Skip to content

Latest commit

 

History

History
118 lines (101 loc) · 4.18 KB

README.md

File metadata and controls

118 lines (101 loc) · 4.18 KB
Raw

pyLAR

pyLAR features Python implmentations of a low-rank atlas-to-image registration(LAR) framework and its applications in medical image analysis and computer vision. The core machine learning technique is Robust PCA.

Subdirectory content:

  • core --- two implementations of RPCA
  • examples -- a couple of ipython notebook examples of running RPCA
  • tests -- testing scripts of the core functionalities
  • low_rank_atlas -- the study of using RPCA in a low-rank atlas buildling framework
  • eval_utils -- the utilitiy scripts to evaluate the results on the low rank atlas building framework

pyLAR contains implementation of the following paper:

@article{Liu14,
    author = {X.~Liu and M.~Niethammer and R.~Kwitt and M.~McCormick and S.~Aylward},
    title = {Low-Rank to the Rescue – Atlas-based Analyses in the Presence of Pathologies},
    year = 2014,
    journal = {MICCAI},

The implementations of two recent proposals for robust PCA can be found in the "core" subdirectory:

@article{Candes11a,
    author = {E.J.~Cand\'es and X.~Li and Y.~Ma and J.~Wright},
    title = {Robust Principal Component Analysis?},
    year = 2011,
    volume = 58,
    number = 3,
    journal = {J. ACM},
    pages = {1-37}}

and

@article{Xu12a,
    author = {H.~Xu and C.~Caramanis and S.~Sanghavi},
    title = {Robust {PCA} via Outlier Pursuit},
    journal = {IEEE Trans. Inf. Theory},
    volume = 59,
    number = 5,
    pages = {3047-3064},
    year = 2012}

Please cite these articles in case you use the code in "core". Note that the original authors of those articles also provide MATLAB code. Further, the objectives of the two works are different: Candes et al.'s approach assumes randomly distributed corruptions throughout the dataset, while Xu et al.'s approach assumes that full observations (i.e., column vectors of the data matrix) and not just single entries are corrupted.

Requirements

Problem Statement(s)

See references (above) for the exact problem formulations of Candes et al. and Xu et al.

Example

An illustrative example for Candes et al.'s RPCA approach is to use a checkerboard image (provided under the examples directory) which is, by definition, low-rank and corrupt that image with randomly distributed outliers. The task is then to recover the low-rank part and thus obtain a clean version of the checkerboard image (as well as the sparsity pattern).

The examples directory contains an example (ex1.py) that demonstrates exactly this scenario. (Note: The example requires SimpleITK's python wrapping for image loading and image writing; it should be easy to replace these parts with your favorite image handling library, though).

Run the code with

python ex1.py checkerboard.png 0.3 /tmp/outlierImage.png /tmp/lowRank.png

Two images will be written: /tmp/outlierImage.png (i.e., the image with outliers) and /tmp/lowRank.png (i.e., the low-rank recovered part).

Using the IPython Notebook

We provide an IPython notebook, pyrpca-Tutorial.ipynb which can be found in the top-level directory of pyrpca. It basically walks a new user through the example implemented in ex1.py.

The following instructions were tested on a Linux machine running Ubuntu 12.04. We assume that you have virtualenv installed, e.g., using apt-get install python-virtualenv. Basically, we create a virtual environment, install all the required packages and eventually run the IPython notebook.

cd ~
mkdir tutorial-env
virtualenv ~/tutorial-env --no-site-packages
~/tutorial-env/bin/pip install ipython
~/tutorial-env/bin/pip install ipython[zmq]
~/tutorial-env/bin/pip install tornado
~/tutorial-env/bin/pip install numpy
~/tutorial-env/bin/pip install matplotlib
~/tutorial-env/bin/easy_install SimpleITK

Next, launch the IPython notebook:

~/tutorial-env/bin/ipython notebook --pylab=inline

To run the low-rank atlas-to-image registration application code, you need to install the [IPL package] (http://www.pythonware.com/products/pil/)