library_name: ilex
tags:
- jax
- equinox
- ilex
- neuroimaging
- whole-brain
license: cc-by-nc-4.0
license_link: https://creativecommons.org/licenses/by-nc/4.0/
gated: auto
extra_gated_prompt: |
HD-BET v2's pretrained weights are distributed under
Creative Commons Attribution-NonCommercial 4.0
International (CC BY-NC 4.0): non-commercial academic,
research, or educational use only; no commercial use of
the weights. The HD-BET code itself is Apache-2.0, but
the weights carry separate non-commercial terms.
By requesting access you affirm that your intended use
is non-commercial academic, research, or educational;
that you will not use or redistribute the weights for
commercial purposes; 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 upstream CC BY-NC weights license: checkbox
extra_gated_button_content: Acknowledge and download
HD-BET (brain extraction) -- HD-BET v2 (single fold_all checkpoint)
Description
HD-BET (Brain Extraction Tool) is an nnUNet-derived 3D segmentation network for brain extraction from multi-sequence MRI (T1w, T2w, FLAIR, CT). This ilex bundle ports HD-BET v2 (Zenodo release_v1.5.0.zip; despite the file name this is the v2 release, per the upstream paths.py note) to JAX / Equinox. The architecture is nnUNet PlainConvUNet -- 6 encoder stages with channel counts (32, 64, 128, 256, 320, 320), per-stage 3x3x3 convs with InstanceNorm + LeakyReLU(0.01), asymmetric strides at the bottleneck ((1, 2, 2) at stage 5), and 5 decoder stages with ConvTranspose3d upsampling and per-decoder-stage deep-supervision seg heads (inference uses the topmost head only). The upstream pipeline performs crop-to- nonzero, resampling to 1 mm isotropic spacing, and z-score normalisation before forward, then runs sliding-window inference at tile_step_size=0.5 with Gaussian- weighted overlap and mirror-axis test-time augmentation; these are scoped for v0 alongside the model package but land in subsequent commits.
Intended use
Brain extraction (skull stripping) of 3D structural MRI volumes (T1w, T2w, FLAIR, CT). Inputs must be in approximate MNI152 orientation; consumers should apply fslreorient2std (or equivalent) before forward. The upstream pipeline performs nonzero cropping, resampling to 1 mm isotropic spacing, and z-score intensity normalisation; vendored alongside the model package in ilex.models.hd_bet for v0. Inference uses sliding-window with Gaussian-weighted overlap (tile_step_size=0.5) and mirror-axis test-time augmentation by default.
Usage
from ilex.models.hd_bet import HDBet
model = HDBet.from_pretrained('ilex-hub/hd_bet.2')
Authors
Isensee F., Schell M., Tursunova I., Brugnara G., Bonekamp D., Neuberger U., Wick A., Schlemmer H. P., Heiland S., Wick W., Bendszus M., Maier-Hein K. H., Kickingereder P.
Citation
Isensee F., Schell M., Tursunova I., Brugnara G., Bonekamp D., Neuberger U., Wick A., Schlemmer H. P., Heiland S., Wick W., Bendszus M., Maier-Hein K. H., Kickingereder P. (2019). Automated brain extraction of multi-sequence MRI using artificial neural networks. Human Brain Mapping 40(17):4952-4964. doi:10.1002/hbm.24750. The nnUNet backbone the upstream wraps: Isensee F., Jaeger P. F., Kohl S. A., Petersen J., Maier-Hein K. H. (2021). nnU-Net: a self- configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2):203-211. doi:10.1038/s41592-020-01008-z
References
- Isensee F., Schell M., Tursunova I., Brugnara G., Bonekamp D., Neuberger U., Wick A., Schlemmer H. P., Heiland S., Wick W., Bendszus M., Maier-Hein K. H., Kickingereder P. (2019). Automated brain extraction of multi-sequence MRI using artificial neural networks. Human Brain Mapping 40(17):4952-4964. doi:10.1002/hbm.24750
- Isensee F., Jaeger P. F., Kohl S. A., Petersen J., Maier-Hein K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18(2):203-211. doi:10.1038/s41592-020-01008-z
License
HF Hub license tag: cc-by-nc-4.0
Effective terms: The HD-BET code (https://github.com/MIC-DKFZ/HD-BET) is distributed under Apache-2.0; the pretrained weights (Zenodo record 14445620 / file release_v1.5.0.zip) are separately distributed by MIC-DKFZ under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). Non-commercial academic / research / educational use only; commercial use of the weights requires separate authorisation from the upstream authors. 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 CC BY-NC terms governing the weights. Note that the dataset.json shipped with the weights additionally declares its licence as "internal, talk to Fabian" -- consistent with the CC BY-NC 4.0 Zenodo declaration but confirming the upstream's expectation that commercial / redistribution use is gated through direct contact with the upstream authors.
Upstream license reference: https://creativecommons.org/licenses/by-nc/4.0/
Copyright
Network architecture and training code: copyright (c) the HD-BET contributors, distributed under the Apache-2.0 license at https://github.com/MIC-DKFZ/HD-BET. Pretrained weights (Zenodo record 14445620 / file release_v1.5.0.zip) are distributed by MIC-DKFZ under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) -- non-commercial academic / research use only; commercial use of the weights requires separate authorisation from the upstream authors. 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 non-commercial terms governing the pretrained weights.
Upstream source
Original weights / reference implementation: https://github.com/MIC-DKFZ/HD-BET
Provenance
This artefact was produced by ilex's
save/load pipeline. The architecture is implemented in
ilex.models.hd_bet.HDBet and the weights have been converted
from their upstream format. See the upstream source above
for the canonical reference.