MLI-lab/cinemri

Implicit Neural Networks with Fourier-Feature Inputs for Free-breathing Cardiac MRI Reconstruction

This repository contains the code for reproducing figures and results in the paper ``Implicit Neural Networks with Fourier-Feature Inputs for Free-breathing Cardiac MRI Reconstruction''.

Citation

@article{kunz_implicit_2023,
    author    = {Johannes F. Kunz and Stefan Ruschke and Reinhard Heckel},
    title     = {Implicit Neural Networks with Fourier-Feature Inputs for Free-breathing Cardiac MRI Reconstruction},
    journal   = {	arXiv:2305.06822},
    year      = {2023}
}

Setup

1. Setup a docker container with support for Nvidia GPUs and pytorch.
2. Install additional packages

chmod +x ./setup/setup.sh
./setup/setup.sh

3. Download the datasets and copy them into the data folder in your project.
4. Configure and run the experiment scripts in the experiments/ folder.

Datasets

The low-resolution high-SNR, the low-resolution low-SNR, and the high-resolution dataset are available on IEEEDataPort, see https://dx.doi.org/10.21227/f057-dw29. The datasets need to be copied into the data folder of the project.

Licence

Reconstructed videos

Low-resolution high-SNR dataset

The video below shows the reconstructions of the low-resolution high-SNR dataset by the FMLP, the KFMLP, and the t-DIP for an acquisition time of $4s$ ($T = 225$). An ECG-gated breath-hold (BH) dataset was reconstructed using classical sparsity-based methods and is shown as visual reference. It can be seen that the reconstruction quality of the FMLP and the t-DIP are on par, whereas the KFMLP suffers aliasing-like artifacts.

Reconstructions by the FMLP using different acquisition lengths

For the video below, the FMLP was trained on three different acquisition lengths $4s$ ($T=225$), $8s$ ($T=512$), and $16s$ ($T=900$). It can be seen that the visual reconstruction quality does not improve notably with increased acquistion time.

Low-resolution low-SNR dataset

The low-resoltion low-SNR dataset was reconstructed by the FMLP, the KFMLP, and the t-DIP for an acquisition time of $4s$ ($T = 225$). It can be seen below that the reconstruction quality of the FMLP and the t-DIP is similar and both methods exhibit similar artifacts. The KFMLP, by contrast, exhibits aliasing-like artifacts and high-frequency noise.

High-resolution dataset

The reconstructions of high-resolution dataset for an acquisition time of $4s$ ($T=225$) are depricted below. Again, the FMLP and the t-DIP are similar in image quality. The KFMLP's reconstruction is obscured substantially by noise.

Licence

All files are provided under the terms of the Apache License, Version 2.0.