Task modality in fMRI code

This repo contains the source code for the calculations done in the forthcoming paper Task modality in fMRI is detected by persistence.

The structure is as follows:

• data contains three subdirectories for getting the fMRI data into the correct format, sorted further by subjectID.
  • raw contains the raw matlab files.
  • preprocessed contains the masked csv files.
  • postprocessed contains the output of the persistent homology calculations from perseus.
• src contains the main scripts for the computation.
  • make_dataset.py contains the data wrangling aspects of the project, with methods for converting matlab files to masked perseus input files.
  • landscapes.py creates and manipulates landscapes for machine learning algorithms.
  • permutation_test.py contains a labelled permutation test.
  • svm.py contains an sklearn Linear SVM.
  • config.py contains global variables, like the list of modality labels for the experiment.
• main.py contains the main scripts used for running the pipeline.

Workflow

The main workflow of the code is as follows:

1. Pre-process the data: convert matlab files to plain text files, apply the mask, and convert into perseus input files.
2. Run perseus on the input files.
3. Convert the perseus output into a numpy.ndarray to be fed into persim.
4. Use persim to compute the Persistence Landscape of each time slice.
5. Label and process (pad) the landscapes for the statistical tests.
6. Apply one of two statistical tests:
   1. The permutation test. Given two modality labels, compute the average PL of each label and then take the sup norm of their difference. Shuffle the labellings, compute new averages and the new sup norm difference. Compare this to the original difference and determine if the shuffled labelling is significant.
   2. SVM. Given two modality labels, construct a linear SVM for each label and validate it using 10-fold cross validation.

This process is outlined in main.py.

Note: The raw data files are large, so the output of the perseus files have been uploaded to data/postprocessed, so the workflow can start directly from Step 3.