Turn it sideways

This Github repository contains the scripts and preliminary results for a classification-based approach at determining the specificity of functional ROIs.

Method summary

The method in questions works by 1) deriving sensitivities during classification analysis in a leave-one-participant-out crossvalidation of any two ROIs, and 2) modelling this time course of sensitivities with the available experimental design to provide a functional description of it. Comparing this approach with the typical "standard" univariate GLM with canonical contrasts on the fMRI data (e.g. faces versus houses) shows that 1) such canonical contrasts may not work well for complex stimulation 2) the functional specificity of ROIs derived from simplistic designs is diverse during complex stimulation

Check out the poster in derivatives/poster for a graphical description of analyses and results.

Repo overview

The repository is a datalad dataset. In theory, the commit history should contain all relevant information about the origin of source data, its preprocessing, the performed analyses and the corresponding results. Nevertheless, the history is long, and scripts underwent constant refactoring. Therefore, this README tries to help by giving an overview of the contents of this repository.

Sourcedata

The directory sourcedata/ contains submodules with the various input files (all linked as datalad subdatasets):

• studyforrest phase 2 movie + block design data
• studyforrest movie annotation data
• studyforrest visualrois data
• custom additional automatic movie annotation with pliers (McNamara et al., 2017), (data yet to be published)
• templates and transformations for the studyforrest dataset

Code

The directory code/ contains all scripts used for the analyses. The important ones are:

• fsl_warp*: warps images from subject- to groupspace
• preproc*: preprocessing workflow
• create_hdf5_ds: creates groupdatasets from all subjects
• cv_clf_glm_main.py: main analysis script for all analyses
• utils.py: Helper function for the main analyses
• *face*: Scripts to preprocesses face annotations

The main analysis script contains commandline-specifyable options to compute partial results (classification, glm, plotting) on the full or only-ROI dataset, containing either no coordinates, additional coordinate information, or only coordinate information, with different classifiers (GLM, SGD, l-SGD), and to reverse the analysis. The create_hdf5_ds.py script contain commandline-specifyable options to zscore normally, not at all, or - in the case of the localizer data - on parameters derived from rest periods with no stimulation.

sub-* directories

Each subjects subdirectory contains - per session (block-design, movie) - custom preprocessed files in subject and group space. Preprocessing was done with a nipype workflow that can be found in the code directory as code/preprocess_locdata.py. The script was executed with datalad-containers run in a singularity image (https://www.singularity-hub.org/collections/1877) Subject space files were warped into a study-specific group space based on templates in the studyforrest dataset, using FSL and the script code/fsl_warp_from-BOLD_to-group_mode-image.py.

Derivatives

The derivatives/ directory contains

• groupdatasets/: large groupdatasets in subject space containing the block design or moviedata together with ROI information, produced with the script code/create_hdf5_ds.py
• ds_groupspace/: Similar groupdatasets, but in group space. These files are the analysis base.
• stimuli/detected_faces_events.tsv: automatically extracted faces per frame for all runs, based on
• stimuli/researchcut/avmovie: the subsets of this large file corresponding to each movie run. This cutting was done with sourcedata/annotations/code/researchcut2segments.py)
• stimuli/researchcut/downsampled_event_files: The former event files as proper event files (with code/faces2events.py), and downsampled to seconds (with code/downsample_face_events.py)
• results/: [OUTDATED -- IGNORE] Results of various flavours of the analysis on either the location data (performed with code/localizer_cv_glm_analysis.py) or the moviedata (performed with code/avmovie_cv_glm_analysis.py).
• groupspace_analysis/: Analyses results performed in groupspace for various flavours (classification --> GLM/GLM --> classification; GNB/SGD/l-SGD; FFA-vs-PPA/FFA-vs-brain; normalized results/unnormalized results; movie/block-design data; Full dataset/stripped dataset). This is the relevant directory to look for results.
• poster/: contains the corresponding poster presented at the OHBM in Rome

NOTE: all of this here is work in progress. Refactoring happens quite often.

Acknowledgements

The analyses rely on open source software (nipype, numpy, pandas, scipy, sklearn, seaborn, matplotlib, pylab, datalad, git, git-annex, FSL, Neurodebian, PyMVPA, pliers, and likely many more). We want to thank the authors, contributers and maintainers of these awesome software projects for making it possible. The analyses further rely on shared, open data. We want to thank the authors and anonymous subjects of these datasets.