TractSeg -- white matter bundle / endings segmentation (2D U-Net + 3-axis ensemble) -- TractSeg tract_segmentation v3 (42 bundles)

Description

JAX / Equinox port of TractSeg (Wasserthal et al., NeuroImage 2018; github.com/MIC-DKFZ/TractSeg). A 2D U-Net with deep supervision that, per slice, predicts binary masks for each white-matter bundle (or each bundle's endpoint regions) from a 9-channel FOD-peak input volume. The upstream's canonical inference path runs the per-slice forward on all three principal slice orientations (axial, coronal, sagittal) and averages the three resulting 3D mask predictions; the ilex TractSegPipeline class wraps that end to end.

Two task variants ship in v0:

  • tract_segmentation_v3 (42 bundles, the upstream's current default)
  • endings_segmentation_v4 (72 endpoint regions)

Both share parameter-identical bodies (~37 M trainable scalars); only the three 1x1 head Conv2ds at output_2, output_3, and conv_5 differ in out_channels to match the task's n_classes.

Intended use

Predict per-bundle binary masks for 42 white-matter bundles from a 9-channel FOD-peak input volume. The TractSegPipeline 3-axis ensemble produces a 4D (42, D, H, W) sigmoid probability volume; consumers threshold + run downstream tractography filtering. Zenodo record 3634539, epoch 266.

Usage

from ilex.models.tract_seg import TractSeg
model = TractSeg.from_pretrained('ilex-hub/tract_seg.tract-segmentation-v3.1')

Authors

Wasserthal J., Neher P. F., Maier-Hein K. H. (MIC-DKFZ, German Cancer Research Center, Heidelberg University)

Citation

Wasserthal J., Neher P. F., Maier-Hein K. H. (2018). TractSeg -- Fast and accurate white matter tract segmentation. NeuroImage 183, 239-253. doi:10.1016/j.neuroimage.2018.07.070.

References

  • Wasserthal J., Neher P. F., Maier-Hein K. H. (2018). TractSeg -- Fast and accurate white matter tract segmentation. NeuroImage 183, 239-253. doi 10.1016/j.neuroimage.2018.07.070.
  • Wasserthal J., Neher P. F., Hirjak D., Maier-Hein K. H. (2020). Multiparametric mapping of white matter microstructure in 100,000 subjects from the UK Biobank.
  • Upstream code, license, training scripts -- github.com/MIC-DKFZ/TractSeg (Apache-2.0).
  • Upstream weights -- Zenodo records 3634539 (tract_segmentation v3, epoch 266) and 3518348 (endings_segmentation v4, epoch 220).

License

HF Hub license tag: apache-2.0

Effective terms: Apache-2.0 (MIC-DKFZ) on both the network code (github.com/MIC-DKFZ/TractSeg) and the pretrained weights (Zenodo records 3634539 and 3518348). The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.

Upstream license reference: https://www.apache.org/licenses/LICENSE-2.0

Copyright

TractSeg is copyright (c) MIC-DKFZ (German Cancer Research Center), Apache-2.0-licensed on both the network code (github.com/MIC-DKFZ/TractSeg) and the pretrained weights (Zenodo records 3634539 + 3518348). The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.

Upstream source

Original weights / reference implementation: https://github.com/MIC-DKFZ/TractSeg

Provenance

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

Downloads last month
16
Safetensors
Model size
37M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support