ilex-hub/neurostorm.ratio0.8.1

NeuroSTORM (4D Swin + Mamba MAE foundation model for fMRI) -- NeuroSTORM mask_ratio=0.8 pretraining checkpoint

Description

NeuroSTORM (Zhang et al., Nature Biomedical Engineering 2026) is a general-purpose foundation model for fMRI analysis. The architecture is a 4D Swin Transformer with Mamba selective-scan state-space models as the per-window mixer (in place of the conventional window attention) and a masked-autoencoder (MAE) pretraining objective.

v0 ships two variants corresponding to the upstream's two released mask-ratio pretraining checkpoints (0.5 and 0.8), sharing parameter-identical architecture (7,712,892 trainable scalars each); only the training-time mask ratio differs.

The 4D operates on the full (96, 96, 96, 20) voxel-time volume (MNI152 brain-only resampling of fMRI BOLD); the resulting embedding is the deepest-stage feature map (288, 2, 2, 2, 20) -- consumed by downstream task heads (sex/age regression, phenotype prediction, disease diagnosis) per the upstream's benchmark.

Intended use

fMRI foundation backbone for representation learning. Input: (1, 96, 96, 96, 20) MNI152 BOLD clip (consumer handles registration + min-max normalisation + quantisation per the upstream pipeline). Output: deepest- stage encoder feature map (288, 2, 2, 2, 20) suitable for downstream task heads (linear probe, attention pool, MLP). The MAE decoder is loaded into the bundle for state-dict round-trip integrity but is not invoked at inference.

Usage

from ilex.models.neurostorm import NeuroSTORM
model = NeuroSTORM.from_pretrained('ilex-hub/neurostorm.ratio0.8.1')

Authors

Zhang Y. et al. (CUHK-AIM-Group, Chinese University of Hong Kong)

Citation

Zhang Y. et al. (2026). Towards a general-purpose foundation model for fMRI analysis. Nature Biomedical Engineering. doi:10.1038/s41551-026-01666-y.

References

• Zhang Y. et al. (2026). Towards a general-purpose foundation model for fMRI analysis. Nature Biomedical Engineering. doi 10.1038/s41551-026-01666-y.
• Gu A., Dao T. (2023). Mamba -- Linear-time sequence modeling with selective state spaces. arXiv 2312.00752.
• Liu Z. et al. (2021). Swin Transformer -- Hierarchical Vision Transformer using Shifted Windows. arXiv 2103.14030.
• Upstream code -- github.com/CUHK-AIM-Group/NeuroSTORM (Apache-2.0); upstream weights -- huggingface.co/zxcvb20001/NeuroSTORM (MIT).

License

HF Hub license tag: mit

Effective terms: MIT (on the released checkpoint weights, per the HF model card at huggingface.co/zxcvb20001/NeuroSTORM). The upstream code (github.com/CUHK-AIM-Group/NeuroSTORM) is separately Apache-2.0; both are fully permissive. The ilex JAX / Equinox port code is licensed under Apache-2.0 / GPL-3.0.

Upstream license reference: https://opensource.org/licenses/MIT

Copyright

NeuroSTORM is copyright (c) CUHK-AIM-Group 2026, Apache-2.0- licensed on the code (github.com/CUHK-AIM-Group/NeuroSTORM, LICENSE file) and MIT-licensed on the released checkpoints (huggingface.co/zxcvb20001/NeuroSTORM, license tag mit). The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0; license restrictions on the upstream artefacts are preserved through the canonical bundle's _ilex.origin = 'pytorch' provenance.

Upstream source

Original weights / reference implementation: https://github.com/CUHK-AIM-Group/NeuroSTORM

Provenance

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