Motion Correction in MRI

Motion correction for MRI reconstruction using joint optimization techniques from the papers titled "Sensitivity Encoding for Aligned Multishot Magnetic Resonance Reconstruction" and DISORDER sampling techniques from a subsequent paper written by the same authors titled "Motion-corrected MRI with DISORDER: Distributed and incoherent sample orders for reconstruction deblurring using encoding redundancy".

Environment Installation

Use a basic python venv with the requirements.txt file

Directory Structure

The directory is just a flat collection of files that have various functions that are used in the two main scripts namely simulation.py which handles any experimental simulations, and recon.py that will take a dat file as input and run the reconstruction end to end.

Preprocessing

• extract_from_dat.py: extracts kspace, ref scan and noise raw data from a .dat file
• coil_compression.py: compresses sensitivity maps and kspace data by reducing number of coils
• espirit.py: implementation of Espirit algorithm for large in memory datasets, used to estimate sensitivity maps
• noise_whiten.py: applies whitening matrix to kspace using noise data aquired from raw dat file data

Core Algorithm

• image_estimation.py: this script contains the GPU accelerated image reconstruction portion of the algorithm, the function named estimate_image_cg everything else is legacy implementations. This function assumes a gpu and has not been tested yet for CPU usage.
• transform_estimation.py: this script contains the motion estimate step of the algorithm contained in the function call estimate_transform, the rest is all legacy code that should not be used.
• transform.py: this script contains the linop classes for applying the rigid transforms. The two in question are called RigidTransformCudaOptimzied and RigidTransformDerivativeCuda the rest is legacy implementations that should not be used.
• joint_recon.py: this script contains the JointRecon class that is the main algorithm and alternated between image estimation and motion estimate returning the fully reconstructed image with estiamted motion parameters.
• pyramid.py: this is an orchestration script which handles downsampling the input to a correct size and running a joint recon and continuing the process till we are back at full resolution.

Miscellaneous

• utils.py: self explanatory file that conatins small utility functions

Usage

Mentioned in the previous section the two main scripts that you will be using are recon.py and simulation.py and they can be simply run via commaline with some agruments supplied. Using the --help functionality will list what possibly inputs are allowed.

The standard way 3D and 4D image data are represented is the following:

• kspace: [ncoils, kx, ky, kz] complex64
• mps: [ncoils, kx, ky, kz] complex64
• ground/image: [x, y, z] complex64
• sampling mask: [nshots, x, y, z] bool
• transforms: [nshots, 6] float32

Another common naming convention used throughout is when a reconstruction is labeled as sense this refers to a simple sense reconstruction wihout motino correction. Joint Recon indicates a reconstruction that estimates motion and automatically corrects the image for it.

python recon.py \
    --dat <path to .dat file> \
    --out <path to output dir\
    --nshots 64 \
    --sampling disorder\
    --iters 100

python simulation.py \
    --ground <path to image file in .npy format>\
    --mps <path to sensitivity maps file in .npy format>\
    --out_dir <path to output dir>\
    --ord disorder\
    --tile_size 16 16\
    --accel 2 \
    --iters 1000\
    --continuous