brainiac.1 / README.md
rciric's picture
Initial upload via tools/push_to_hf.py (architecture: ilex.models.brainiac.BrainIAC)
ac9ff15 verified
metadata
library_name: ilex
tags:
  - jax
  - equinox
  - ilex
  - neuroimaging
  - foundation
license: other
license_name: brainiac-research-only
license_link: https://github.com/AIM-KannLab/BrainIAC/blob/main/LICENSE
gated: auto
extra_gated_prompt: |
  The BrainIAC family is distributed under the BrainIAC
  Research-Only License (Mass General Brigham, 2026). Access is
  granted solely for non-commercial academic research and
  educational use. Commercial use -- including clinical decision
  support, clinical workflows, products, or services for a fee --
  is expressly reserved by MGB and requires separate written
  authorization.

  Acceptance is per-repository on HuggingFace Hub; HF does not
  extend acceptance across all BrainIAC variants automatically.
  The terms below are identical for every BrainIAC variant in
  the ilex-hub catalog: if you have already accepted them for
  another variant, the same answers apply here.

  By requesting access you affirm that your intended use falls
  within the Research-Only terms and that you have read the
  license linked from this model card.
extra_gated_fields:
  Name: text
  Affiliation: text
  Email: text
  Intended use: text
  I acknowledge the BrainIAC Research-Only License terms: checkbox
extra_gated_button_content: Acknowledge and download

BrainIAC (3D ViT-B/16 foundation backbone) -- SimCLR-pretrained backbone (768-dim CLS embedding)

Description

BrainIAC ViT-3D foundation backbone ported to JAX / Equinox from the AIM-KannLab reference PyTorch implementation. The network is a ViT-B/16 with cubic 16x16x16 patches over a 96x96x96 single-channel input, pretrained with SimCLR on a large multi-cohort brain-MRI corpus. Built on monai.networks.nets.ViT with classification=False, so the architecture has no CLS token; the forward pass returns a 768-dimensional embedding read out from the first patch token after the global LayerNorm (matching upstream ViTBackboneNet.forward's features[0][:, 0]). Downstream heads consume this embedding (brain age, IDH classification, MR-sequence multiclass, OS regression, stroke detection, MCI classification, segmentation via UNETR); ilex v0 ships the backbone only.

Intended use

Contrast-agnostic feature embedding for 3D brain MRI volumes at 96x96x96. Inputs must be skull-stripped, atlas-registered, and intensity-normalised per the upstream preprocessing pipeline (HD-BET skull strip + NIHPD atlas registration + MONAI NormalizeIntensityd, all out of scope for ilex v0).

Usage

from ilex.models.brainiac import BrainIAC
model = BrainIAC.from_pretrained('ilex-hub/brainiac.1')

Authors

Tak D., Garomsa B. A., Zapaishchykova A., et al.

Citation

Tak D., Garomsa B. A., Zapaishchykova A., et al. (2026). A generalizable foundation model for analysis of human brain MRI. Nature Neuroscience. doi:10.1038/s41593-026-02202-6

References

  • Tak D., Garomsa B. A., Zapaishchykova A., et al. (2026). A generalizable foundation model for analysis of human brain MRI. Nature Neuroscience. doi:10.1038/s41593-026-02202-6

License

HF Hub license tag: other HF Hub license slug: brainiac-research-only

Effective terms: BrainIAC Research-Only License (c) 2026 Mass General Brigham. Permission is granted to use, copy, modify, and distribute the software and weights solely for non-commercial academic research and educational purposes. Commercial applications (products, services for a fee, clinical workflows, clinical decision support, healthcare operations) are expressly reserved and require separate written authorization from MGB. The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0 for the port code itself, but the ilex license does NOT override or re-license the upstream Research-Only terms governing the weights or the architecture; non-commercial use only continues to apply.

Upstream license reference: https://github.com/AIM-KannLab/BrainIAC/blob/main/LICENSE

Copyright

Network architecture, training code, and pretrained weights: copyright (c) 2026 Mass General Brigham, distributed under the BrainIAC Research-Only License (non-commercial academic research and educational use only; commercial rights expressly reserved -- see https://github.com/AIM-KannLab/BrainIAC/blob/main/LICENSE). JAX / Equinox port code: copyright (c) the ilex authors, released under the Apache-2.0 / GPL-3.0 dual license used by ilex itself. The ilex port's license covers only the original Equinox port code; it does NOT override, supersede, or re-license the upstream BrainIAC Research-Only License. The pretrained weights and the BrainIAC architecture (regardless of the JAX / Equinox re-expression here) remain governed by the upstream research-only terms; commercial use of either is expressly reserved by Mass General Brigham and requires separate written authorization from MGB.

Upstream source

Original weights / reference implementation: https://github.com/AIM-KannLab/BrainIAC

Provenance

This artefact was produced by ilex's save/load pipeline. The architecture is implemented in ilex.models.brainiac.BrainIAC and the weights have been converted from their upstream format. See the upstream source above for the canonical reference.