ilex-hub/synthdist.1

SynthDist (cortical-surface signed-distance-function U-Net) -- SynthDist v1.0 (synthsurf_v10_230420)

Description

A 5-level neurite U-Net (SynthSeg-lineage architecture) that predicts a 9-channel signed-distance map at the input spatial resolution from a single T1-weighted (or contrast-clipped CT) brain volume. The network is the SynthDist component of FreeSurfer 8.x's recon-all-clinical stream -- it replaces the iterative Bayesian surface placement of standard recon-all with a learned distance-map predictor, enabling topologically-accurate cortical surface reconstruction from clinical scans of arbitrary orientation, resolution, and contrast. Channels 0/1/2/3 of the output are the LH-WM / LH-Pial / RH-WM / RH-Pial signed distances (negative inside the surface, zero on the surface, positive outside); the trunk emits five additional channels that the recon-all-clinical wrapper does not consume. The architecture is identical in body to SynthSeg / SynthSR / FSMSeg (same ilex.nimox.architectures.UNet3dNeurite primitive), differing only in depth (nb_levels=5), output width (nb_labels=9), and head activation (linear, since distances are unbounded reals).

Intended use

Inference-only prediction of a 9-channel cortical-surface signed-distance map from one 3D brain volume (T1 of any contrast, or CT clipped to [0, 80] HU), conformed to 1mm isotropic resolution and padded to multiples of 32 voxels. The consumer is recon-all-clinical, which uses channels 0-3 (LH-WM, LH-Pial, RH-WM, RH-Pial signed distances) to drive topologically-accurate cortical surface placement.

Usage

from ilex.models.synthdist import SynthDist
model = SynthDist.from_pretrained('ilex-hub/synthdist.1')

Authors

Gopinath K., Greve D. N., Das S., Arnold S., Magdamo C., Iglesias J. E.

Citation

Gopinath K., Greve D. N., Das S., Arnold S., Magdamo C., Iglesias J. E. (2023). Cortical analysis of heterogeneous clinical brain MRI scans for large-scale neuroimaging studies. arXiv:2305.01827.

References

• Gopinath K., Greve D. N., Das S., Arnold S., Magdamo C., Iglesias J. E. (2023). Cortical analysis of heterogeneous clinical brain MRI scans for large-scale neuroimaging studies. arXiv:2305.01827. (Original SynthDist publication; describes the distance-map prediction step that replaces iterative Bayesian surface fitting in the standard FreeSurfer pipeline.)
• Billot B., Greve D. N., Puonti O., Thielscher A., Van Leemput K., Fischl B., Dalca A. V., Iglesias J. E. (2023). SynthSeg: segmentation of brain MRI scans of any contrast and resolution without retraining. Medical Image Analysis, 83:102789. doi:10.1016/j.media.2022.102789. (Same SynthSeg-lineage UNet body is used by SynthDist; recon-all-clinical composes SynthSeg + SynthSR + SynthDist as a single stream.)

License

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

Effective terms: FreeSurfer Software License: academic / non-commercial use only. Pending explicit upstream confirmation that re-hosting on HF Hub is acceptable. Resolve before flipping published: true.

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

Copyright

Network architecture, training code, and pretrained weights: copyright (c) the upstream authors (Athinoula A. Martinos Center for Biomedical Imaging / Massachusetts General Hospital), distributed as part of the FreeSurfer software bundle under the FreeSurfer Software License (FSLA; permissive academic / non- commercial research offering). See https://surfer.nmr.mgh.harvard.edu/fswiki/FreeSurferSoftwareLicense for the binding terms. 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://surfer.nmr.mgh.harvard.edu/

Provenance

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