Universal Tract: Whole-Brain Tractography

Updated 6 July 2026

Universal Tract is a whole-brain, tract-specific framework that uses learned tract orientation mapping to generate bundle-aware representations of white matter pathways.
It integrates tract segmentation, orientation mapping, and constrained probabilistic tracking using multi-plane strategies to produce anatomically valid tractograms.
The approach achieves robust performance with minimal annotations, significantly improving Dice scores and runtime compared to traditional methods.

Searching arXiv for the papers on arXiv and nearby work to ground the article. arxiv_search(query="Tract orientation mapping bundle-specific tractography TractSeg white matter tract segmentation one annotated subject", max_results=10) arxiv_search(query="(Wasserthal et al., 2018)", max_results=5) arxiv_search(query="(Wasserthal et al., 2019)", max_results=5) Universal Tract is not presented in the cited literature as a formally standardized method name. In diffusion MRI and bundle-specific tractography, it is best understood as a context-dependent designation for a broadly applicable, whole-brain, tract-specific framework that learns from fiber orientation distribution function peaks and produces bundle-aware representations for many white matter pathways across subjects and acquisition conditions (Wasserthal et al., 2018, Wasserthal et al., 2019, Xu et al., 2023). In this usage, the concept is anchored by tract orientation mapping (TOM), its extension to combined tract segmentation and orientation mapping, and a subject-level one-shot tract segmentation framework based on registration-based pseudo-subject synthesis and uncertainty-based refining. This suggests an operational notion of universality centered on tract coverage, automation, minimal annotation, and cross-subject robustness rather than a formal universality theorem.

1. Concept and scope

Within the white matter analysis literature represented here, the strongest “universal” reading is whole-brain rather than ontological. One line of work learns tract-specific orientation maps and tract-specific segmentations for 72 bundles and then uses them to generate anatomically constrained tractograms automatically (Wasserthal et al., 2019). Another line addresses the extreme supervision regime in which only one subject is annotated, yet the method still performs whole-brain tract segmentation with 72 tract labels on held-out Human Connectome Project subjects (Xu et al., 2023). The earlier TOM formulation focuses on 20 bundles in 105 Human Connectome Project subjects and emphasizes low angular errors, top Dice values, and runtimes that avoid whole brain tractography, atlas registration, or clustering (Wasserthal et al., 2018).

The cited one-shot segmentation paper is explicit that this should be understood as a subject-level one-shot white matter tract segmentation framework rather than a fully universal tract model in the strong theoretical sense (Xu et al., 2023). The same source states that the term “universal” should be used carefully and that a more accurate description is a subject-level one-shot whole-brain tract segmentation framework. A plausible implication is that “Universal Tract” denotes a design objective: a single pipeline that is broadly applicable within tract segmentation and bundle-specific tractography, rather than a claim of cross-dataset universality guarantees for all scanners, domains, or anatomical conventions.

2. Tract orientation mapping as the foundational representation

TOM introduces the central representational shift. Rather than performing tractography first and isolating bundles afterward, it learns a mapping from the original fiber orientation distribution function peaks to a list of tract orientation maps, also abbreviated TOM, where each map represents one known tract and each voxel contains no more than one orientation vector (Wasserthal et al., 2018). These tract orientation maps can act as a prior or even as direct input for tractography. The mapping is learned with an encoder-decoder fully-convolutional neural network architecture.

The 2019 extension makes this formulation more explicit in terms of input and output representation. The model uses the three principal FOD peaks per voxel, yielding 9 input channels in total, and regresses a 3D vector per voxel per tract (Wasserthal et al., 2019). TOM is therefore a learned voxelwise vector field that represents the voxel-wise principal orientation of one tract. Conventional FOD-based tracking uses tract-agnostic peaks that summarize local diffusion structure; TOM instead learns a tract-specific remapping from the original FOD peaks to a single orientation field for each tract. In technical terms, the orientation field becomes a bundle-specific anatomical prior.

The regression target is optimized with cosine similarity:

$\operatorname{loss}(\hat{y}, y) = -\frac{1}{N}\sum_{i=0}^{N} \frac{\left|\langle \hat{y}_i, y_i\rangle\right|}{\|\hat{y}_i\|_2 \,\|y_i\|_2}.$

The same 2019 work notes that an earlier variant also learned peak length jointly, but that the final method performs better when tract segmentation and peak-angle learning are separated (Wasserthal et al., 2019). This separation clarifies the role of TOM: it is the orientation backbone of bundle-specific tractography, not a substitute for tract extent or endpoint validity.

3. Integrated bundle-specific tractography

The combined framework augments TOM with two additional learned outputs per bundle: a tract mask and a start/end-region mask (Wasserthal et al., 2019). The tract segmentation model identifies the spatial extent of the bundle, while the start/end-region model identifies where valid streamlines are allowed to originate and terminate. This architecture turns TOM from a local directional prior into part of a constrained bundle-specific tracking system. After TOM-based tracking, streamlines are filtered so that any streamline leaving the tract mask is removed, and any streamline that does not start and end in the learned start/end masks is discarded.

A custom probabilistic tracking algorithm is introduced specifically for TOM. Since TOM provides one orientation per voxel rather than a full distribution, the method samples from a Gaussian distribution with fixed standard deviation centered on each peak. The direction sampling is described by

$\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$

with $\sigma = 0.15$ , where $\boldsymbol{\mu}$ is the TOM-predicted peak vector at that voxel. The step size is 0.7 voxels, final streamlines are interpolated with B-splines, seeds are placed randomly inside the tract mask until up to 2000 streamlines per tract are generated, and streamlines shorter than 50 mm are rejected. The paper states that $\sigma = 0.15$ was chosen conservatively as a lower bound on fiber orientation dispersion, and that a larger standard deviation such as 0.3 yields much spurious tracking.

The full pipeline begins with constrained spherical deconvolution followed by peak extraction. For tract segmentation and start/end segmentation, predictions are made from axial, coronal, and sagittal slices and then averaged; for TOM, the best results were obtained using only coronal slices. Tract segmentation probabilities are thresholded at 0.5, and TOM uses linear outputs with peaks shorter than 0.3 discarded. For 72 bundles, the TOM task is split into four models because predicting all 72 tracts at once did not converge. Ground-truth construction is also nontrivial: tract masks are generated from streamline sets, start/end regions are derived from streamline endpoints using DBSCAN, a random forest, binary closing, and dilation, and TOM targets are created by clustering streamline orientations with Mean Shift and taking the mean of the largest cluster (Wasserthal et al., 2019).

4. One-shot whole-brain tract segmentation

A complementary interpretation of Universal Tract emerges from the one-shot setting, where whole-brain tract segmentation is attempted with only one annotated subject (Xu et al., 2023). In this framework, each subject is represented as a 9-channel 3D volume of size $144\times144\times144\times9$ , corresponding to fODF peak directions in sagittal, axial, and coronal orientations. The task is whole-brain tract segmentation with 72 tract labels, and evaluation is performed on 21 held-out Human Connectome Project subjects.

The framework has three stages: registration-based peak augmentation (RPA), TractSeg-A training with uncertainty estimation, and uncertainty-based refining (URe). Let the only labeled subject be $\{x,l\}$ , with $x$ the labeled fODF-peak volume and $l$ the tract segmentation label, and let $y^{(i)}$ be an unlabeled subject. A deep registration model produces a voxel-wise displacement field,

$\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$ 0

with deformation

$\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$ 1

The labeled subject and label are warped to the target subject space as

$\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$ 2

The registration network is trained with a smoothness loss,

$\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$ 3

and a similarity loss,

$\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$ 4

combined as

$\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$ 5

One pseudo-subject is generated for each of the 83 unlabeled training subjects, yielding 83 pseudo-subjects.

TractSeg-A is trained only on the original labeled subject. As in TractSeg, the subject is decomposed into 2D slices in sagittal, axial, and coronal planes, and the three plane-wise predictions are averaged during inference. The segmentation loss is binary cross entropy:

$\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$ 6

Its purpose is not final prediction but voxel-level confidence estimation on pseudo-subjects. URe then converts frozen TractSeg-A outputs $\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$ 7 into an uncertainty map,

$\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$ 8

and uses it to reweight the pseudo-label loss:

$\mathbf{v} \sim \mathcal{N}(\boldsymbol{\mu}, \sigma^2 I),$ 9

The paper notes that the notation is somewhat inconsistent, but conceptually this is the segmentation loss on pseudo data weighted to suppress low-confidence voxels. At test time, only TractSeg-B is used, and the final output is a $\sigma = 0.15$ 0 tract map obtained by averaging predictions across the three planes (Xu et al., 2023).

5. Empirical performance, runtime, and generalization

The empirical program attached to the Universal Tract idea is defined by accuracy, completeness, robustness, and speed. The original TOM paper compares its method to four state-of-the-art bundle recognition methods on 20 different bundles in a total of 105 Human Connectome Project subjects and reports anatomically convincing results even for difficult tracts, low angular errors, unprecedented runtimes, and top Dice values (Wasserthal et al., 2018).

The 2019 combined framework broadens this substantially to 72 bundles across high quality, low quality, and phantom data, with comparison against seven benchmark methods: TractQuerier, RecoBundles, streamline atlas registration, custom atlas registration, FSL atlas registration, MRtrix atlas registration, multiple-mask registration, together with orientation-only baselines such as peak atlas and best original peak (Wasserthal et al., 2019). Quantitatively, it significantly outperforms all reference methods on segmentation Dice score across all three datasets. On HCP Quality it is about 14 Dice points better on average, on Clinical Quality about 18 points better, and on Phantom about 22 points better. For orientation, RecoBundles and TractQuerier can have slightly better angular error on HCP Quality because they find clean core streamlines in a high-quality whole-brain tractogram, but their error increases sharply on lower-quality data, whereas the proposed method’s angular error increases only by about 1 degree from HCP Quality to Clinical Quality. Runtime is reported as 137× faster than the reference methods on HCP Quality and 50× faster on Clinical Quality. The same paper further reports generalization on 17 additional datasets with different scanners, b-values, numbers of directions, resolutions, and pathologies, including Alzheimer’s disease with enlarged ventricles, multiple sclerosis lesions inside bundles, brain volume loss, schizophrenia, autism, pediatric and elderly populations, OASIS, IXI, COBRE, Rockland, and HCP 7T. Rigid registration to MNI space was only needed to match axis orientation conventions, not for the method itself.

The one-shot framework provides a distinct quantitative perspective on minimal-annotation universality (Xu et al., 2023).

Method	Mean Dice
U-Net	$\sigma = 0.15$ 1
TractSeg	$\sigma = 0.15$ 2
Ours (RPA)	$\sigma = 0.15$ 3
Ours (RPA+URe)	$\sigma = 0.15$ 4

The full RPA+URe method improves mean Dice by 3.56% over RPA alone, by 24.16% over U-Net, and by 29.82% over TractSeg. It achieves the best Dice on all 72 tracts, and the standard deviation is also the lowest, suggesting more stable performance. Visual results are described as showing more complete segmentations, including thin tracts, compared with the baselines.

The main misconception is to read “universal” as a formal claim of unrestricted generality. The one-shot segmentation paper explicitly rejects that interpretation and states that the method does not claim a universal segmentation architecture with explicit cross-dataset universality guarantees (Xu et al., 2023). The more precise characterization is a broadly applicable tract segmentation framework that demonstrates strong generalization across subjects under an extremely limited annotation regime. Likewise, the combined TOM framework is best read as a single learned framework that can infer tract-specific anatomy from diffusion MRI across many bundles, datasets, scanners, and disease states, while remaining firmly within the tractography setting (Wasserthal et al., 2019).

A second source of confusion is terminological rather than methodological. A separate geometry literature studies circular tractrices in $\sigma = 0.15$ 5 and develops what it describes as a unified tract geometry for circular directrices (Gorkavyy et al., 2022). There, the central objects are tractrices and circular pseudospheres, not neuroanatomical tracts. The paper states that it does not explicitly use the phrase “universal tract,” but it does provide a unified framework across the regimes $\sigma = 0.15$ 6, $\sigma = 0.15$ 7, and $\sigma = 0.15$ 8. This is an unrelated usage of “tract” and should not be conflated with bundle-specific tractography.

Taken together, the cited literature supports a restrained but technically substantive interpretation. Universal Tract denotes, in practice, a bundle-aware computational program in which tract-specific orientation, extent, and endpoint constraints are learned directly from diffusion-derived peaks, combined with efficient inference or tractography, and extended in some settings to subject-level one-shot supervision. The universality lies in broad tract coverage and robust applicability, not in a formal universal model valid without qualification.