OHBM-2016

Submission to OHBM 2016 on functional lateralization using the neurovault dataset.

Installation

1. clone this repository
2. pip install -r requirements.txt

Note: if using virtualenv, run: fix-osx-virtualenv env/ (if env/ is your virtual env root)

Analyses

Run each script with --help to view all script options.

• analysis.py - Downloads images, computes components, runs sparsity analyses. Key metrics include:
  • HPAI (Hemispheric Participation Asymmetry Index): (L-R)/L+R for # of voxels above a given threshold for each component
  • SSS (Spatial Similarity Score): correlation measure to compare R and L in wb components
  • ACNI (Anti-correlated network index): proportion of anti-correlated network, or the voxels with negative weight (and with magnitude above a given threshold) in a given component.
• match.py - Downloads images, computes components, compares/matches & plots components.
• qc.py - Downloads images, visualizes them for quality control purposes.

Outputs

For analysis.py / match.py:

• ica_nii - directory containing Nifti1 label maps for each of 20 ICA components when run on left-only, right-only, and both hemispheres.
• ica_map - Png images showing each component above (20 for each ICA run) when run on left-only, right-only, and both hemispheres.

For qc.py:

• qc - directory of images showing 16 nii files for review. To exclude images / collections, use fetch_neurovault's filtering procedures

For match.py:

• ica_nii/[dataset]/ - directory containing Nifti1 image maps for n ICA components when run on left-only (L), right-only (R), and both hemispheres (wb) for the given dataset.
• ica_nii/{dataset}/[n] - directory containing PNG images for each of n ICA component
• ica_imgs/[dataset]/analyses/[n] - Matching matrices (wb_{{L/R}}_simmat.png) showing match scores between wb and L/R components for n ICA decomposition.
• ica_imgs/[dataset]/analyses/n/{{forced/unforced}}-match/ - directory containing PNG images for pairs of matched wb and R/L, and combined RL. Forced-match force one-to-one matching, while unforced-match pairs best-matching R/L/RL for every wb component. For each pair, term comparison is also saved (each comparison as PNG and summary as csv).

For analysis.py:

• ica_imgs/[dataset]/analysis/1_wb_HPAI.png - Graph of the HPAI in wb ICA images for a given range of n_components. Size of dot = # voxels above a global threshold.

• ica_imgs/[dataset]/analysis/2_vcSparsity_comparison_{{L/R}}.png - Graph of the sparsity over a range of n_components for wb, R, and L components: if there is no asymmetric activity, the contrast should be similar in wb and R/L, resulting in similar voxel count when comparing for each hemisphere (confirm this by performing the same analysis with dummy half-brain composed of randomly selected voxels). Increased contrast in unilateral components suggest ‘masking’ of lateralized activity by wb analysis.

• ica_imgs/[dataset]/analysis/3_l1Sparsity_comparison_{{L/R}}.png - Same idea as above but calculate sparsity using L1 rather than voxel count with an arbitrary threshold.

• ica_imgs/[dataset]/analysis/4_Matching_results_box.png - Matching results: Graph of the change in matching performance (average dissimilarity score and proportion of unmatched unilateral components) over a range of n_components. High-quality matching suggests refinement via hemi-analysis; low-quality matching suggests masking via wb analysis.

• ica_imgs/[dataset]/analysis/5_wb_RL_SSS_{{box/dots}}.png - Graph of SSS over a range of n_components, comparing SSS of wb and the matched RL. Decrease suggests asymmetric organization revealed by unilateral ICA.

• ica_imgs/[dataset]/analysis/6_ACNI_comparison.png - Graph of ACNI over a range of n_components for L/R and wb analyses.

• ica_imgs/[dataset]/analysis/[n_components]/1_PosNegHPAI_{{n_components}}.png - Relationship between positive and negative HPAI, for each n_component. The robust presence of sparsity on the negative side indicates anticorrelated networks being pulled out as a single component; this graph shows the relationship between HPAI of the anticorrelated networks.

• ica_imgs/[dataset]/analysis/[n_components]/2_HPAIvsSAS_{{n_components}}.png - Graph of the relationship between HPAI and SAS score for a particular n_component decomposition.