ilex-hub/gdc_net.3d-supervised.1

GDCNet -- calibrationless geometric distortion correction of EPI data -- GDCNet 3D supervised (unet arch)

Description

GDCNet (Jimeno et al., 2024; arXiv:2402.18777) is a deep-learning model that predicts a 1-channel voxel displacement map (VDM) along the phase-encoding (PE) direction of an EPI volume from a (source, target) pair. The VDM drives a downstream geometric distortion correction without requiring a separate fieldmap or reverse-PE calibration acquisition.

Architecture: a vxm-family 3D U-Net (encoder filters 16/32/32/32, decoder filters 32/32/32/32, tail convs 32/16/16; InstanceNorm-free; LeakyReLU(0.2) intra-block) consuming a 2-channel (source, target) input at the published 64x64x32 grid. The U-Net body produces a 16-channel feature map; a final Conv3d with kernel (1, 1, 3) collapses the 16-channel features into a 1-channel VDM along the PE (last spatial) axis.

v0 ships three 3D variants (supervised + semi-supervised + self-supervised). The architecture is parameter-identical across variants (~326k trainable scalars); only the optional spatial-transform step and the training corpus differ.

Intended use

Supervised arch -- forward returns the predicted VDM only. Use when downstream pipelines accept a raw displacement field as input to TopUp / applywarp. Input grid: (1, 64, 64, 32). Output: (1, 64, 64, 32) VDM along the channel-0 (PE) direction of a (vdm, 0, 0) flow.

Usage

from ilex.models.gdc_net import GDCNet
model = GDCNet.from_pretrained('ilex-hub/gdc_net.3d-supervised.1')

Authors

Manso Jimeno M., Vaughan J. T., Geethanath S. (Columbia University, Department of Biomedical Engineering)

Citation

Manso Jimeno M., Vaughan J. T., Geethanath S. (2024). GDCNet -- Calibrationless geometric distortion correction of echo planar imaging data using deep learning. arXiv:2402.18777. Zenodo deposit: doi:10.5281/zenodo.10257231.

References

• Manso Jimeno M., Vaughan J. T., Geethanath S. (2024). GDCNet -- Calibrationless geometric distortion correction of echo planar imaging data using deep learning. arXiv:2402.18777.
• Manso Jimeno M. (2023). GDCNet -- Calibrationless geometric distortion correction of echo planar imaging data using deep learning. Zenodo deposit. doi 10.5281/zenodo.10257231.
• Upstream code (no LICENSE file in repo) -- github.com/imr-framework/gdcnet.

License

HF Hub license tag: cc-by-4.0

Effective terms: CC-BY-4.0 (Creative Commons Attribution 4.0). The trained .h5 weights are deposited on Zenodo (zenodo.org/records/10257231) under CC-BY-4.0. The upstream code (github.com/imr-framework/gdcnet) has no LICENSE file; the ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0. Redistribution requires attribution to Manso Jimeno, Vaughan, and Geethanath (2024).

Upstream license reference: https://creativecommons.org/licenses/by/4.0/

Copyright

Network architecture (github.com/imr-framework/gdcnet) is unlicensed in the upstream repository -- the trained .h5 weights deposited on Zenodo (zenodo.org/records/10257231, DOI 10.5281/zenodo.10257231) are CC-BY-4.0 (Creative Commons Attribution 4.0). The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0; the CC-BY terms on the upstream weights are preserved through the canonical bundle's _ilex.origin = 'tf' provenance.

Upstream source

Original weights / reference implementation: https://github.com/imr-framework/gdcnet

Provenance

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