BLIPSrecon

Code for reproducing results in the BLIPS paper:

Anish Lahiri, Guanhua Wang, Sai Ravishankar, Jeffrey A. Fessler, (2021). "Blind Primed Supervised (BLIPS) Learning for MR Image Reconstruction." IEEE Transactions on Medical Imaging http://doi.org/10.1109/TMI.2021.3093770; arXiv preprint arXiv:2104.05028.

The code is made up of two components:

• Blind dictionary learning (MATLAB version 2020+)
• Supervised learning (with PyTorch > 1.7.0).
• Additionally, we used BART to generate the dataset.

Blind Dictionary Learning-based Image Reconstruction

The MATLAB code performs without issues on MATLAB 2020+.

Run batchSOUP_DLMRI_randmask.m to reconstruct images from an input directory.

The input directory should consist of .mat files for individual slices, and include I1 (ground truth), S (sensitivity map) and Q1 (sampling masks). The option to save the dictionary learning output to a directory is available in the script upon uncommenting line 53. The code saves IOut (dictionary learning reconstruction), y (k-space), I1 (ground truth), S (sensitivity map) and Q1 (sampling masks), along with paramsout for performace metrics as needed.

subbatchMRIreconstruction_multicoil.m specifies the parameter choices (such as number of inner and outer iterations, sparsity penalty weight, etc.) for dictionary learning reconstruction from multicoil MR measurements.

dictionarylearningMRIreconstruction_multicoil.m generates a reconstruction from undersampled multi-coil k-space measurements using SOUP-DIL dictionary learning-based regularization.

SOUPDIL.m is the inner dictionary learning and sparse coding function that takes overlapping patches in an initial image and learns a dictionary and a set of sparse coefficients to represent these patches.

Update

Now we provide SOUP-DIL on Python, via MIRTorch. This Jupyter notebook illustrates the usage.

Deep Supervised Learning-based Image Reconstruction

The supervised learning approach used a training set generated from the fastMRI project.

• Preprocessing.ipynb provides an example of the data-preprocessing.
• MODL_DLMRI_Knee_vd.sh gives an example of training the neural network.
• The file requirements.txt denotes related Python packages.