LifespanStrip (age-aware lifespan brain extraction) -- LifespanStrip v1 (Lifespan_Skull_Stripping.pt)

Description

CNN+ViT+STN+flow composite that performs age-aware brain extraction by registering an age-specific atlas onto a subject T1w MRI. Wang et al. (Nat Biomed Eng 9:700-715, 2025) the standalone "Atlases-empowered Lifespan Skull Stripping" model from DBC-Lab (UNC). Architecture: three-scale ViT trunks (hidden=768/128/16, num_heads=12/8/4) feeding a CNN-encoder/decoder U-Net (channels 16->32->64->128->256, in-house UnetBasicBlock variant with InstanceNorm3d(affine=False), no residual skip), plus a rigid STN (Conv stack + 71M-param FC layer + manual cos/sin Euler-angle matrix assembly) and a VoxelMorph-style flow head (Conv encoder/decoder + VecInt scaling-and-squaring SVF integration + dense warp via map_coordinates). 159M params total; largest single block is the FC localizer (256 x 13**3 -> 128 features = 72M params). Age-awareness is achieved at inference by selecting one of 11 atlas pairs (Month0/3/6/9/12/18/24/Adolescent/ Adult/Elder/Year65-80) based on a user-supplied age_in_month; a single shared set of network weights covers the full lifespan. Atlases ship via lytemaps as TemplateFlow-spec data, not in the ilex code repo. The two-pass inference pipeline (coarse mask -> crop + center -> refine) lives in ilex.models.lifespan_strip.pipeline.

Intended use

Brain-extract a 3D T1-weighted MRI volume conditional on an age-specific atlas. Input is a single-channel 128x128x128 volume at 2 mm isotropic, intensity in [0, 1], plus the matching atlas image + binary mask at the same spec. The caller selects one of 11 atlas buckets (Month0/3/6/9/12/18/24/Adolescent/Adult/Elder/Year65-80) based on the subject's age. Output is a six-tuple per NET.forward (logits, moving_trans, y_source, pos_flow, x_reg, atlas_mask_formable); the primary output is logits[1] (soft brain probability), thresholded at 0.5 after the two-pass pipeline for the binary mask. The preprocessing pipeline (RAI orient + N4 + resample + histogram-match + min-max) is the caller's concern (delegated to ilex.models.lifespan_strip.pipeline).

Usage

from ilex.models.lifespan_strip import LifespanStrip
model = LifespanStrip.from_pretrained('ilex-hub/lifespan_strip.1')

Authors

Li Wang, Yue Sun, Jakob Seidlitz, Richard A. I. Bethlehem, Aaron Alexander-Bloch, Lena Dorfschmidt, Gang Li, Jed T. Elison, Weili Lin

Citation

Wang L., Sun Y., Seidlitz J., Bethlehem R. A. I., Alexander-Bloch A., Dorfschmidt L., Li G., Elison J. T., Lin W., Wang L. (2025). A lifespan-generalizable skull-stripping model for magnetic resonance images that leverages prior knowledge from brain atlases. Nature Biomedical Engineering 9:700-715.

References

  • Wang L., Sun Y., Seidlitz J., Bethlehem R. A. I., Alexander-Bloch A., Dorfschmidt L., Li G., Elison J. T., Lin W., Wang L. (2025). A lifespan-generalizable skull-stripping model for magnetic resonance images that leverages prior knowledge from brain atlases. Nature Biomedical Engineering 9, 700-715. DOI: 10.1038/s41551-024-01337-w.
  • Upstream repo: https://github.com/DBC-Lab/Atlases-empowered_Lifespan_Skull_Stripping
  • Architecture: monai.networks.nets.ViT (hidden=768/128/16, perceptron pos-embed) + in-house UnetBasicBlock/UnetUpBlock (networks/blocks.py) + rigid STN with cos/sin Euler matrix assembly + VoxelMorph layers (layers.py: VecInt, SpatialTransformer, ResizeTransform).

License

HF Hub license tag: mit

Effective terms: MIT (Developing Brain Computing Lab, UNC -- Wang et al.) on both the network code and the Lifespan_Skull_Stripping.pt weights. The 11 age-specific atlas pairs are shipped via lytemaps under the same MIT terms; we host them on OSF as TemplateFlow-spec data, not in the ilex repo. The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.

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

Copyright

Network architecture and pretrained weights -- copyright (c) Developing Brain Computing Lab (DBC-Lab, UNC), released under the MIT 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.

Upstream source

Original weights / reference implementation: https://github.com/DBC-Lab/Atlases-empowered_Lifespan_Skull_Stripping

Provenance

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

Downloads last month
2
Safetensors
Model size
0.2B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support