Source-linked AI summary
FiberGeoText: A Vision-Language Model for Population- Level Organization of Superficial White Matter
Yuqian Chen, R. Jarrett Rushmore, Guikun Chen, Fan Zhang, Edward Yeterian, Nikos Makris, Yogesh Rathi, Lauren J. O'Donnell
TL;DR
Short-range SWM connections are difficult to organize across individuals because trajectories and cortical folding vary, and spatial proximity alone may not define correspondence. FiberGeoText combines trajectory, cortical anatomy from multiple parcellations, and shape in a multimodal representation. It organizes clusters coherently across individuals and recovers most learned clusters in unseen subjects.
Problem
Short SWM fibers remain under-characterized, and their short trajectories and variable cortical folding make correspondence across individuals challenging.
Source: §1 · pp. 2–3Method
FiberGeoText jointly represents streamline geometry, text-encoded cortical anatomy from multiple parcellations, and explicit shape features.
Source: §1 · p. 4, §2.2.1 · p. 7, §2.2.2 · p. 8Results
FGT organizes short SWM streamlines into clusters coherent in cortical connectivity and morphology across individuals, recovering an average of 96.7% of the 5,000 learned clusters in testing subjects without reclustering.
Source: §4 · pp. 16–17Takeaways & Limitations
The findings indicate that SWM correspondence is not adequately captured by spatial proximity alone.
Source: §4 · p. 17There is no direct ground truth for short-range SWM correspondence, and tractography or cortical parcellation errors may influence the learned organization.
Source: §4 · pp. 17–18
Abstract
from arXiv · showhide
The superficial white matter (SWM), a critical brain region for cognition across the lifespan and brain disease, contains abundant short-range association fibers whose organization remains incompletely characterized, in part because the short trajectories and highly variable cortical folding make correspondence across individuals challenging. Anatomically corresponding connections may vary in spatial location across individuals and therefore may not be adequately defined by geometric proximity alone. We introduce FiberGeoText (FGT), a vision-language model (VLM) for organizing short-range superficial white matter (SWM) streamlines reconstructed from ultra-high-resolution diffusion MRI into population-level clusters. FGT jointly represents three complementary properties of each streamline: its three-dimensional trajectory, its cortical anatomical context, and its shape. Cortical endpoint information from multiple parcellation schemes is expressed as text and encoded using a pretrained large language model (LLM), enabling heterogeneous anatomical descriptions to contribute to a common continuous representation. We evaluated FGT on acquired submillimeter 0.76 mm diffusion MRI data. Compared with state-of-the-art (SOTA) methods, FGT produced substantially greater cortical parcel coherence, within-cluster shape consistency, cluster-size consistency, and cross-subject correspondence. The trained model also generalizes well to unseen subjects with an average of 96.7% of the 5,000 learned clusters recovered, and high consistency of cluster structure between training and testing data. Together, these findings demonstrate that integrating geometric, anatomical, and shape information by learning multimodal deep embeddings with a VLM model enables robust learning of population-consistent SWM organization despite interindividual anatomical variability.
1 Introduction
Short superficial white matter fibers remain under-characterized, and their variability makes population-level correspondence challenging. FiberGeoText addresses this by integrating geometry, shape, and language-encoded cortical anatomy from multiple parcellations.
1 Introduction
- Short trajectories and variable cortical folding make it difficult to identify corresponding SWM connections across individuals.The SWM is important for cognition, lifespan development, and vulnerability to aging and disease.Source: §1 · p. 2
- Short-range SWM bundles remain under-characterized despite their abundance, and their geometric diversity complicates population-level characterization.Existing clustering studies prioritized longer bundles, and none had incorporated ultra-high-resolution submillimeter dMRI.Source: §1 · pp. 2–3
- Most methods rely on a single coarse cortical atlas, while endpoint labels alone miss trajectory and shape differences among fibers.Complementary parcellations can provide distinct descriptions of connected regions, motivating joint consideration of anatomy, trajectory, and shape.Source: §1 · p. 3
- FiberGeoText maps endpoint descriptions from multiple parcellations into a language-derived representation integrated with geometry and shape features.The framework learns population-consistent representations of short-range SWM connections from ultra-high-resolution dMRI.Source: §1 · p. 4
2 Methods
FiberGeoText represents each SWM streamline using its trajectory, cortical context, and shape, then derives population-level clusters for assigning consistent connection identities to new subjects.
2 Methods
- Each streamline is characterized by its three-dimensional trajectory, cortical anatomical context, and shape.Cortical context is expressed as text derived from multiple parcellations.Source: §2 · p. 4
- A point-based neural network encodes trajectory and an LLM encodes anatomical text, with both combined with shape features in a joint embedding..Source: §2 · p. 4
- The learned embeddings are clustered to provide connection identities that can be assigned to streamlines from new subjects.Figure 1 is identified as an overview of the framework.Source: §2 · p. 4, Figure 1 · p. 5
2.1 Imaging data and multimodal streamline information
The study used submillimeter diffusion MRI, reconstructed short association streamlines, characterized cortical endpoints across parcellations, and measured streamline shape using length and U-ratio.
2.1.1 High-resolution dMRI
- Ten healthy adults were scanned with diffusion MRI reconstructed to isotropic 0.76 mm³ resolution.The acquisition used a 3T scanner and also collected high-resolution structural T1-weighted images.Source: §2.1.1 · p. 6
2.1.2 SWM tractography
- Parallel transport tractography retained 6–50 mm streamlines with both endpoints in cortex and confined to one hemisphere.Eight subjects were used for training and two held out for testing; 150,000 streamlines per training subject and 1,000,000 per testing subject were sampled.Source: §2.1.2 · p. 6
2.1.3 Cortical anatomical information
- Cortical endpoint regions were labeled using three atlases with different parcellation schemes.The endpoint label intersected in each cortical atlas was recorded for each streamline endpoint.Source: §2.1.3 · pp. 6–7
2.1.4 Streamline shape information
- Streamline shape was characterized by length and U-ratio, the endpoint distance divided by streamline length.Lower U-ratio values indicate more curved, U-shaped streamlines, while values near 1 indicate straighter trajectories.Source: §2.1.4 · p. 7
2.2 FiberGeoText multimodal deep embedding framework
FiberGeoText learns a joint representation of streamline geometry, cortical anatomy, and shape, then uses population-level clusters to assign consistent connection identities across subjects.
2.2 FiberGeoText multimodal deep embedding framework
- FiberGeoText independently encodes streamline trajectory and cortical anatomy, combines both embeddings with shape information, and clusters the resulting joint representation.New-subject streamlines can be assigned to stored centroids without rerunning k-means.Source: §2.2 · p. 7
2.2.1 Streamline geometry encoding
- A DGCNN encodes each streamline’s point-cloud trajectory, with left-right coordinates made absolute for bilateral clustering.Self-supervised training predicts pairwise geometric distances to bring similar trajectories closer in the embedding space.Source: §2.2.1 · p. 7
2.2.2 Cortical anatomy encoding using an LLM
- Endpoint labels from three parcellations are converted to text and encoded with a pretrained LLM to create an anatomy embedding.Contrastive training treats streamlines sharing an unordered region pair in at least one parcellation as positive pairs and balances positive and negative samples.Source: §2.2.2 · pp. 7–8
2.2.3 Multimodal integration of geometry, anatomy, and shape
- Geometry and anatomy embeddings are concatenated with streamline length and U-ratio after normalization to [0, 1].The resulting representation reflects trajectory similarity, cortical anatomical similarity, and streamline shape.Source: §2.2.3 · p. 8
2.2.4 Population-level organization and inference
- K-means clustering with k=5,000 creates population-level cluster centroids used to assign unseen-subject streamlines to existing identities.Inference uses the closest centroid rather than rerunning k-means for each new subject.Source: §2.2.4 · p. 8
3 Experiments and Results
The experiments assess FiberGeoText across anatomical coherence, cluster size and shape, compactness, cross-subject consistency, and generalization, using eight training and two held-out subjects. FGT improves several consistency measures over DFC and FFClust, with a geometric-compactness tradeoff, and recovers most learned clusters in unseen subjects.
3.1 Evaluation Metrics
- The evaluation measures CPC, cluster-size and shape variability, training-subject representation, unseen-subject cluster recovery, geometric compactness, and embedding-to-centroid mismatch.Higher CPC, PoS, and PoC indicate greater coherence or representation; lower CoVs and RNWCSS indicate more uniform clusters or compactness, while distortion near 1 indicates similar training and testing distances.Source: §3.1 · pp. 9–10
3.2 Experimental Results
- The study used 1,200,000 streamlines from eight subjects for training and clustering, then evaluated generalization on two held-out subjects with 1,000,000 streamlines each.Experiments on cluster number, method comparison, and ablations used training data; the held-out subjects were used for generalization evaluation.Source: §3.2 · p. 10
3.2.1 Effect of the Number of Clusters
- As k increased from 500 to 10,000, CPC rose and shape consistency and compactness improved, while CoV of NoS rose and PoS fell from 1.0 to 0.857; k = 5,000 was selected.Source: §3.2.1 · p. 10, Table 1 · p. 11
3.2.2 Comparison with SOTA Fiber Clustering Methods
- FGT had higher CPC across DK, HOA, and MMP, lower CoVs for NoS, U-ratio, and length, and higher PoS than DFC and FFClust; DFC had the lowest RNWCSS.Example clusters also showed greater within-cluster shape similarity and endpoint convergence for FGT.Source: §3.2.2 · p. 11, Table 2 · p. 12
3.2.3 Ablation study
- FGT achieved the highest Mean Normalized Score; adding shape to Fiber+Text lowered U-ratio and length CoVs and raised PoS, while moderately reducing CPC.Text-only, Fiber+Text, and FGT all had higher CPC than Fiber-only across the three parcellations.Source: §3.2.3 · pp. 12, 14, Table 3 · p. 14
3.2.4 Generalization to Unseen Subjects
- PoC was 94.88% for sub-01 and 98.36% for sub-02; distortion ratios near 1 indicated similar embedding-to-centroid distances in testing and training.The qualitative examples also show corresponding clusters recovered in both unseen testing subjects.Source: §3.2.4 · pp. 14–15, Table 4 · p. 15, Figure 3 · p. 16
4 Discussion
FGT organizes short-range SWM streamlines by jointly representing geometry, cortical anatomy, and shape. Its clusters transfer to unseen subjects and balance anatomical and morphological correspondence against purely geometric compactness, though validation and data scope remain limited.
4 Discussion
- FGT permits greater intersubject spatial variation to align cortical anatomy and morphology, rather than treating geometric compactness as sufficient correspondence.DFC achieved lower RNWCSS, while FGT achieved substantially higher cortical parcel coherence and shape consistency.Source: §4 · p. 17
- Cortical parcel coherence was approximately 96%, 88%, and 44% for Desikan–Killiany, Harvard–Oxford, and Glasser-MMP, respectively.The finer MMP subdivisions provide a more stringent endpoint-agreement measure; lower pooled coherence may partly reflect intersubject fine-scale anatomical variability.Source: §4 · p. 17
- Text-embedded endpoint descriptions let parcellations contribute to a common representation without requiring shared nomenclature or region boundaries.The approach also supports interpretation across complementary labeling schemes and granularity levels.Source: §4 · p. 17
- Ablations showed that anatomy improved endpoint coherence, geometry improved spatial compactness, and length plus U-ratio improved within-cluster morphological consistency.No single information source captured all aspects of SWM correspondence; explicit shape is particularly useful for short trajectories spanning relatively straight to strongly curved forms.Source: §4 · p. 17
- Approximately 96.7% of the 5,000 learned clusters were recovered on average in unseen subjects without reclustering, with distortion ratios close to training data.Inference encodes a new streamline and assigns it to an existing cluster centroid.Source: §4 · p. 17
- Short-range SWM correspondence lacks direct ground truth, and evaluation remains constrained by tractography and parcellation uncertainties and the need for larger, more diverse high-resolution datasets.The current framework concatenates geometry, anatomy, and shape before clustering; future work could investigate learned fusion or weighting and additional validation.Source: §4 · pp. 17–18
5 Conclusion
The conclusion presents FGT as a way to organize short-range SWM connections consistently across individuals by integrating trajectory, cortical anatomy, and shape. Its population-level organization accommodates interindividual spatial variability and is supported by recovery of most learned clusters in unseen subjects.
5 Conclusion
- FGT integrates trajectory, cortical anatomy, and shape to organize short-range SWM connections consistently across individuals while accommodating interindividual spatial variability.Recovery of most learned clusters in unseen subjects supports a multimodal definition of SWM correspondence rather than one based on spatial proximity alone.Source: §5 · p. 18