FD-Net -- unsupervised forward-distortion correction of reversed-PE EPI pairs -- FD-Net 42-subjects checkpoint

Description

FD-Net (Alkilani et al., Magn Reson Med 2023) is an unsupervised deep-learning model for EPI susceptibility distortion correction. Given a reversed phase-encoding pair of EPI slices (LR, RL), the model jointly predicts:

  • a corrected image (the underlying anatomically-correct slice),
  • a displacement field along the PE direction.

The forward-distortion physics model is baked into the architecture as a K-matrix sinc-interpolation operator (K_UNIT); during training, the model is supervised by the reconstruction loss between the K_UNIT-forward-distorted outputs and the input pair, without requiring ground-truth fieldmaps. Plus a 5-layer rigid-alignment regressor (LocNet)

  • bicubic Hermite STN that does rigid alignment of the +PE output against the input.

v0 ships one variant -- the 42-subjects training checkpoint (the upstream's primary release). The 4-/8-/12-subject ablation checkpoints + the fine-tuned-on-fMRI checkpoint will ship in follow-up versions.

Intended use

Reversed-PE EPI distortion correction trained on the 42-subjects HCP-derived corpus (the upstream's primary release). Input is a (LR, RL) pair of (1, 144, 168) slices intensity-normalised to [0, 1]; output is a (4, 144, 168) stack of (output_LR, output_RL, output_field, output_image) per the upstream's full_scale head. Multi-blur training outputs and LocNet rigid params are exposed via FDNet.multi_blur / FDNet.rigid_params.

Usage

from ilex.models.fd_net import FDNet
model = FDNet.from_pretrained('ilex-hub/fd_net.42-subjects.1')

Authors

Zaid Alkilani A., Cukur T., Saritas E. U. (ICON Lab, Bilkent University, Ankara, Turkey)

Citation

Zaid Alkilani A., Cukur T., Saritas E. U. (2023). FD-Net -- An unsupervised deep forward-distortion model for susceptibility artifact correction in EPI. Magnetic Resonance in Medicine, 1-17. doi:10.1002/mrm.29851.

References

  • Zaid Alkilani A., Cukur T., Saritas E. U. (2023). FD-Net -- An unsupervised deep forward-distortion model for susceptibility artifact correction in EPI. Magnetic Resonance in Medicine, 1-17. doi 10.1002/mrm.29851.
  • Zaid Alkilani A., Cukur T., Saritas E. U. (2022). A Deep Forward-Distortion Model for Unsupervised Correction of Susceptibility Artifacts in EPI. Proceedings of the 30th Annual Meeting of ISMRM, London, paper 0959.
  • Upstream code + weights -- github.com/icon-lab/FD-Net (MIT).

License

HF Hub license tag: mit

Effective terms: MIT (Massachusetts Institute of Technology License). The upstream code + TF Checkpoint weights are released under MIT at github.com/icon-lab/FD-Net. The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.

Upstream license reference: https://opensource.org/licenses/MIT

Copyright

FD-Net is copyright (c) ICON Lab 2023, MIT-licensed on both the network code (github.com/icon-lab/FD-Net) and the released TF Checkpoint weights (under network/weights). The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.

Upstream source

Original weights / reference implementation: https://github.com/icon-lab/FD-Net

Provenance

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

Downloads last month
5
Safetensors
Model size
1.74M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support