NKI-AI/direct-vsharp-multianatomy

DIRECT — vSHARP multi-anatomy

Pretrained vSHARP models for multi-coil MRI reconstruction. Each release is an inference-ready pair:

vsharp_<name>.yaml   # inference-only DIRECT config
vsharp_<name>.pt     # weights

Paper: arXiv:2309.09954 · Framework: DIRECT

Datasets & provenance

Official portals (accept each dataset’s Data Sharing Agreement where required):

Model	Official source	Notes
vsharp_brain	fastMRI brain (multi-coil Cartesian)	Same portal as knee; original challenge data / papers via NYU FAIR fastMRI.
vsharp_knee	fastMRI knee (multi-coil Cartesian)	Fully sampled multi-coil knee MRI (Zbontar et al.).
vsharp_prostate	fastMRI prostate · Sci Data · code	T2 multi-coil. Raw volumes are (averages, slices, coils, readout, phase) with three averages at GRAPPA R=2 (odd/even line interleaving across averages). For these models, each average was GRAPPA-reconstructed, then the averages were combined into a single fully sampled multi-coil volume (slices, coils, readout, phase) using the official T2 GRAPPA pipeline (prostate_t2_recon.py).
vsharp_breast	fastMRI breast · Radiol AI · code	Native acquisition is radial GRASP DCE. For these models, radial k-space was regridded to Cartesian multi-coil volumes before training / inference with DIRECT.
vsharp_cardiac	CMRxRecon 2023 cine · Synapse syn51471091 · Sci Data	Multi-coil cine cardiac MRI from the 2023 challenge release.
vsharp_universal	fastMRI brain/knee/prostate/breast + CMRxRecon 2023 / 2024 / 2025	Same preprocessing as the anatomy-specific models above. CMRxRecon hub: cmrxrecon.github.io; Synapse portals for later challenges are linked from that site.

Training protocol

These models were not trained at a single fixed acceleration or scheme. All used the same mixed schedule:

	Values
Accelerations (R)	2, 4, 6, 8, 10
Center fractions (ACS)	0.16, 0.08, 0.06, 0.04, 0.02 (paired with (R): 2→0.16, 4→0.08, 6→0.06, 8→0.04, 10→0.02)
Sampling schemes	FastMRIEquispaced, FastMRIRandom, Gaussian1D, Gaussian2D, VariableDensityPoisson, Radial

vsharp_universal was trained jointly across the anatomies / challenges above; anatomy-specific models were trained on one dataset each.

Files & default inference masks

Released YAMLs are inference-only (no training / validation blocks). Each pins one acceleration and one ACS fraction — DIRECT’s mask sampler draws randomly from lists, so multi-(R) lists belong in training only.

File pair	Default mask	Default (R) / ACS
vsharp_brain.{yaml,pt}	FastMRIRandom	4× / 0.08
vsharp_knee.{yaml,pt}	FastMRIEquispaced	4× / 0.08
vsharp_prostate.{yaml,pt}	FastMRIEquispaced	4× / 0.08
vsharp_breast.{yaml,pt}	Radial	4× / 0.08
vsharp_cardiac.{yaml,pt}	FastMRIEquispaced	4× / 0.08
vsharp_universal.{yaml,pt}	FastMRIEquispaced	4× / 0.08

Changing acceleration or scheme

Edit inference.dataset.transforms.masking and keep both lists length 1:

inference:
  dataset:
    transforms:
      masking:
        name: FastMRIEquispaced   # any training scheme above
        accelerations: [8]        # single R
        center_fractions: [0.04]  # matching ACS

Target (R)	accelerations	center_fractions
2×	[2]	[0.16]
4×	[4]	[0.08]
6×	[6]	[0.06]
8×	[8]	[0.04]
10×	[10]	[0.02]

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-vsharp-multianatomy --local-dir ./vsharp_multianatomy

direct predict ./predictions \
  --cfg ./vsharp_multianatomy/vsharp_knee.yaml \
  --checkpoint ./vsharp_multianatomy/vsharp_knee.pt \
  --data-root /path/to/fastmri/knee/multicoil_val \
  --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

@article{yiasemis2024vsharp,
  title={vSHARP: variable Splitting Half-quadratic {ADMM} algorithm for Reconstruction of inverse Problems},
  author={Yiasemis, George and Moriakov, Nikita and S{\'a}nchez, Clara I. and Sonke, Jan-Jakob and Teuwen, Jonas},
  journal={Magnetic Resonance Materials in Physics, Biology and Medicine},
  year={2024},
  doi={10.1007/s10334-024-01189-0}
}