Bayesian Non-Parametric Image Segmentation

This project implements a Gibbs Sampler for performing pixel-wise classification (image-segmentation) of vector-valued voxel representations in a joint Hierarchical Dirichlet Process-Markov Random Field bayesian non-parametric model.

The model was developed for learned intra-tumor segmentation. The model places each voxel across a population of patients' images into one of a variable number of classes. Utilization of the HDP guarantees correspondence of a class definition for all patients, allowing the algorithm to select the appropriate number of distinct voxel representations based on population information.

The MRF constraint is placed on the class assignment to reinforce the assumption that under a revealing representation (feature vector) voxels (collections of cells) tend to compose themselves into larger clusters rather than spread them selves into small, disjoint clusters.

The algorithm draws samples from conditional distributions over the full state space vector to infer the latent class assignment parameters and implicit number of unique cell classifications.

Getting Started

Running the code for the first time is easy, just follow these steps after you have installed all of the dependencies

1. Clone the repo:

   git clone https://github.com/ryanneph/bnp-tumorseg

2. Enter the project directory and run on one of the sample datasets:

   python run.py --dataset=balloons

• Check out all of the available options with:

  python run.py --help

On every iteration, the t-index and k-index mode maps are saved to ./figures/<dataset>/, and logfiles are saved to ./logs/. Presently, the full trace histories for t-index, k-index, and beta random variables are stored to file in ./blobs/<dataset> at the end of the sampling stage.

Resuming a Previous Sampling Run

Basic support for resuming a sampling session from a previously terminated session is available by specifying the cli argument: --resume-from=<blob.pickle> using one of the recorded data "checkpoints" which are written to ./blobs/<dataset> a the end of each session (whether maxiter is reached or the user terminates early with <ctrl-c>) Currently, by resuming a sampling session, one can modify the concentration parameter values, the MRF constraint weights, and the burnin period. The data used for the resuming session will be identical to that used in the checkpoint blob. (no new data added to the dataset directory will be registered at this time).

Dependencies

• NumPy
• SciPy
• matplotlib
• pymedimage (install using pip install --upgrade git+git://github.com/ryanneph/pymedimage.git)
• choldate
• pushbullet.py [optional] - for getting success/failure notifications on your mobile devices (see notifications below for more details - coming soon)