ilex-hub/wmh_synthseg.1

WMH-SynthSeg (joint anatomy + white-matter-hyperintensity seg) -- WMH-SynthSeg v10 (231110 checkpoint)

Description

WMH-SynthSeg (Laso et al. 2024), ported to JAX / Equinox from the upstream PyTorch release shipped with FreeSurfer. A SynthSeg-style domain-randomised 3D U-Net that jointly segments brain anatomy and white-matter hyperintensities from MRI of any contrast and resolution, including low-field portable scanners. The backbone is the UNet3D (DoubleConv, GroupNorm, nearest-interp upsampling) from Adrian Wolny's pytorch-3dunet library. The network emits 39 raw output channels: 33 segmentation classes (background, the WMH label, optic chiasm, and the standard left/right subcortical, cortical, and ventricular structures) plus 6 auxiliary regression channels from the joint training objective; segmentation uses the first 33 (softmax + argmax).

Intended use

Joint segmentation of brain anatomy and white-matter hyperintensities, plus per-structure / WMH volumetry, from heterogeneous-contrast and low-field MRI. Inputs are single-channel intensity volumes normalised to [0, 1] and resampled to 1 mm isotropic, with each spatial dim a multiple of 16 (the upstream pipeline pads to a multiple of 32). The v0 bundle is the network forward only: it returns 39 raw output channels; segmentation is softmax + argmax over the first 33 (WMH posterior is the channel for FreeSurfer label 77). The mri_WMHsynthseg pre/post-processing (resample, crop, flip-TTA, volumetry CSV) is not vendored in v0.

Usage

from ilex.models.wmh_synthseg import WMHSynthSeg
model = WMHSynthSeg.from_pretrained('ilex-hub/wmh_synthseg.1')

Authors

Laso P., et al.

Citation

Laso P., Cerri S., Sorby-Adams A., Guo J., Matteen F., Goebl P., Wu J., Li H., Young S. I., Billot B., Puonti O., Rosen M. S., Kirsch J., Strisciuglio N., Wolterink J. M., Eshaghi A., Barkhof F., Kimberly W. T., Iglesias J. E. (2024). Quantifying white matter hyperintensity and brain volumes in heterogeneous clinical and low-field portable MRI. IEEE ISBI 2024. doi:10.1109/ISBI56570.2024.10635502; arXiv:2312.05119. Backbone: Wolny A., et al. (2020). Accurate and versatile 3D segmentation of plant tissues at cellular resolution. eLife 9:e57613.

References

• Laso P., Cerri S., Sorby-Adams A., Guo J., Matteen F., Goebl P., Wu J., Li H., Young S. I., Billot B., Puonti O., Rosen M. S., Kirsch J., Strisciuglio N., Wolterink J. M., Eshaghi A., Barkhof F., Kimberly W. T., Iglesias J. E. (2024). Quantifying white matter hyperintensity and brain volumes in heterogeneous clinical and low-field portable MRI.
• Backbone library (pytorch-3dunet): Wolny A., et al. Accurate and versatile 3D segmentation of plant tissues at cellular resolution. eLife 2020. https://github.com/wolny/pytorch-3dunet
• Distributed with FreeSurfer as mri_WMHsynthseg.

License

HF Hub license tag: other HF Hub license slug: freesurfer-software-license

Effective terms: The pytorch-3dunet backbone code is MIT (Adrian Wolny). The WMH-SynthSeg pretrained weights are distributed with FreeSurfer under the FreeSurfer Software License: academic / non-commercial use only. Pending explicit upstream confirmation that re-hosting the weights on HF Hub is acceptable. Resolve before flipping published: true. The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0; that does not override the upstream FreeSurfer weight terms.

Upstream license reference: https://surfer.nmr.mgh.harvard.edu/fswiki/FreeSurferSoftwareLicense

Copyright

Network architecture (pytorch-3dunet): copyright (c) Adrian Wolny, MIT License. Pretrained WMH-SynthSeg weights: copyright (c) the WMH-SynthSeg / FreeSurfer authors, distributed under the FreeSurfer research-only license. JAX / Equinox port: copyright (c) the ilex authors, released under the Apache-2.0 / GPL-3.0 dual license used by ilex itself. The ilex license does not override the upstream FreeSurfer weight terms.

Upstream source

Original weights / reference implementation: https://surfer.nmr.mgh.harvard.edu/fswiki/WMH-SynthSeg

Provenance

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