The accompanying codes provide an implementation in PyTorch of the methods in "Compressed sensing with a jackknife and a bootstrap" by Aaron Defazio, Mark Tygert, Rachel Ward, and Jure Zbontar.
The codes depend on Matplotlib, NumPy, Pillow, PyTorch, Scikit-image, and SciPy.
If using Anaconda of
Download Anaconda,
installation is possible via issuing the commands
conda install numpy;
conda install matplotlib;
conda install scipy;
conda install Pillow;
conda install scikit-image;
conda install pytorch torchvision -c pytorch
If using pip of Install pip,
installation is possible via issuing the commands
pip3 install numpy;
pip3 install matplotlib;
pip3 install scipy;
pip3 install Pillow;
pip3 install scikit-image;
pip3 install torch torchvision
For help, try
python bootstrap.py --help;
python bootstrap2.py --help;
python jackknife.py --help;
python jackknife2.py --help;
python admm.py --help;
python admm2.py --help;
Each includes a toy example for a unit test that runs when issuing the commands
python bootstrap.py;
python bootstrap2.py;
python jackknife.py;
python jackknife2.py;
python admm.py;
python admm2.py;
python radialines.py;
bootstrap.py and bootstrap2.py implement the bootstrap for estimating
errors in reconstructions from compressed sensing, the former for horizontal
sampling, and the latter for radial sampling.
jackknife.py and jackknife2.py implement the jackknife for estimating
errors in reconstructions from compressed sensing, the former for horizontal
sampling, and the latter for radial sampling.
admm.py and admm2.py perform compressed sensing via the alternating
direction method of multipliers for maximizing fidelity to Fourier-domain
measurements minus a total-variation penalty.
ctorch.py extends PyTorch to complex numbers.
radialines.py generates pixels on a Cartesian grid that intersect radial
lines emanating from the origin at specified angles.
Whereas bootstrap.py and jackknife.py should be useful for clinical
practice, bootstrap2.py and jackknife2.py are hacks full of code
duplication from bootstrap.py and jackknife.py, useful only for
prototyping. The "2" versions include admm2.py and radialines.py, which are
similarly useful only for experimentation. Radially retained sampling patterns
used in clinical practice do not fall on Cartesian grids. The "2" versions give
a proof of principle, nothing more.
brains.sh is a bash script that generates all figures in the paper available
at PDF
brains.sh processes the scans of brains in the subdirectory brains and
saves the results to a new subdirectory, brainsout.
Running brains.sh is required to process the LaTeX source in the subdirectory
tex; fbooja.tex depends on the results stored to the new subdirectory
brainsout.
License
This fbooja software is licensed under the LICENSE file (the MIT license) in the root directory of this source tree.