NKI-AI/direct-cvpr2022-recurrentvarnet

DIRECT — CVPR 2022 RecurrentVarNet

Checkpoints from the DIRECT release related to Recurrent Variational Network (Yiasemis et al., CVPR 2022).

• Paper: arXiv:2111.09639
• Project: projects/cvpr2022_recurrentvarnet
• Previous hosting: files.aiforoncology.nl/direct-project/cvpr_recurrentvarnet

Datasets

Use in paper	Dataset	Link
Main experiments & ablations	Calgary-Campinas brain (multi-coil)	sites.google.com/view/calgary-campinas-dataset
Challenge masks	Poisson-disk undersampling masks	Challenge site · Hub NKI-AI/direct-mri-masks
Additional experiments (Appendix)	fastMRI brain AXT1	fastmri.med.nyu.edu · fastmri.org/dataset

Calgary-Campinas split used in the paper: 47 / 10 / 10 volumes (train / val / test) after discarding outer slices. Subsampling uses the challenge Poisson-disk masks. fastMRI AXT1 experiments use random Cartesian undersampling.

Models

<name>.yaml
<name>.pt

Calgary-Campinas (default names without prefix): RecurrentVarNet (shared weights and ablations *_noRSI, *_noSER, *_T11) plus comparison baselines (rim, xpdnet, unet, …), with a single pinned acceleration under inference.dataset.transforms.masking (and crop_outer_slices: true).

fastMRI AXT1 packs use the fastmri_axt1_ prefix (see below).

fastMRI AXT1 brain (appendix / rebuttal)

Additional CVPR 2022 experiments trained on fastMRI brain AXT1:

fastmri_axt1_<model>.yaml
fastmri_axt1_<model>.pt

Models: recurrentvarnet, recurrentvarnet_ablation, lpd, rim, unet, varnet.

Inference YAMLs default to 4× random Cartesian undersampling (center_fractions: [0.08]). Commented 8× settings (accelerations: [8], center_fractions: [0.04]) are under inference.dataset.transforms.masking. Keep exactly one active acceleration list.

direct predict ./predictions \
  --cfg ./cvpr_rvn/fastmri_axt1_recurrentvarnet.yaml \
  --checkpoint ./cvpr_rvn/fastmri_axt1_recurrentvarnet.pt \
  --data-root /path/to/fastmri/brain/multicoil_val \
  --num-gpus 1

Install DIRECT

git clone https://github.com/NKI-AI/direct.git
cd direct
conda create --name direct python=3.12
conda activate direct
pip install meson-python meson ninja
pip install --no-build-isolation -e ".[dev]"

Usage

hf download NKI-AI/direct-cvpr2022-recurrentvarnet --local-dir ./cvpr_rvn

direct predict ./predictions \
  --cfg ./cvpr_rvn/recurrentvarnet_shared_weights.yaml \
  --checkpoint ./cvpr_rvn/recurrentvarnet_shared_weights.pt \
  --data-root /path/to/calgary_campinas \
  --num-gpus 1

The first argument is the prediction output directory.

Citation

If you use these models or DIRECT, please cite the DIRECT toolkit and the method paper(s) below.

DIRECT

@article{DIRECTTOOLKIT,
  title={DIRECT: Deep Image REConstruction Toolkit},
  author={Yiasemis, George and Moriakov, Nikita and Karkalousos, Dimitrios and Caan, Matthan and Teuwen, Jonas},
  journal={Journal of Open Source Software},
  volume={7},
  number={73},
  pages={4278},
  year={2022},
  doi={10.21105/joss.04278},
  url={https://doi.org/10.21105/joss.04278}
}

Method

@inproceedings{yiasemis2022recurrentvarnet,
  title={Recurrent Variational Network: A Deep Learning Inverse Problem Solver applied to the task of Accelerated {MRI} Reconstruction},
  author={Yiasemis, George and Sonke, Jan-Jakob and S{\'a}nchez, Clara I. and Teuwen, Jonas},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2022}
}