FiberLM

FiberLM is an attention-based Transformer model for mouse-brain diffusion-MRI (dMRI) tractography, developed by Ray Wen, Jiangyang Zhang, and Zifei Liang (corresponding author) at the NYU Grossman School of Medicine and released as a bioRxiv preprint in May 2026. The model addresses a long-standing limitation of dMRI tractography: streamlines reconstructed from local diffusion orientations are prone to false-positive and false-negative connections, particularly in regions of crossing or kissing fibers, because the algorithm has no anatomical prior beyond the local diffusion signal.

FiberLM tackles this by learning the statistical structure of real axonal trajectories. It is trained on a whole-brain streamline dataset derived from viral tracer experiments in the Allen Mouse Brain Connectivity Atlas (AMBCA), a mesoscale resource that maps mouse forebrain projections through thousands of anterograde tracer injections. By treating streamlines as sequences and applying self-attention, the model learns the properties of both local and long-range axonal paths, then uses this tracer-informed prior to guide step-by-step streamline propagation through new dMRI volumes.

The work sits within a broader effort at NYU and elsewhere to use mesoscopic tract-tracing data to constrain dMRI tractography in the mouse brain. Where earlier approaches used tracer data to improve fiber orientation distribution estimation, FiberLM applies a sequence model directly to the streamline-generation step, offering a learned, anatomy-aware alternative to conventional probabilistic tracking.

#Key Features

• Tracer-guided supervision: Training streamlines are derived from AMBCA viral tracer experiments, giving the model a ground-truth-anatomy prior rather than relying solely on the local diffusion signal.
• Self-attention over streamlines: A Transformer architecture lets the model capture both local geometry and long-range axonal trajectory context when propagating each streamline.
• Reduced spurious connections: Reported results indicate improved spatial agreement with AMBCA streamlines and fewer false-positive and false-negative connections than conventional probabilistic tractography.
• Fixed-checkpoint inference: Once trained, the model is applied from a fixed checkpoint to new ex vivo mouse dMRI data, predicting anatomically plausible trajectories without per-dataset retraining.

#Technical Details

FiberLM is a streamline-propagation model built on an attention-based Transformer that consumes streamline sequences and predicts the next propagation step. Supervision comes from a whole-brain streamline dataset constructed from thousands of individual AMBCA tracer injection experiments, aggregated into a representation of mouse forebrain connectivity. Through self-attention, the model learns to encode local and long-range axonal trajectory properties, which it then uses to guide tractography. The authors report that FiberLM improves spatial agreement between dMRI tractography results and AMBCA streamlines and reduces both false-positive and false-negative connections relative to conventional probabilistic tractography. The preprint does not report a public release of code or pretrained weights, and specific hyperparameters such as parameter count are not summarized here pending the full text.

#Applications

FiberLM is aimed at neuroimaging and connectomics researchers working with high-resolution ex vivo mouse-brain dMRI, where accurate reconstruction of axonal pathways is essential for studying structural connectivity, white-matter organization, and circuit-level anatomy in models of development and disease. By incorporating a tracer-derived prior, it offers a way to obtain more anatomically faithful tractograms than diffusion-only methods, which is valuable for cross-validating dMRI connectomes against gold-standard tracer connectivity and for building reference atlases of mouse brain wiring.

#Impact

FiberLM illustrates how sequence models with self-attention can be combined with gold-standard anatomical tracer data to constrain dMRI tractography, a problem where purely local, diffusion-driven methods routinely produce spurious connections. Its scope is deliberately narrow: it is specific to mouse-brain dMRI and was trained on AMBCA mouse connectivity, so it does not transfer directly to human imaging or other species. As a bioRxiv preprint (license cc_no) that has not yet undergone peer review, with no public code or weights at the time of release, its broader adoption and independent validation remain to be established. The work was supported by NIH grant R01NS102904.