Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Understanding the Preprocessed Data

Authors
Affiliations
University of Oxford
University of Toronto / University of Cambridge

What’s in the Pre-Processed Data?

Before we begin analyzing brain networks, it is useful to understand what is actually contained within the preprocessed dataset.

If you would like to explore the original files yourself, you can download the preprocessed MRI data from the LEMON dataset website:

https://fcon_1000.projects.nitrc.org/indi/retro/MPI_LEMON/downloads/download_MRI.html

The preprocessed data for a single participant is approximately 3 GB, so feel free to download one participant and explore the files while following this guide. However, this is not necessary!

A typical participant directory will look something like this:

sub-032301
├── anat
│   └── sub-032301_ses-01_acq-mp2rage_brain.nii.gz
└── func
    ├── sub-032301_ses-01_task-rest_acq-AP_run-01_MNI2mm.nii.gz
    ├── sub-032301_ses-01_task-rest_acq-AP_run-01_QC.pdf
    ├── sub-032301_ses-01_task-rest_acq-AP_run-01_confounds.txt
    └── sub-032301_ses-01_task-rest_acq-AP_run-01_native.nii.gz

What Do These Files Mean?

The dataset is organized into two main folders:

The file extension .nii.gz (a Neuroimaging Informatics Technology Initiative or NIfTI file) is the standard format used in neuroimaging. It stores MRI data in a compressed form and is one of the most common formats you will encounter when working with MRI datasets.

Structural MRI Files

anat/sub-032301_ses-01_acq-mp2rage_brain.nii.gz

This file contains a three-dimensional image of the participant’s brain anatomy after preprocessing. Researchers use these images to study brain structure, align participants to a common template, and define anatomical regions.

MRI images are generated by applying carefully designed sequences of magnetic fields and radiofrequency pulses. Different sequences manipulate the MRI signal in different ways, allowing researchers to emphasize different properties of tissue. MP2RAGE (Marques et al., 2010) is one example of a structural MRI sequence that uses a particular combination of magnetic field manipulations and radiofrequency pulses to produce high-quality images of brain anatomy with clear contrast between different tissue types.

Functional MRI Files

func/sub-032301_ses-01_task-rest_acq-AP_run-01_MNI2mm.nii.gz

This file contains the participant’s resting-state fMRI data.

Unlike a structural MRI scan, which is a single 3D image, a functional MRI scan is a 4D image:

You can think of a resting-state fMRI scan as a sequence of hundreds of 3D brain images collected one after another over time. Each image captures a snapshot of brain activity at a particular moment.

3D/Volumetric Image: Pixel but 3D?

MRI images are 3D images (i.e., 2D images stacked on top of each other).

In normal photograph, the smallest unit is a ‘pixel’. For a 3D MRI image, the smallest unit is called a voxel (short for “volume element”).

A voxel is similar to a pixel in a photograph, except that it occupies a small three-dimensional volume (i.e., a cube) within the brain rather than a two-dimensional location on a screen (i.e., a square).

A typical resting-state fMRI scan may contain hundreds of thousands of voxels. For each voxel, the scanner records how the MRI signal changes over time.

This means that a single resting-state scan contains an enormous amount of data.

Quality Control Reports

func/sub-032301_ses-01_task-rest_acq-AP_run-01_QC.pdf

This file contains quality-control information generated during preprocessing.

Researchers often inspect these reports to ensure that the data quality is acceptable before performing further analyses.

For the LEMON dataset, the researchers have already inspected their pre-processed data and discarded 8 participants. See “Technical Validation” section.

Confounds Files

func/sub-032301_ses-01_task-rest_acq-AP_run-01_confounds.txt

This file contains measurements that may introduce unwanted variation into the fMRI signal.

Examples include:

Later in the competition, you will see that many analysis pipelines use these confounds to remove unwanted sources of variation before estimating functional connectivity (a type of analysis).

Native and Standard-Space Data

You may notice both these two neuroimaging files in func folder:

..._native.nii.gz

and

..._MNI2mm.nii.gz

We know these are neuroimaging files because they are NIfTI (.nii) files.

The native brain image remains in the participant’s original brain space, whereas the MNI2mm brain image has been transformed into a common reference (template) space used throughout neuroimaging research.

Why do we align each participant’s brain to a common space you may ask?

Because every human brain differs slightly in size, shape, and anatomy, scientists cannot directly compare brain activity or anatomical locations across individuals without a standard reference system. MNI space addresses this by mathematically transforming and aligning each participant’s brain to a shared template, allowing corresponding brain regions to be compared across participants.

The Montreal Neurological Institute (MNI) space and Talairach space are two of the most commonly used standardized three-dimensional coordinate systems in human neuroimaging. These coordinate systems allow researchers to report and compare findings using a common anatomical framework.

For the analyses performed in this competition, we will primarily use the MNI2mm images—that is, the images that have already been transformed into MNI space.

Viewing Neuroimaging Files

Throughout this competition, we will primarily use Python and the Nilearn package to work with neuroimaging data.

There are many other neuroimaging software packages available, including:

Each tool has its own strengths and weaknesses. Many neuroimaging tools were developed primarily for Linux and macOS systems, whereas Python-based workflows have become increasingly popular because they are flexible, accessible, and integrate well with modern data science tools.

For this reason, we will focus primarily on Python-based approaches throughout the competition.