMRM 2015 phantom. Its expected behavior is the field-wide pattern of about 90% of bundles found, roughly four invalid bundles per valid one, and about 36% volume recovery (The challenge of mapping the human connectome based on diffusion tractography).
The strongest existing method in the file is TractoTransformer. It combines a 3D convolutional encoder of the local neighborhood with a decoder-only transformer that conditions each step on the whole streamline history. It reports 84% valid connections and a 75% F1 score on the ISMRM 2015 phantom (TractoTransformer: Diffusion MRI Streamline Tractography using CNN and Transformer Networks). It is the strongest competitor because its history context could, in principle, learn which exit continues which entry, which is exactly our question. We rerun it from released or reimplemented weights on the same phantom and on the same TractoInferno splits. We then apply our routing module on top of its streamlines, so the comparison isolates the bottleneck step. Beating it means improving its invalid-bundle count and overreach at equal or better valid connections and F1. For TractoInferno, TractoTransformer reports the highest overlap and Dice among compared methods, but no reported result was found on Lacuna for its exact values, so we must regenerate them.
The key ablation is the same pipeline with $\alpha = 0$. This keeps continuity and mass balance but removes topography, and amounts to an unbalanced linear optimal-transport router. A second ablation also removes the mass terms, which reduces to greedy smoothest-continuation (Hungarian) matching. All variants share fixel estimates, seeds, step size, and a matched hyperparameter search budget on a held-out phantom variant.
Why not just…?
Why not just let a sequence model learn the routing, as TractoTransformer does? This is the real threat. A transformer with full path history sees the entry configuration and could implicitly learn to keep bundles intact. Two facts make us doubt it suffices. First, it decodes with argmax over 724 directions, so it commits to one branch without representing the competing ones. Second, it is trained to imitate reference tractograms, which carry the systematic errors the phantom study documents. The decisive experiment is stacking our router on its output. If invalid bundles do not fall, the learned history already captures topography and our module adds nothing.
Why not just use the continuity and mass terms, an ordinary linear transport or Hungarian match? This is cheaper and convex. If the $\alpha = 0$ ablation matches the full model on invalid bundles, the topography hypothesis is wrong and the simpler router should be preferred. We report this comparison first.
Why not just sharpen the local fiber estimates with better deconvolution? The phantom study's noise-free condition is the cleanest answer: perfect local data still produced systematic invalid bundles (The challenge of mapping the human connectome based on diffusion tractography). Better fixels improve our inputs but leave the matching undetermined. We still rerun with the strongest available deconvolution, to check that our gains do not shrink to zero as fixels improve.
Why not just constrain tracking with anatomical priors from T1-weighted images? Anatomy-to-tract mapping shows macro-anatomy predicts bundle shape (Anatomy-to-tract mapping infers white matter pathways without diffusion streamline propagation). However, cortical folding says little about which of two bundles inside a deep crossing continues where. On TractoInferno we compare bundle Dice and overreach against that approach, with the expectation that the two are complementary rather than substitutes.
Proposed Experiments
Primary test: ISMRM 2015 phantom. We report valid and invalid bundles, valid connections, overlap, overreach, bundle volume recovery, and F1 for every baseline and ablation, using five seeds each. The decisive outcome is the full model against $\alpha = 0$ on invalid bundles. A reduction in invalid bundles with unchanged valid bundles supports the claim. No reduction falsifies it.
Fiberfox crossing phantoms. These test the mechanism. We predict gains on order-preserving and rigidly rotating crossings, and losses on interleaved crossings. Seeing that loss confirms the method does what we say rather than acting as a generic regularizer.
TractoInferno across sites. We report Dice, overlap, and overreach on the 30 curated bundles against regenerated baselines. We also check that gains hold across scanner sites.
Downstream connectome check on the HCP 100-subject cohort. We rebuild connectomes with and without bottleneck routing and measure cross-subject WL-align matching accuracy (Network alignment and similarity reveal atlas-based topological differences in structural connectomes). The reported 100% figure is for self-alignment to a permuted copy, not across subjects, so the cross-subject baseline must be computed by us. If removing systematic false edges makes subjects' graphs more alignable, the vertex-correspondence community has a reason to care about how tractography treats crossings. If it does not, that is also informative about how much crossing errors actually distort connectome comparisons.
References
- The challenge of mapping the human connectome based on diffusion tractography
- Ground-Truth Effects in Learning-Based Fiber Orientation Distribution Estimation in Neonatal Brains
- Spatially regularized super-resolved constrained spherical deconvolution (SR 2 -CSD) of diffusion MRI data
- Equivariant spatio-hemispherical networks for diffusion MRI deconvolution
- DeepFixel: Crossing white matter fiber identification through spherical convolutional neural networks
- COPT: Coordinated Optimal Transport on Graphs
- Fused Gromov-Wasserstein Graph Mixup for Graph-level Classifications
- Graph Alignment via Dual-Pass Spectral Encoding and Latent Space Communication
- Network alignment and similarity reveal atlas-based topological differences in structural connectomes
- Evolvable Graph Diffusion Optimal Transport with Pattern-Specific Alignment for Brain Connectome Modeling
- From Your Block to Our Block: How to Find Shared Structure Between Stochastic Block Models over Multiple Graphs
- Anatomy-to-tract mapping infers white matter pathways without diffusion streamline propagation
- Graph Alignment Kernels using Weisfeiler and Leman Hierarchies
- TractoTransformer: Diffusion MRI Streamline Tractography using CNN and Transformer Networks