GenTract: Generative Global Tractography

TL;DR

GenTract reframes tractography as a conditional global generative task that maps the full diffusion MRI volume to a tractogram by learning $p( ext{streamlines}| ext{dMRI})$ and generating all coordinates in parallel. It uses a two-part architecture: an Anatomical Conditioning Encoder that produces a global context tensor $oldsymbol{z}$ from the SH-fODF representation, and a Conditional Transformer that outputs streamline coordinates conditioned on $oldsymbol{z}$ under either the Diffusion objective $ abla$ or Flow Matching objective $ abla$. Compared with state-of-the-art baselines, GenTract achieves a precision improvement of a factor of $2.1$ over TractOracle on high-quality data and shows robustness to low-resolution and noise by orders of magnitude relative to closest competitors, while maintaining fast inference times. By removing the need for seeding masks and mitigating autoregressive error propagation, GenTract offers a practical, high-precision global tractography solution with potential for clinical diffusion MRI adoption.

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 2.1x 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.

Figures (5)

• 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.
• Figure 2: Overview of the GenTract framework. A: VAEs encode fODF coefficients into latent representations. B(1) and B(2): The training protocol for Diffusion and FM models respectively, using the learned $z^{(i)}$ as input. Losses are back-propagated through both the generative model and the class-conditioned encoder $\mathcal{E}^c$. C: The inference process, where streamlines are generated by sampling from Gaussian noise conditioned on $\mathbf{z}$.
• Figure 3: AFQ streamlines retained (AFQ % P) vs. Computational Time by Inference Step for the Diffusion ($M=8$, $n=256$) model. The size of each bubble is proportional to the number of DDIM inference steps used.
• Figure 4: Qualitative result showing SLFR segmented by BundleSeg for each method, on original, noised, and LR+Noise for a HCP test subject.
• Figure 5: Computational time comparison for all methods. Bars show inference time (seconds) (mean $\pm$ std). The std for GenTract is very small.