GenTract: Generative Global Tractography
Alec Sargood, Lemuel Puglisi, Elinor Thompson, Mirco Musolesi, Daniel C. Alexander, [1.5ex] Hawkes Institute, Department of Computer Science, University College London, UK · 2025
Read GenTract: Generative Global Tractography by Alec Sargood, Lemuel Puglisi, Elinor Thompson, Mirco Musolesi, Daniel C. Alexander, [1.5ex] Hawkes Institute, Department of Computer Science, University College London, UK (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-13183.txt.
Introduction
Alec Sargood and four co-authors generate brain wiring maps directly, instead of tracing them step by step. Tractography infers white-matter pathways from diffusion MRI. Local methods follow fiber orientations one step at a time and accumulate errors, producing many false pathways, especially in noisy or low-resolution scans. Global methods optimize whole streamline collections against the data but cost heavily. GenTract reframes the problem as generation: learn a direct mapping from the diffusion image to complete, anatomically plausible streamlines, trying both diffusion-based and flow-matching generators.
Against state-of-the-art baselines, GenTract's precision was roughly twice the next-best methods on research-grade data, and its advantage grew to a factor of three and a half on low-resolution, noisy data. Clinical scans are rarely research-grade, so graceful degradation on imperfect data is the point.
Catalog record
| CREATOR | Alec Sargood, Lemuel Puglisi, Elinor Thompson, Mirco Musolesi, Daniel C. Alexander, [1.5ex] Hawkes Institute, Department of Computer Science, University College London, UK |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-13183
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
GenTract: Generative Global Tractography Alec Sargood, Lemuel Puglisi, Elinor Thompson, Mirco Musolesi, Daniel C. Alexander, [1.5ex] Hawkes Institute, Department of Computer Science, University College London, UK Abstract Tractography is the process of inferring the trajectories of white-matter pathways in the brain from diffusion magnetic resonance imaging (dMRI). Local tractography methods, which construct streamlines by following local fiber orientation estimates stepwise through an image, are prone to error accumulation and high false positive rates, particularly on noisy or low-resolution data. In contrast, global methods, which attempt to optimize a collection of streamlines to maximize compatibility with underlying fiber orientation estimates, are computationally expensive. To address these challenges, we introduce GenTract, the first generative model for global tractography. We frame tractography as a generative task, learning a direct mapping from dMRI to complete, anatomically plausible streamlines. We compare both diffusion-based and flow matching paradigms and evaluate GenTract’s performance against state-of-the-art baselines. Notably, GenTract achieves precision 2.1 × \times higher than the next-best method, TractOracle. This advantage becomes even more pronounced in challenging low-resolution and noisy settings, where it outperforms the closest competitor by an order of magnitude. By producing tractograms with high precision on research-grade data while also maintaining reliability on imperfect, lower-resolution data, GenTract represents a promising solution for global tractography. Figure 1 : Overview of our proposed GenTract methodology. We provide a learned embedding of, \mathbf{z} , representing whole dMRI information. Starting from Gaussian noise ( \epsilon ), all coordinates of individual streamlines are generated in parallel. Generated streamlines are collated to form the output tractogram. 1 Introduction The intricate network of white matter pathways forming the human connectome provides the structural architecture for all neural communication ( 9 ) . The ability to map these connections in vivo with diffusion-weighted Magnetic Resonance Imaging (dMRI) has offered insights into brain development ( 6 ) , aging ( 20 ) , and the pathophysiology of neurological disorders ( 7 ) , while also serving as a tool for neurosurgical planning ( 15 ) . However, current methods to reconstruct these pathways, known as tractography, face fundamental challenges that have limited their accuracy and reliability ( 22 ; 21 ) . Tractography algorithms use information from a dMRI scan, which comprises a set of Diffusion Weighted Images (DWIs), each sensitive to water diffusion in a distinct direction. Because water diffuses more readily along brain fibers than across them, the full set of DWIs can be used to estimate the distribution of fiber orientations within each voxel. The dominant tractography paradigm, local tracking 5 ; 35 ; 18 ; 33 ; 24 , uses these voxel-wise orientations to trace the trajectory of the white matter fibers step-by-step in the form of ‘streamlines’: lines in 3D space that represent the best estimates of the underlying fiber trajectories. However, global connectivity mapping using the local search strategy makes these methods prone to error accumulation, resulting in high false-positive connections ( 21 ; 30 ) . This problem is especially pronounced in clinical dMRI scans, which are often limited to acquisition times of a few minutes leading to lower spatial resolution and signal-to-noise ratio compared to the longer, high-resolution protocols used in research. In addition, these methods depend on heterogeneous preprocessing pipelines, especially for constructing seeding masks that define streamline initiation points. This process, which may involve anything from manual segmentation to thresholding of tissue masks, introduces substantial operator- and method-dependent variability, undermining reproducibility ( 16 ; 33 ) . Global tractography 4 seeks to address these limitations by inferring the most probable streamline configurations across the entire brain simultaneously through a single optimization problem. Compared with local methods, global approaches reduce dependence on seeding masks and leverage broader anatomical context. However, existing global approaches depend on traditional optimization schemes that are computationally inefficient, susceptible to suboptimal solutions, and often fail to generate complete tractograms, thereby limiting their practical utility ( 21 ; 14 ; 29 ) . Contributions. In this paper, we introduce GenTract ( Gen erative Tract ography), reframing tractography as a generative task, conditioned on global information, as illustrated in Figure 1 . GenTract is designed to resolve the limitations of both paradigms: it tackles the reliance of local tracking on step-wise information, while solving the computational inefficiency and optimization failures of existing global methods. Our primary contributions are: 1. We are the first to formulate tractography as a global, conditional generative task conditioned on the whole set of DWIs for a given individual. GenTract treats tractography as a sampling process, where the model draws streamlines from a learned conditional distribution, producing all streamline coordinates simultaneously. This overcomes the sequential error propagation of step-wise methods, and also eliminates the need for a seeding mask, reducing method-dependent variability. 2. We propose a modular generative architecture for global tractography compatible with multiple generative paradigms. We provide the first implementation and analysis of both Diffusion and Flow Matching for the task of global tractography, establishing the effectiveness of both schemes. 3. We conduct a comparative analysis of GenTract against traditional and state-of-the-art (SOTA) tractography paradigms. This analysis demonstrates that GenTract achieves state-of-the-art level of precision on high-fidelity data while exhibiting greater robustness in noisy, low-resolution settings 2 Related Works Local Methods. Classical local tractography methods, such as SD Stream 5 and iFOD2 35 , track streamlines by using voxel-wise diffusion information to take the next tracking step. However, they break down in complex fiber configurations like crossing or kissing ‘bundles’ 29 : collections of fibers that form neuroanatomical pathways, connecting two regions of the brain. Machine learning (ML) based techniques have increasingly advanced the field of tractography, improving accuracy over classical methods 40 . For example, the authors in 24 propose the first ML based approach, employing random forest classifiers to model the relationship between local image data and sequential tracking steps. More recently, TrackToLearn 34 reframes tractography as a reinforcement learning problem, where agents start from seed points and are guided by rewards that encourage geometrically plausible trajectories and alignment with the diffusion signal. Despite being able to produce more complete connectivity mappings, these methods still produce many false positives, especially in low-resolution data where blurred fiber orientations cause tracking drift. To address this issue, TractOracle 33 builds on TrackToLearn by incorporating anatomical priors into the reward function, providing agents with additional contextual information, aiding tracking and lowering false positives. Similarly, DDTracking 18 , a supervised diffusion-based tracking algorithm, uses neighboring voxel information to generate more reliable next-step predictions. However, these methods still propagate streamlines in an autoregressive manner and are thus limited in their robustness to underlying data perturbations. Global Methods. Existing global tractography methods 4 ; 13 address the error accumulation of local methods by treating fiber reconstruction as a single inverse problem, optimizing the entire whole-brain fiber configuration simultaneously to find the solution that best explains the underlying diffusion data. This concept is introduced in 23 , which models short fiber segments as ‘spins’ and finds the lowest-energy state for the entire system. This is improved by 8 , where these segments can move and rotate to better fit complex crossings. A major limitation is computation time, which is addressed by the authors in 28 , who introduce an efficient algorithm that