fmralign-benchmark-mockup

This repository contains code for reproducing results from the recent paper:

An empirical evaluation of functional alignment using inter-subject decoding.

Thomas Bazeille*, Elizabeth DuPre*, Jean-Baptiste Poline, & Bertrand Thirion.

doi: 10.1101/2020.12.07.415000

This module provides:

• code to replicate easily all main and supplementary experiments from this paper
• a fetcher for IBC dataset derivatives used (comprising 2 out of 5 decoding tasks studied)

Requirements

Replication of the main benchmarking results already require significant computational power (around 150+CPU hours) and RAM (50+Go). Parallelism is provided in this repository to ease replication on clusters, if available.

Dependencies :

• fmralign
• nibabel>=3.1
• numpy>=1.18
• matplotlib
• pandas
• scipy
• scikit-learn

Installation

First, make sure you have installed all the dependencies listed above. Then you can install fmralign-benchmark by running the following commands:

git clone https://github.com/thomasbazeille/fmralign-benchmark-mockup
cd fmralign-benchmark-mockup
pip install -e .

To reproduce results from the Searchlight Hyperalignment method, you'll also need to install PyMVPA. You can do so with the following commands, assuming you are still in the ``fmralign-benchmark-mockup`` directory:

cd ..
git clone https://github.com/PyMPVA/PyMVPA
cd PyMVPA
pip install -e .

You can confirm that both packages have installed correctly by opening a Python terminal and running the following commands:

import fmralignbench
import mvpa2

Getting started

In order to make code runnable :

1. Please modify ``conf.py`` to provide :

• a ROOT_FOLDER to download initial data and store derivatives
• N_JOBS the number of CPU cores usable for replication.

2) Execute a file from the experiments folder (which includes code to re-execute all of the main and supplemental experiments included in the manuscript):

python experiments/experiment_1-2.py

Experiments description

• experiment_1-2.py replicates the whole-brain and ROI-based level of analysis (170 CPU hours, 30+Go RAM)
• experiment_3.py replicates the qualitative comparison of alignment methods on IBC data (a few CPU hours, few hours, 30+Go RAM)
• supplementary_2-3.py replicates the supplemental experiments investigating the impact of parcellation and smoothing
• supplementary_4.py replicates the supplemental experiment comparing surface- and volume-based results for piecewise Procrustes (13Go download(high-resolution data), 45 CPU hours, 60Go RAM)

Replication outputs

Experiment 1

Experiment 2

Experiment 3

Supplementary results