FiberLM: A Transformer-Based Model for Mouse Brain Diffusion MRI Tractography Guided by Viral Tracer Data.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Connectomics

Background:

• Diffusion MRI (dMRI) tractography maps brain structural connectivity non-invasively.
• Conventional tractography suffers from significant false-positive and false-negative connections.
• Existing deep learning methods often inherit limitations from conventional tractography training data.

Purpose of the Study:

• Introduce FiberLM, an attention-based Transformer model for mouse brain tractography.
• Improve the accuracy of mapping axonal trajectories using high-resolution dMRI data.
• Leverage viral tracer data for robust model training.

Main Methods:

• Trained FiberLM on a whole-brain streamline dataset derived from Allen Mouse Brain Connectivity Atlas (AMBCA) viral tracer data.
• Utilized self-attention mechanisms to learn local and long-range axonal trajectory properties.
• Applied FiberLM to ex vivo high-resolution mouse brain dMRI data.

Main Results:

• FiberLM significantly reduced false-positive and false-negative connections compared to conventional methods.
• Achieved improved spatial agreement with tracer-defined pathways.
• Generated whole-brain connectomes more closely approximating AMBCA results.

Conclusions:

• FiberLM demonstrates superior performance in mouse brain tractography.
• The model offers a potential tool for accurate reconstruction of mouse brain structural connectomics.
• This approach overcomes limitations of conventional dMRI tractography and deep learning methods relying on similar data.