DIRECT: Deep Image REConstruction Toolkit

Installation • Quick Start • Documentation • Model Zoo

DIRECT is a Python, end-to-end pipeline for solving Inverse Problems emerging in Imaging Processing. It is built with PyTorch and stores state-of-the-art Deep Learning imaging inverse problem solvers such as denoising, dealiasing and reconstruction. By defining a base forward linear or non-linear operator, DIRECT can be used for training models for recovering images such as MRIs from partially observed or noisy input data. DIRECT stores inverse problem solvers such as the vSHARP, Learned Primal Dual algorithm, Recurrent Inference Machine and Recurrent Variational Network, which were part of the winning solutions in Facebook & NYUs FastMRI challenge in 2019, the Calgary-Campinas MRI reconstruction challenge at MIDL 2020 and the CMRxRecon challenge 2023. For a full list of the baselines currently implemented in DIRECT see here.

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

Projects

In the projects folder baseline model configurations are provided for each project.

Baselines and trained models

We provide a set of baseline results and trained models in the DIRECT Model Zoo. Baselines and trained models include the vSHARP, Recurrent Variational Network (RecurrentVarNet), the Recurrent Inference Machine (RIM), the End-to-end Variational Network (VarNet), the Learned Primal Dual Network (LDPNet), the X-Primal Dual Network (XPDNet), the KIKI-Net, the U-Net, the Joint-ICNet, and the AIRS Medical fastmri model (MultiDomainNet).

License and usage

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

Citing DIRECT

If you use DIRECT in your own research, or want to refer to baseline results published in the DIRECT Model Zoo, please use the following BiBTeX entry:

@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}
}