Brain2Vec: open-source 3D foundation model for structural brain MRI released
Researchers pre-trained a self-supervised model on 74,425 T1-weighted MRI scans, reporting gains in brain age prediction and benchmark rankings, with code and models openly available.
In plain words
A foundation model is an AI system pre-trained on large amounts of data so it can be adapted to many smaller tasks. Brain2Vec was pre-trained on tens of thousands of structural brain MRI scans and can either be fine-tuned for a specific job or used as-is to produce numerical summaries of scans. The authors report it improved accuracy and training speed in their tests and released it openly.
What was announced
The team behind Brain2Vec, described in an arXiv abstract, says the model addresses a common problem in medical imaging: labeled data is scarce and training deep learning models end-to-end is computationally demanding. The model was pre-trained self-supervised on 74,425 T1-weighted MRI scans pooled from multiple studies, and can serve as a weight-initialization backbone for fine-tuning or as a standalone feature extractor.
In the reported evaluations, fine-tuning from Brain2Vec's pretrained weights improved final accuracy and convergence speed on a brain age prediction task compared with random initialization. Its frozen embeddings beat anatomically defined volumetric features on four of five downstream classification tasks and were competitive with a state-of-the-art foundation model trained on substantially more data. On the independent BrainFMBench benchmark, it ranked first among six pretrained brain MRI foundation models on 9 of 15 sex, age, and BMI prediction tasks across five cohorts. The authors also report the model encodes information from spatially localized, neuroanatomically meaningful regions, and that altering subsets of latent features produces monotonic, region-specific changes in reconstructed anatomy. Code and trained models are available on GitHub, and features can also be derived through the NiChart cloud platform without local installation.
What it does not show
- All results come from the authors' own abstract; no independent evaluation or peer review is reflected in this excerpt.
- The excerpt covers only the abstract, so task details, methods, cohort definitions and failure cases are not visible.
- Benchmark rankings and accuracy comparisons are as reported by the authors and are not independently confirmed.
For security and IT teams
- Access
- Brain2Vec can be installed locally from its open-source GitHub package or used via the NiChart cloud platform, where features can be derived without local installation or specialized infrastructure.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.