This demonstration (in submission for the studyforrest competition) shows how surface-based whole brain second-order representational similarity can be visualized interactively at a random-effect group level using PyMVPA, Python, and AFNI / SUMA.

Demos and download materials


Using the pipeline described below, the user can select a node on the surface. Almost instantaneously, the elements of the representational similarity matrix (time-by-time) associated with the node are correlated (by AFNI's 3dGroupInCorr) with the representational similarity matrices of all nodes on the surfaces. This is done for each subject separately, resulting in a correlation map for each subject. The correlation maps are averaged into a group map, thresholded using a t-test and an a-priori threshold, and visualized on the surface.

Instructions to run the demo

Get the precomputed results and make sure to have a working AFNI installation on a unix-like platform, then:

  • cd group
  • In one terminal window, run 3dGroupInCorr -setA ALL.grpincorr.niml -suma
  • In another terminal window, run suma -spec mh_ico16_fsavg.spec -niml
  • In the SUMA surface window, use ctrl+shift+rclick (where rclick indicates a click with the right-hand mouse button) multiple times to select different nodes.


Data was preprocessed according to the following pipeline:

  • Input data was freesurfer and BOLD data from the studyforrest dataset.

  • FreeSurfer surfaces were preprocessed using pymvpa2-prep-afni-surf from the PyMVPA toolbox, with the "moco" EPI image as reference. This module provides two-hemisphere surfaces in standard space at various resolutions. Run with:

    pymvpa2-prep-afni-surf \
       -r moco_ref_ss+orig \
       -d ref/sub[snum] \
       -d freesurfer/sub[snum]/surf

    with snum from 011 to 020 and moco_ref_ss+orig the output of AFNI's 3dSkullStrip applied to moco_ref.nii.gz.

  • reads data from each run (in the example: one run is used) and stores a down-sampled version (in the example: 4 timesteps) as a PyMVPA dataset. Run with [snum]

  • computes surface-based neighborhood. Run with [snum] 16 100 for an low-res ico value of 16 (see below) and about 100 voxels in each searchlight.

  • helper module to compute dissimilarity measures

  • runs a surface-based searchlight using the neighborhood and downsampled data with the DSCorrMeasure. For each searchlight, a time-by-time similarity matrix is computed, and the upper diagonal elements are flatten into a vector. This vector is then treated as a pseudo-timeseries. Run with [snum] [snum] 16 100 for an low-res ico value of 16 (see below) and about 100 voxels in each searchlight.

  • pseudo-timeseries data was aggregated and preprocessed across subjects using the (in PyMVPA mvpa2/support/afni). Run with:

    python \
        -N 5128 -g ALL -p -g _foo -i fn1 fn2 ... fnK

    where 5128=(16**2*10+2)*2 is the number of nodes, and the fn... are the .niml.dset files from the output of the previous step.

  • data was visualized interactively using 3dGroupInCorr and SUMA. When in SUMA a node is selected as a 'source' node, the correlation between the pseudo-timeseries of the source node and all nodes on the surface is computed, for each subject. The average correlation is visualized on the surface, while the map is thresholded using an a-priori defined threshold and node-wise one-sample t-test against zero.

Demo video

The demo video was based using the following parameters:

  • low-res ico value: 16 (5128 nodes). This surface is used for visualization.
  • high-res ico value: 128 (327684). This surface is used to delineate the searchlights (curved cylinders between the pial and white matter surfaces).
  • searchlight size: 100 voxels (features).
  • participants: BOLD data studyforrest dataset were used.
  • input data: first run of studyforrest dataset of participants 11-20, with the data binned in sets of 4 consecutive timepoints.
  • visualization was based on the FreeSurfer surface template.

Possible extensions and future work

  • depending on available computational resources, higher resolution surfaces and longer time series can be used for a better estimate of similarity. (server issues around the time of the deadline prevented computing these for the submission).
  • if multiple repetitions of the same stimulus are measured for each subject, the correlations can be computed across repetitions. This could decrease potential confounds caused by local (shared voxels) or global (shared common signal) similarity values across locations.
  • whole-brain Hyperalignment preprocessing can be used with this method
  • clustering methods can be applied to find networks with shared second-order representational similarity, and these can be compared with results from other methods (e.g. resting state).

About this work

This article has been contributed by Nick Oosterhof as a submission to the real-life cognition contest. The author can be contacted by email at nikolaas dot oosterhof at unitn dot it.

All referenced source code and materials are copyright (c) 2014, Nikolaas N. Oosterhof and are made available under the terms of the MIT license.