ilex-hub/supersynth.1

SuperSynth (multi-task brain-imaging foundation U-Net) -- SuperSynth (August 2025 checkpoint; backbone + 9 heads)

Description

SuperSynth (Iglesias / Liu et al. 2025), ported to JAX / Equinox from the upstream PyTorch release shipped with FreeSurfer (mri_super_synth). A modality- and resolution-agnostic 3D brain U-Net that emits a set of predictions from a single shared "frugal" backbone. This port is the network forward of frugal_models.ssynth_inference: the backbone plus 9 task heads -- reg (MNI deformation), seg (56-class segmentation incl. WM lesions, limbic and extracerebral structures), T1/T2/FLAIR (super-resolution synthesis), and LP/LW/RP/RW (left/right pial/white surface-distance maps). The separate AutoQC sub-network and the upstream pre/post-processing are out of scope for v0.

Intended use

Multi-task brain-imaging foundation model: from a single single-channel 3D brain scan (any contrast / resolution; the upstream resamples to 1 mm and normalises, spatial dims a multiple of 16) it returns 9 head outputs -- reg (3-ch MNI deformation), seg (56-class segmentation), T1/T2/FLAIR (super-resolution synthesis), and LP/LW/RP/RW (left/right pial/white surface-distance maps). The v0 bundle is the network forward only; the AutoQC sub-net and the upstream pre/post-processing (resample, label remapping, surface placement, atlas registration) are not vendored.

Usage

from ilex.models.supersynth import SuperSynth
model = SuperSynth.from_pretrained('ilex-hub/supersynth.1')

Authors

Iglesias J. E.; Liu et al.

Citation

Iglesias J. E., Liu et al. (2025). SuperSynth (FreeSurfer mri_super_synth). https://surfer.nmr.mgh.harvard.edu/fswiki/SuperSynth. Backbone family (pytorch-3dunet): Wolny A., et al. Accurate and versatile 3D segmentation of plant tissues at cellular resolution. eLife 2020;9:e57613.

References

• SuperSynth (FreeSurfer): https://surfer.nmr.mgh.harvard.edu/fswiki/SuperSynth ; code: https://github.com/freesurfer/freesurfer/tree/dev/mri_super_synth
• Backbone family (pytorch-3dunet): Wolny A., et al. eLife 2020. https://github.com/wolny/pytorch-3dunet

License

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

Effective terms: FreeSurfer Software License: academic / non-commercial use only. SuperSynth ships with FreeSurfer (mri_super_synth); the weights are lazy-fetched from the FreeSurfer FTP on first run. 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 and pretrained weights: copyright (c) the SuperSynth / 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/SuperSynth

Provenance

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