SynDisCo -- MASILab fieldmap-free EPI distortion correction (b0 + BOLD) -- SynBOLD-DisCo -- fMRI BOLD distortion correction (5-fold)
Description
SynDisCo is a JAX / Equinox port of MASILab's two
distortion-correction nets -- Synb0-DisCo (dMRI b0) and
SynBOLD-DisCo (fMRI BOLD) -- consolidated as a single
multi-variant family class because the two upstream repos ship
byte-identical model.py architectures and differ only by
training corpus.
Each modality is a 5-fold ensemble of a 3D U-Net with 3
pooling stages, InstanceNorm3d + LeakyReLU(0.01) intra-block
norm/activation, ConvTranspose3d throughout the decoder (not
the more typical Conv+Upsample split), and bare 1x1
Conv3d/ConvTranspose3d at the bottleneck and head. Input is a
2-channel atlas-space (padded 80x96x80) volume stacking the
distorted target (b0 or BOLD) with a co-registered T1w
anatomical; output is a single-channel synthesised undistorted
target ready to drive FSL TopUp for susceptibility distortion
correction without reverse-PE acquisition or separate
fieldmap.
v0 ships two bundles (one per modality), each holding the
upstream's full 5-fold weights stacked on a leading fold axis.
The forward applies eqx.filter_vmap across the folds and means
the output -- byte-equivalent to the upstream pipeline's
fslmaths -Tmean ensemble step.
Intended use
fMRI sibling of the b0 variant. Synthesise an undistorted BOLD from T1w + the distorted (mean) BOLD for TopUp recovery; same atlas-space input format and ensemble policy as the b0 variant. Use when an fMRI acquisition lacks a reverse-PE / fieldmap calibration scan but a T1w is available.
Usage
1from ilex.models.syn_disco import SynDisCo
2model = SynDisCo.from_pretrained('ilex-hub/syn_disco.bold.1')
Authors
Schilling K. G., Blaber J., Huo Y., Newton A., Hansen C., Nath V., Yao T., Newlin N. R., Cai L. Y., Yang Q., Landman B. A., et al. (MASILab, Vanderbilt University; Synb0-DisCo + SynBOLD-DisCo upstreams)
Citation
Schilling K. G., Blaber J., Huo Y., Newton A., Hansen C., Nath V., et al. (2020). Synthesized b0 for diffusion distortion correction (Synb0-DisCo). PLoS One 15(7), e0236418. doi:10.1371/journal.pone.0236418. Yao T., Newlin N. R., Cai L. Y., Yang Q., Williams O., Beason-Held L. L., Resnick S. M., Schilling K. G., Landman B. A. (2023). SynBOLD-DisCo -- Synthetic BOLD images for distortion correction of fMRI without additional calibration scans. Imaging Neuroscience 1, 1-13.
References
- Schilling K.G., Blaber J., Huo Y., Newton A., Hansen C., Nath V., Shafer A.T., Williams O., Resnick S.M., Rogers B., Anderson A.W., Landman B.A. (2020). Synthesized b0 for diffusion distortion correction (Synb0-DisCo). PLoS One 15(7):e0236418. doi:10.1371/journal.pone.0236418.
- Yao T., Newlin N.R., Cai L.Y., Yang Q., Williams O., Beason- Held L.L., Resnick S.M., Schilling K.G., Landman B.A. (2023). SynBOLD-DisCo -- Synthetic BOLD images for distortion correction of fMRI without additional calibration scans. Imaging Neuroscience 1, 1-13.
- Synb0-DisCo upstream code, weights, license - github.com/MASILab/Synb0-DISCO (MIT).
- SynBOLD-DisCo upstream code, weights, license - github.com/MASILab/SynBOLD-DisCo (MIT).
License
HF Hub license tag: mit
Effective terms: MIT (copyright (c) MASILab, Vanderbilt University) on both the upstream code and the pretrained weights of Synb0-DisCo (github.com/MASILab/Synb0-DISCO) and SynBOLD-DisCo (github.com/MASILab/SynBOLD-DisCo). The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.
Copyright
Synb0-DisCo and SynBOLD-DisCo upstreams are copyright (c) MASILab,
Vanderbilt University, MIT-licensed. The JAX / Equinox port code
in this package is separately licensed under Apache-2.0 / GPL-3.0
per the wider ilex license boundary; the MIT terms on the upstream
weights are preserved through the canonical bundle's _ilex.origin
= 'pytorch' provenance.
Upstream source
Original weights / reference implementation:
https://github.com/MASILab/Synb0-DISCO
Provenance
This artefact was produced by
ilex's
save/load pipeline. The architecture is implemented in
ilex.models.syn_disco.SynDisCo and the weights have been converted
from their upstream format. See the upstream source above
for the canonical reference.