TractFM: Foundation model learns brain wiring patterns directly from tractograms
One model that works across 3 tractography pipelines and 5 brain datasets to parcellate pathways and predict phenotypes.
Diffusion MRI tractography is the only noninvasive method to map white-matter pathways in the living brain, generating tractograms—large unordered sets of 3D streamlines. Existing methods separate streamline classification from subject-level prediction, relying on hand-crafted features and failing to learn reusable representations that connect streamline anatomy with whole-brain variation. To bridge this gap, researchers from Harvard, EPFL, and other institutions developed TractFM, a tractogram foundation model that processes all streamlines from a subject in a single forward pass using a permutation-equivariant encoder.
TractFM was pretrained on dense anatomical tract parcellation (assigning labels to individual streamlines) and then frozen for downstream tasks. Across three different tractography algorithms (e.g., deterministic, probabilistic) and five independent dMRI datasets, its representations generalized without retuning. TractFM achieved accurate tract parcellation and predicted subject age and sex with high fidelity, demonstrating that whole-brain geometric context learned once can transfer across pipelines, datasets, and prediction tasks. This represents a significant step toward unified representation learning in brain imaging.
- TractFM combines a local streamline encoder with a permutation-equivariant tractogram encoder for joint whole-brain processing in one forward pass.
- Pretrained on dense anatomical tract parcellation, it yields both streamline-level embeddings for segmentation and compact subject-level descriptors for phenotype prediction.
- Achieves accurate tract parcellation and age/sex prediction across 3 tractography algorithms and 5 independent dMRI datasets without fine-tuning.
Why It Matters
A unified brain-wiring model could automate clinical tractography analysis, accelerating diagnosis of neurological disorders and improving cross-study reproducibility.