mikgroup/extreme_mri

Code for Extreme MRI

Extreme MRI

This repo contains scripts to reproduce some experiments in Extreme MRI. It uses the Python package sigpy.

Colab Demos

To start, we recommend trying the demo notebooks on Google Colab:

• DCE:

• Lung:

Data

Example DCE and lung datasets can be found on Zenodo:

• DCE:

• Lung:

For more description about how the data was acquired, please see the manusciprt. For description about the variable names, please see the Variables section below.

Command line scripts

Installation

Install sigpy using pip:

pip install sigpy

GPU is recommended for running extreme MRI reconstruction. To do so, you will need to install cupy, a numpy-like package for CUDA, either through conda or pip:

pip install cupy

Variables

• ksp: kspace data array of shape [# of channels, # of TRs, readout lengths].
• coord: kspace coordinate array of shape [# of TRs, readout lengths, # of dimensions].
• dcf: density compensation factor of shape [# of TRs, readout lengths].
• mps: sensitivity maps of shape [# of channels, nx, ny, nz]
• img: reconstructed image of shape [# of frames, nx, ny, nz]

Processing pipeline

To run the reconstruction, the general pipeline is to:

• run setup script to the folder containing numpy arrays (ksp.npy, coord.npy, dcf.npy).
• automatically select FOV to account for leakage from slab selection.
• perform gridding reconstruction to look at image.
• estimate sensitivity maps using JSENSE.
• running the low rank reconstruction.

Example Usages

Auto FOV and estimate sensitivity maps

source setup.sh path/to/folder/
python autofov.py $ksp $coord $dcf --device 0
python jsense_recon.py $ksp $coord $dcf $mps --device 0

Estimate respiratory signal and soft-gating weights with TR of 7.7 ms

python estimate_resp.py $ksp 0.0077 $resp
python soft_gating_weights.py $resp $sgw

Low rank reconstruction with 500 frames

python multi_scale_low_rank_recon.py $ksp $coord $dcf $mps 500 $img --device 0

Low rank reconstruction with 20 frames and soft-gating weights

python multi_scale_low_rank_recon.py $ksp $coord $dcf $mps 20 $img --device 0 --sgw_file $sgw

Motion resolved reconstruction with 5 bin

python motion_resolved_recon.py $ksp $coord $dcf $mps $resp 5 $mrimg --device 0

extreme_mri's Issues

Is this code able to reproduce the amazing images in your paper?

Hi! Thanks a lot for sharing your code, and congratulation for the amazing work described in your recent paper!

I was looking at the colab notebook, and it seems to me that there are some simplifications there with the final results not really approaching what we can see in the paper. Is there an "easy" way to change the notebook to reproduce your result for the pedriatic dce-mri dataset, or what you shared is just a simpler example?

Thanks!