Tractogram foundation model

Researchers introduce TractFM, a foundation model that learns reusable representations from whole-brain diffusion MRI tractography data by combining local streamline encoding with permutation-equivariant processing. The model demonstrates strong transfer learning capabilities across different tractography algorithms, datasets, and prediction tasks, achieving accurate tract parcellation and demographic predictions without task-specific fine-tuning.

TractFM represents a meaningful advancement in neuroimaging AI by addressing a fundamental limitation in how brain white-matter pathways are analyzed. Traditional approaches compartmentalize the problem: streamline classifiers focus on geometric patterns in isolation while subject-level predictions rely on manually engineered features. This fragmentation prevents learning unified representations that capture both local anatomical details and global brain organization patterns across populations.

The foundation model approach mirrors successful strategies in computer vision and NLP, where pretraining on large-scale unlabeled or weakly-labeled data produces representations transferable to diverse downstream tasks. TractFM's architecture elegantly handles the unique structure of tractography data—unordered sets of three-dimensional curves—through permutation-equivariant design, ensuring the model produces consistent outputs regardless of streamline ordering.

The validation results span three tractography algorithms and five independent dMRI datasets, demonstrating genuine generalization rather than dataset-specific performance. The model's ability to predict age and sex from frozen representations indicates it learns biologically meaningful patterns correlated with brain development and sexual dimorphism, validating that learned representations capture real neurobiology.

For neuroscience and neuroimaging research, this work enables more efficient analysis pipelines and could accelerate development of biomarkers for neurological and psychiatric conditions. The foundation model approach reduces computational burden on individual research groups and democratizes access to sophisticated analysis methods. Future applications may include disease prediction, clinical outcome prognosis, and understanding how brain connectivity relates to cognitive and behavioral traits.

• →TractFM learns unified representations connecting individual streamline geometry with whole-brain anatomical organization through permutation-equivariant architecture.
• →The foundation model generalizes across three different tractography algorithms and five independent datasets without task-specific retraining.
• →Frozen representations accurately predict demographic variables like age and sex, demonstrating biological relevance of learned features.
• →Pretraining on tractogram parcellation yields both streamline-level and subject-level embeddings useful for complementary analysis tasks.
• →Foundation model approach reduces reliance on hand-crafted neuroimaging features and enables efficient knowledge transfer across research pipelines.