MR data processing and image reconstruction.

This project is a faster moving continuation/fork of MRpro. Some of the additional features of MRtwo will eventually get backported to MRpro, all new features of MRpro will be included in MRtwo.

In most cases, you can replace mrpro by mr2 in you code and everything works.

• Source code: https://github.com/fzimmermann89/mr2
• Documentation: https://fzimmermann89.github.io/mr2/
• Bug reports: https://github.com/fzimmermann89/mr2/issues
• Try it out:

Main features

• Data handling Custom dataclasses for fast data subsetting, sorting, rearranging
• Neural Network Block Common blocks and networks used for ML based MRI reconstruction
• ISMRMRD support MRtwo supports ismrmrd-format for MR raw data.
• PyTorch All data containers utilize PyTorch tensors to ensure easy integration in PyTorch-based network schemes.
• Cartesian and non-Cartesian trajectories MRtwo can reconstruct data obtained with Cartesian and non-Cartesian (e.g. radial, spiral...) sapling schemes. MRtwo automatically detects if FFT or nuFFT is required to reconstruct the k-space data.
• Pulseq support If the data acquisition was carried out using a pulseq-based sequence, the seq-file can be provided to MRtwo and the used trajectory is automatically calculated.
• Signal models A range of different MR signal models are implemented (e.g. T1 recovery, WASABI).
• Regularized image reconstruction Regularized image reconstruction algorithms including Wavelet-based compressed sensing or total variation regularized image reconstruction are available.

Examples

In the following, we show some code snippets to highlight the use of MRtwo. Each code snippet only shows the main steps. A complete working notebook can be found in the provided link.

Simple reconstruction

Read the data and trajectory and reconstruct an image by applying a density compensation function and then the adjoint of the Fourier operator and the adjoint of the coil sensitivity operator.

# Read the trajectory from the ISMRMRD file
trajectory = mr2.data.traj_calculators.KTrajectoryIsmrmrd()
# Load in the Data from the ISMRMRD file
kdata = mr2.data.KData.from_file(data_file.name, trajectory)
# Perform the reconstruction
reconstruction = mr2.algorithms.reconstruction.DirectReconstruction(kdata)
img = reconstruction(kdata)

Full example: https://github.com/fzimmermann89/mr2/blob/main/examples/scripts/direct_reconstruction.py

Estimate quantitative parameters

Quantitative parameter maps can be obtained by creating a functional to be minimized and calling a non-linear solver such as ADAM.

# Define signal model
model = MagnitudeOp() @ InversionRecovery(ti=idata_multi_ti.header.ti)
# Define loss function and combine with signal model
mse = MSE(idata_multi_ti.data.abs())
functional = mse @ model
[...]
# Run optimization
params_result = adam(functional, [m0_start, t1_start], n_iterations=n_iterations, learning_rate=learning_rate)

Full example: https://github.com/fzimmermann89/mr2/blob/main/examples/scripts/qmri_sg_challenge_2024_t1.py

Pulseq support

The trajectory can be calculated directly from a provided pulseq-file.

# Read raw data and calculate trajectory using KTrajectoryPulseq
kdata = KData.from_file(data_file.name, KTrajectoryPulseq(seq_path=seq_file.name))

Full example: https://github.com/fzimmermann89/mr2/blob/main/examples/scripts/comparison_trajectory_calculators.py

Development

pip install -e ".[dev]"

MRtwo requires PyTorch 2.3.1 or newer. CI currently validates the minimum supported CPU build, PyTorch 2.3.1, and the recommended CPU build, PyTorch 2.10.0.

Docker

A notebook-capable image for the recommended environment (Python 3.12 and PyTorch 2.10) is available from GHCR:

docker pull ghcr.io/fzimmermann89/mr2_py312:latest