How can vision-language models organize superficial white matter streamlines?

Why spatial proximity alone cannot define SWM correspondence

Superficial white matter sits between the deep white matter and the cortex and carries dense short-range association fibers connecting within and between gyri. The paper notes that the most prominent U-shaped fibers account for roughly twice the volume of the long-range bundles that have traditionally dominated diffusion MRI tractography research. Yet most SWM clustering studies have prioritized longer bundles, generally 30 to 85 mm, leaving shorter connections under-characterized.

The obstacle is not only imaging resolution. Short trajectories sit close to highly variable cortical folding, so anatomically corresponding connections may occupy different spatial locations across individuals. A clustering method that groups streamlines mainly by geometric proximity can therefore merge or split connections that are anatomically the same. The paper states this directly: SWM correspondence should not be defined by spatial proximity alone.

FiberGeoText is built around that claim. It organizes streamlines of 6–50 mm reconstructed from submillimeter diffusion MRI into population-level clusters, and it treats cortical anatomy and shape as information that must enter the representation before clustering, not as labels attached afterward.

Three inputs are fused into one streamline embedding

For each streamline, the framework encodes three complementary properties. A point-based neural network encodes the three-dimensional trajectory. A pretrained large language model encodes cortical anatomical context, which is written as text describing the region labels intersected by each endpoint. Two explicit shape measures are added: length in millimeters, and the U-ratio, defined as the Euclidean distance between endpoints divided by streamline length, where lower values indicate more curved U-shaped fibers and values near 1 indicate straighter trajectories.

The text route is what lets heterogeneous atlases coexist. The study used three cortical parcellations with different granularity: Desikan–Killiany with 34 regions per hemisphere, Harvard–Oxford 2.0 with 73, and the HCP multi-modal parcellation with 180. Because each endpoint description is expressed as text and embedded by the language model, the schemes do not need a shared nomenclature or common region boundaries to contribute to one continuous representation.

Before concatenation, the geometry embedding, anatomy embedding and shape measures were normalized to the range [0, 1] to control their relative contribution to distance. The joint representation is therefore shaped simultaneously by where a streamline travels, which cortical regions it connects, and its local morphology.

Clusters become fixed identities that transfer to new subjects

The joint embeddings of training streamlines were partitioned with k-means, and the study selected k = 5,000 clusters for its final model. The resulting centroids are stored as discrete connection identities. During inference, a new streamline is encoded with the pretrained encoders, combined with its length and U-ratio, and assigned to the nearest centroid. No re-clustering is needed for a new subject, which is what preserves cluster identities across individuals.

The cluster count was itself a tradeoff rather than a default. As k rose from 500 to 10,000, cortical parcel coherence improved for all three atlases, and the coefficients of variation for U-ratio and length fell, indicating tighter within-cluster shape consistency. Geometric compactness also improved. But the coefficient of variation of cluster size rose from 0.354 to 0.526, and the proportion of subjects represented fell from 1.0 to 0.857. Finer clusters were more internally coherent but less consistently present across training subjects, and k = 5,000 was chosen on that balance.

FiberGeoText trades geometric compactness for anatomical coherence

Against DFC, the authors' earlier deep fiber clustering framework, and FFClust, an unsupervised method recently applied to build an SWM atlas, FiberGeoText achieved the highest cortical parcel coherence across all three parcellations. At k = 5,000, coherence was approximately 96% for Desikan–Killiany, 88% for Harvard–Oxford and 44% for the Glasser-MMP scheme. It also showed lower variability in cluster size and lower coefficients of variation for U-ratio and length, meaning greater within-cluster shape consistency, and the highest proportion of subjects represented.

The comparison is not a clean sweep, and the paper is explicit about why. DFC achieved the lowest value on the geometric compactness measure, while FiberGeoText was higher. The authors interpret this as an expected tradeoff of the multimodal formulation: FiberGeoText permits more spatial geometric variability inside a cluster so that anatomically corresponding streamlines can be grouped even when their trajectories do not align exactly across individuals.

The MMP figure needs the same care. The paper attributes the lower 44% at least in part to that atlas's substantially finer cortical subdivision, which makes endpoint agreement a more stringent test. It also reports that MMP-based coherence was higher within individual unseen subjects than in pooled training data, suggesting some of the population-level shortfall reflects intersubject variability in fine-scale cortical anatomy rather than poor within-subject organization.

Ablations show no single input carries the organization

The ablation experiments separate the contributions. Cortical anatomy strongly increased endpoint coherence, but used alone it produced clusters with poorer geometric and shape consistency. Streamline geometry provided substantially greater spatial compactness. Adding length and U-ratio markedly improved within-cluster morphological consistency.

The paper's conclusion from this is that no individual information source was sufficient to capture all aspects of SWM correspondence. That matters for how readers should interpret the headline comparison: the gain over DFC is not simply more anatomy or more shape, but the combination of trajectory, cortical context and morphology in one space. The authors also suggest explicit shape information is particularly useful for short SWM trajectories, whose morphology spans a continuum from relatively straight intra-gyral fibers to more strongly curved U-shaped configurations.

Generalization holds on two held-out subjects, with caveats

The evaluation used 0.76 mm isotropic diffusion MRI from 10 healthy adults, mean age 36 ± 7.5 years, five male and five female. Eight subjects supplied 150,000 streamlines each for training, 1,200,000 in total, and two entirely held-out subjects supplied 1,000,000 streamlines each for testing. On those unseen subjects, an average of 96.7% of the 5,000 learned clusters were recovered without re-clustering, and cluster structure was highly consistent between training and testing data.

That result supports transfer of the learned organization, but the sample is small and the test set is two subjects. The paper itself lists further limits. There is no direct ground truth for short-range SWM correspondence; the evaluations quantify consistency of tractography geometry, shape, cortical anatomy and cross-subject occurrence, and the authors call for validation using anatomical dissection, tracer information where available, or functional and clinical associations. The framework also inherits uncertainty from both tractography reconstruction and cortical parcellation, and errors in either source may influence the learned organization.

The fusion step is another boundary. The current framework concatenates geometry, anatomy and shape before clustering; the authors point to learned multimodal fusion or weighting strategies as future work, along with additional anatomical information such as cytoarchitectonic or functionally relevant parcellations, and evaluation in larger and more diverse high-resolution datasets.

The useful threshold is coherence, not compactness

The paper's most decision-relevant result is a deliberate reversal of the usual criterion. If the goal is a population-level SWM organization in which clusters mean the same connection across people, then a method that wins on purely geometric compactness can still be the wrong choice. FiberGeoText's higher geometric variability within clusters is the mechanism by which it gains cortical parcel coherence and shape consistency, and the authors present that exchange as the point rather than a cost to be minimized.

The practical reading follows from the paper's own variables. When comparing SWM clustering approaches, the informative measures here are cortical parcel coherence across parcellations of differing granularity, within-cluster shape consistency, cluster-size consistency and cross-subject representation, read together with the cluster count that produced them. The k = 5,000 setting is where this study found its balance, with coherence still rising at finer granularity but cross-subject representation falling. For short-range fibers, the paper's evidence favors correspondence built from trajectory, cortical anatomy and shape jointly over correspondence built from location alone.

Sources used in this answer