Moner

This is the official implementation of our ICLR 2025 paper, "Moner: Motion Correction in Undersampled Radial MRI with Unsupervised Neural Representation" [OpenReview].

Fig. 1: Overview of the proposed Moner model.

Visualization

Fig. 2: Qualitative results of NuIFFT and Moner on 5 representative samples of the fastMRI dataset at AF=2 and MR=[-5, +5].

File Tree

Moner
│  config.yaml                  # configuration file.
│  dataset.py                   # dataloader
│  main.py                      # running script for training
│  model.py                     # motion correction module
│  readme.md                    # readme file
│  results.ipynb                # results report
│  train.py                     # training function
│  utils.py                     # tools
│
├─data
│      gt.nii                   # gt MR images
│      kdata_360_5.nii          # undersampled, motion-corrupted k-data
│      nuifft_360_5.nii         # NuIFFT MRI reconstructions
│      rot_360_5.txt            # gt motion parameter (rotation angle)
│      shift_x_360_5.txt        # gt motion parameter (X-axis shift)
│      shift_y_360_5.txt        # gt motion parameter (Y-axis shift)
│
├─fig                           # visualization
│      fig_method.jpg
│      fig_results.gif
│
├─model
│      model_360_5_0~4.pkl      # Moner checkpoints
│
└─out
        recon_360_5.nii         # Moner MRI reconstructions
        recon_reg_360_5.nii     # Moner MRI reconstructions after applying rigid registruction
        rot_360_5.txt           # estimated motion parameter (rotation angle)
        shift_x_360_5.txt       # estimated motion parameter (X-axis shift)
        shift_y_360_5.txt       # estimated motion parameter (Y-axis shift)

Main Requirements

To run this project, you will need the following packages:

• PyTorch
• tinycudann
• torchkbnufft
• SimpleITK
• tqdm
• numpy
• other dependencies

Training and Checkpoints

To train Moner from scratch, navigate to the project directory and run the following command in your terminal:

python main.py

This will train the Moner model on rigid motion-corrupted, undersampled k-space data from the fastMRI dataset at AF=2x and MR=[-5, +5] (./data/kdata_360_5.nii). The file kdata_360_5.nii is generated from the raw k-space data using the 1D IFFT operator, implemented in the function ifft1d in the utils.py file. The well-trained model checkpoints will be stored in ./model, and the corresponding MR images and motion parameters will be saved in ./out.

Evaluation

To evaluate the results, run the results.ipynb notebook, which provides the quantitative results of the reconstructed MR images and the estimated motion parameters.

For the 5 projections (./data/kdata_360_5.nii), the quantitative results of the MR images are:

Method	PSNR	SSIM
NuIFFT	21.78	0.328
Moner	32.53	0.935

Others

NIFTI files (.nii) can be viewed using the ITK-SNAP software, available for free download at: http://www.itksnap.org/pmwiki/pmwiki.php?n=Downloads.SNAP4

Citation

If you find our work useful in your research, please cite:

@inproceedings{
wu2025moner,
title={Moner: Motion Correction in Undersampled Radial {MRI} with Unsupervised Neural Representation},
author={Qing Wu and Chenhe Du and Xuanyu Tian and Jingyi Yu and Yuyao Zhang and Hongjiang Wei},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=OdnqG1fYpo}
}