NKI-AI/direct-uniform

DIRECT — UNIFORM multi-anatomy vSHARP

UNIFORM (MIDL 2025) is a unified deep learning framework for reconstructing undersampled multi-coil MRI across diverse anatomical sites and contrasts, built on vSHARP inside the DIRECT toolkit.

📄 Paper: UNIFORM: A Unified Deep Learning Framework for Multi-organ and Multi-contrast MRI Reconstruction · PDF
🏗️ Method: vSHARP (MRI, 2025) · arXiv:2309.09954
💻 Code: projects/UNIFORM

Figure 1 (MIDL 2025): one vSHARP model trained on fastMRI brain / knee / prostate and CMRxRecon cardiac data; evaluated at 2×, 4×, 6×, and 8× acceleration; zero-shot SSL on breast in the paper.

What is in this repo?

File	Role
uniform_vsharp.pt	Pretrained weights — use with the YAMLs below
uniform_brain.yaml	Brain inference (default 4× FastMRIRandom, ACS 0.08)
uniform_knee.yaml	Knee inference (default 4× FastMRIEquispaced, ACS 0.08)
uniform_prostate.yaml	Prostate inference (default 4× FastMRIEquispaced, ACS 0.08)
uniform_cardiac.yaml	Cardiac / CMRxRecon inference (default 4× FastMRIEquispaced, ACS 0.08)

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

pip install huggingface_hub
hf download NKI-AI/direct-uniform --local-dir ./uniform

direct predict ./predictions/brain \
  --cfg ./uniform/uniform_brain.yaml \
  --checkpoint ./uniform/uniform_vsharp.pt \
  --data-root /path/to/fastmri/brain/multicoil_val \
  --filenames-filter projects/UNIFORM/lists/test/brain_4x.lst \
  --num-gpus 1

The first argument to direct predict is the prediction output directory. Pass basenames via --filenames-filter (path to a .lst file under --data-root); unlike training/validation, inference does not read filenames_lists from the YAML.

Changing acceleration

Edit inference.dataset.transforms.masking and uncomment one pair — keep both lists length 1 (DIRECT samples randomly from lists; multi-(R) lists are for training only):

masking:
  name: FastMRIEquispaced   # brain YAML defaults to FastMRIRandom
  # accelerations: [8]
  # center_fractions: [0.04]
  accelerations: [4]
  center_fractions: [0.08]

Target (R)	accelerations	center_fractions
2×	[2]	[0.1]
4×	[4]	[0.08]
6×	[6]	[0.06]
8×	[8]	[0.04]

Datasets

Anatomy	Source	Contrasts (paper)
Brain	fastMRI multi-coil	T1w, T2w, FLAIR
Knee	fastMRI multi-coil	PD with & without fat suppression
Prostate	fastMRI prostate	T2w
Cardiac	CMRxRecon 2023	Cine, T1w, T2w (use ValidationSet/FullSample; flatten to P0XX_cine_*.mat)

Citation

If you use this model, please cite UNIFORM, vSHARP, and the DIRECT toolkit.

@inproceedings{Yiasemis_UNIFORM,
  title     = {{UNIFORM}: A Unified Deep Learning Framework for Multi-organ and Multi-contrast {MRI} Reconstruction},
  author    = {Yiasemis, George and Ferm, Jonatan and Moriakov, Nikita and Mann, Ritse M. and Sonke, Jan-Jakob and Teuwen, Jonas},
  booktitle = {Medical Imaging with Deep Learning},
  year      = {2025},
  url       = {https://openreview.net/forum?id=I13Y1nU6gs}
}

@article{Yiasemis_2025_vSHARP,
  title   = {vSHARP: Variable Splitting Half-quadratic ADMM algorithm for reconstruction of inverse-problems},
  author  = {Yiasemis, George and Moriakov, Nikita and Sonke, Jan-Jakob and Teuwen, Jonas},
  journal = {Magnetic Resonance Imaging},
  volume  = {115},
  pages   = {110266},
  year    = {2025},
  doi     = {10.1016/j.mri.2024.110266}
}

@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}
}