DIRECT: Deep Image REConstruction Toolkit

Installation · Quick start · Documentation · Model zoo · Papers

DIRECT is a PyTorch toolkit for accelerated MRI reconstruction. It takes undersampled multi-coil k-space through sampling, reconstruction, optional registration, metrics, and pretrained baselines — end to end.

Challenge-winning models shipped in DIRECT include vSHARP (CMRxRecon 2023; also used in the 2024 challenge), RecurrentVarNet (Calgary-Campinas / MIDL 2020), and RIM (fastMRI 2019).

Zero-filled reconstruction, Compressed-Sensing (CS) reconstruction using the BART toolbox, Reconstruction using a RIM model trained with DIRECT

Features

• MRI data and sampling. Multi-coil static, dynamic, and multislice volumes; coil-sensitivity estimation; and a library of Cartesian, radial, spiral, Poisson, Gaussian, and k-t masks. A learned Adaptive Dynamic Sampler (ADS) can also choose lines or pixels under a fixed acceleration budget.
• Reconstruction models. vSHARP, RecurrentVarNet, VarNet, RIM / CIRIM, LPDNet, XPDNet, IterDualNet, ConjGradNet, Joint-ICNet, KIKI-Net, MultiDomainNet, VarSplitNet, U-Net (2D / 3D), MEDL, and transformer reconstructors (ViT, UFormer) in image or k-space.
• Training paradigms. Fully supervised learning, self-supervised SSDU, and JSSL (joint supervised + self-supervised). Distributed multi-GPU training, mixed precision, and TensorBoard logging.
• Conditional and joint pipelines. Modulated convolutions condition an unrolled network on acceleration and ACS fraction. Optional registration (learned or classical) aligns dynamic frames with reconstruction.
• Validation and reuse. MRI metrics (SSIM, pSNR, NMSE, VIF, HFEN, …), YAML configs, direct train / direct predict, and a model zoo on Hugging Face.

Install

PyPI package name is direct-recon (import as direct):

pip install direct-recon

Development install with uv:

git clone https://github.com/NKI-AI/direct.git
cd direct
uv sync

See the installation guide for Docker and conda.

Projects and model zoo

Reproducible experiment configs live under projects/. Pretrained .yaml / .pt pairs are on Hugging Face (NKI-AI) and listed in the model zoo.

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

direct predict ./predictions \
    --cfg ./calgary/rim_5x.yaml \
    --checkpoint ./calgary/rim_5x.pt \
    --data-root /path/to/calgary_campinas \
    --num-gpus 1

License

DIRECT is not intended for clinical use. It is released under the Apache 2.0 License.

Citing DIRECT

If you use DIRECT, please cite the toolkit paper. Method-specific BibTeX entries are collected on the papers page.

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