PyParMRI
Collection of magnetic resonance imaging (MRI) reconstruction tools written in Python with PyTorch.
Currently implemented methods:
-
GeneRalized Autocalibrating Partial Parallel Acquisition (GRAPPA) [1]. Implementation follows similar lines to the description provided in [2].
[1] Griswold, Mark A., et al. "Generalized autocalibrating partially parallel acquisitions (GRAPPA)." Magnetic Resonance in Medicine, 47.6 (2002): 1202-1210. [2] Uecker, Martin, et al. "ESPIRiT—an eigenvalue approach to autocalibrating parallel MRI: where SENSE meets GRAPPA." Magnetic resonance in medicine 71.3 (2014): 990-1001. -
Sum-Of-Squares (SOS) image - standard image combination techniques for parallel images (PI) MRI.
Notes about the implementations:
-
Implementation was written with PyTorch - should be differentiable, but this has not been tested thoroughly.
-
GRAPPA and SOS image reconstructions are GPU supported - this can be very useful when reconstructing from large k-space raw data files.
-
Two use code examples are provided:
a. main_tester.py - Can be run on GPU, or through multi-CPU processing (this feature is not fully tested).
b. main_tester_MPI.py - Can be run with MPI, but this is not fully tested.
-
Using the Python implementation of mapVBVD, data_io folder provides some simple scripts to read Siemens TWIX data files. However, the rsulting images do not contain the proper header.
-
GRAPPA recosntruction code supports multi-coil (multi-channel) 3D acquisitions (2D multi-slice acquisitions should also be supported, but this has not been tested yet). Data format should follow (4D tensor):
# Columns (Freq. encode), # Channels (Coils), # PE Lines, # Partitions.This should be automatically fullfiled if you read from a TWIX .dat file. GRAPPA kernels are 2D, but processing is batch performed for all frequency encode columns.
Requirements:
- PyTorch 1.11.0
- Numpy 1.22.1
- pyMapVBVD 0.4.8 - https://pypi.org/project/pyMapVBVD/
- mpi4py 3.1.3 - https://pypi.org/project/mpi4py/ (if you want to test multi-CPU processing)