NKI-AI/direct-e2e-ads-recon

DIRECT — End-to-End Adaptive Sampling & Reconstruction

Pretrained models for joint learning of an adaptive (k)-space sampler and a reconstruction network on dynamic MRI (MIDL 2026).

• Paper (PMLR): End-to-End Co-Optimization of Adaptive k-space Sampling and Reconstruction for Dynamic MRI
• OpenReview: forum
• Related: arXiv:2403.10346
• Companion (with registration): NKI-AI/direct-e2e-ads-recon-reg

Dataset

Data	CMRxRecon multi-coil cardiac cine
Challenge / site	cmrxrecon.github.io
Task	Adaptive sampling + reconstruction under a fixed acceleration budget

Models

Each experiment is a .yaml + .pt pair (inference YAML active at 4×; other rates are commented under masking):

Name	Reconstruction	Sampler
vsharp_ads_1d	vSHARP	ADS 1D, unified mask
medl_ads_1d	MEDL	ADS 1D, unified
vsharp_ads_1d_frame	vSHARP	ADS 1D, frame-specific
medl_ads_1d_frame	MEDL	ADS 1D, frame-specific
vsharp_ads_1d_init2	vSHARP	ADS 1D unified + init
medl_ads_1d_init2	MEDL	ADS 1D unified + init
vsharp_ads_1d_frame_init2	vSHARP	ADS 1D frame-specific + init
medl_ads_1d_frame_init2	MEDL	ADS 1D frame-specific + init
vsharp_ads_2d	vSHARP	ADS 2D, unified
medl_ads_2d	MEDL	ADS 2D, unified
vsharp_ads_2d_frame	vSHARP	ADS 2D, frame-specific
medl_ads_2d_frame	MEDL	ADS 2D, frame-specific

Full training configs: projects/e2e_ads_recon.

Acceleration rates

Hub YAMLs pin one active accelerations / center_fractions pair (default val-4x). To run another trained rate, uncomment the matching block under masking (and comment out the active lists). For *init2*, also update target_acceleration.

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-e2e-ads-recon --local-dir ./e2e_ads_recon

direct predict ./predictions \
  --cfg ./e2e_ads_recon/vsharp_ads_1d.yaml \
  --checkpoint ./e2e_ads_recon/vsharp_ads_1d.pt \
  --data-root /path/to/cmrxrecon \
  --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{yiasemis2026e2eads,
  title={End-to-End Co-Optimization of Adaptive k-space Sampling and Reconstruction for Dynamic {MRI}},
  author={Yiasemis, George and Moriakov, Nikita and Sonke, Jan-Jakob and Teuwen, Jonas},
  booktitle={Medical Imaging with Deep Learning},
  year={2026}
}