Extracting signals from a brain parcellation

Here we show how to extract signals from a brain parcellation and compute a correlation matrix.

We also show the importance of defining good confounds signals: the first correlation matrix is computed after regressing out simple confounds signals: movement regressors, white matter and CSF signals, … The second one is without any confounds: all regions are connected to each other. Finally we demonstrated the functionality of load_confounds to flexibly select confound variables from fMRIPrep outputs while following some implementation guideline of fMRIPrep confounds documentation https://fmriprep.org/en/stable/outputs.html#confounds.

One reference that discusses the importance of confounds is Varoquaux and Craddock[1].

This is just a code example, see the corresponding section in the documentation for more.

Note

If you are using Nilearn with a version older than 0.9.0, then you should either upgrade your version or import maskers from the input_data module instead of the maskers module.

That is, you should manually replace in the following example all occurrences of:

from nilearn.maskers import NiftiMasker

with:

from nilearn.input_data import NiftiMasker

Retrieve the atlas and the data

from nilearn import datasets

dataset = datasets.fetch_atlas_harvard_oxford("cort-maxprob-thr25-2mm")
atlas_filename = dataset.maps
labels = dataset.labels

print(f"Atlas ROIs are located in nifti image (4D) at: {atlas_filename}")

# One subject of brain development fMRI data
data = datasets.fetch_development_fmri(n_subjects=1, reduce_confounds=True)
fmri_filenames = data.func[0]
reduced_confounds = data.confounds[0]  # This is a preselected set of confounds

Extract signals on a parcellation defined by labels

Using the NiftiLabelsMasker

from nilearn.maskers import NiftiLabelsMasker

masker = NiftiLabelsMasker(
    labels_img=atlas_filename,
    standardize="zscore_sample",
    standardize_confounds=True,
    memory="nilearn_cache",
    verbose=5,
)

# Here we go from nifti files to the signal time series in a numpy
# array. Note how we give confounds to be regressed out during signal
# extraction
time_series = masker.fit_transform(fmri_filenames, confounds=reduced_confounds)

Compute and display a correlation matrix

from nilearn.connectome import ConnectivityMeasure

correlation_measure = ConnectivityMeasure(
    kind="correlation",
    standardize="zscore_sample",
)
correlation_matrix = correlation_measure.fit_transform([time_series])[0]

# Plot the correlation matrix
import numpy as np

from nilearn import plotting

# Make a large figure
# Mask the main diagonal for visualization:
np.fill_diagonal(correlation_matrix, 0)
# The labels we have start with the background (0), hence we skip the
# first label
# matrices are ordered for block-like representation
plotting.plot_matrix(
    correlation_matrix,
    figure=(10, 8),
    labels=labels[1:],
    vmax=0.8,
    vmin=-0.8,
    title="Confounds",
    reorder=True,
)

Extract signals and compute a connectivity matrix without confounds removal

After covering the basic of signal extraction and functional connectivity matrix presentation, let’s look into the impact of confounds to fMRI signal and functional connectivity. Firstly let’s find out what a functional connectivity matrix looks like without confound removal.

time_series = masker.fit_transform(fmri_filenames)
# Note how we did not specify confounds above. This is bad!

correlation_matrix = correlation_measure.fit_transform([time_series])[0]

np.fill_diagonal(correlation_matrix, 0)

plotting.plot_matrix(
    correlation_matrix,
    figure=(10, 8),
    labels=labels[1:],
    vmax=0.8,
    vmin=-0.8,
    title="No confounds",
    reorder=True,
)

Load confounds from file using a flexible strategy with fmriprep interface

The load_confounds function provides flexible parameters to retrieve the relevant columns from the TSV file generated by fMRIPrep. load_confounds ensures two things:

1. The correct regressors are selected with provided strategy, and

2. Volumes such as non-steady-state and/or high motion volumes are masked out correctly.

Let’s try a simple strategy removing motion, white matter signal, cerebrospinal fluid signal with high-pass filtering.

from nilearn.interfaces.fmriprep import load_confounds

confounds_simple, sample_mask = load_confounds(
    fmri_filenames,
    strategy=["high_pass", "motion", "wm_csf"],
    motion="basic",
    wm_csf="basic",
)

print("The shape of the confounds matrix is:", confounds_simple.shape)
print(confounds_simple.columns)

time_series = masker.fit_transform(
    fmri_filenames, confounds=confounds_simple, sample_mask=sample_mask
)

correlation_matrix = correlation_measure.fit_transform([time_series])[0]

np.fill_diagonal(correlation_matrix, 0)

plotting.plot_matrix(
    correlation_matrix,
    figure=(10, 8),
    labels=labels[1:],
    vmax=0.8,
    vmin=-0.8,
    title="Motion, WM, CSF",
    reorder=True,
)

Motion-based scrubbing

With a scrubbing-based strategy, load_confounds returns a sample_mask that removes the index of volumes exceeding the framewise displacement and standardized DVARS threshold, and all the continuous segment with less than five volumes. Before applying scrubbing, it’s important to access the percentage of volumns scrubbed. Scrubbing is not a suitable strategy for datasets with too many high motion subjects. On top of the simple strategy above, let’s add scrubbing to our strategy.

confounds_scrub, sample_mask = load_confounds(
    fmri_filenames,
    strategy=["high_pass", "motion", "wm_csf", "scrub"],
    motion="basic",
    wm_csf="basic",
    scrub=5,
    fd_threshold=0.5,
    std_dvars_threshold=1.5,
)

print(
    f"After scrubbing, {sample_mask.shape[0]} "
    f"out of {confounds_scrub.shape[0]} volumes remains"
)
print("The shape of the confounds matrix is:", confounds_simple.shape)
print(confounds_scrub.columns)

time_series = masker.fit_transform(
    fmri_filenames, confounds=confounds_scrub, sample_mask=sample_mask
)

correlation_matrix = correlation_measure.fit_transform([time_series])[0]

np.fill_diagonal(correlation_matrix, 0)

plotting.plot_matrix(
    correlation_matrix,
    figure=(10, 8),
    labels=labels[1:],
    vmax=0.8,
    vmin=-0.8,
    title="Motion, WM, CSF, Scrubbing",
    reorder=True,
)

The impact of global signal removal

Global signal removes the grand mean from your signal. The benefit is that it can remove impacts of physiological artifacts with minimal impact on the degrees of freedom. The downside is that one cannot get insight into variance explained by certain sources of noise. Now let’s add global signal to the simple strategy and see its impact.

confounds_minimal_no_gsr, sample_mask = load_confounds(
    fmri_filenames,
    strategy=["high_pass", "motion", "wm_csf", "global_signal"],
    motion="basic",
    wm_csf="basic",
    global_signal="basic",
)
print("The shape of the confounds matrix is:", confounds_minimal_no_gsr.shape)
print(confounds_minimal_no_gsr.columns)

time_series = masker.fit_transform(
    fmri_filenames, confounds=confounds_minimal_no_gsr, sample_mask=sample_mask
)

correlation_matrix = correlation_measure.fit_transform([time_series])[0]

np.fill_diagonal(correlation_matrix, 0)

plotting.plot_matrix(
    correlation_matrix,
    figure=(10, 8),
    labels=labels[1:],
    vmax=0.8,
    vmin=-0.8,
    title="Motion, WM, CSF, GSR",
    reorder=True,
)

Using predefined strategies

Instead of customizing the strategy through load_confounds, one can use a predefined strategy with load_confounds_strategy. Based on the confound variables generated through fMRIPrep, and past benchmarks studies (Ciric et al.[2], Parkes et al.[3]): simple, scrubbing, compcor, ica_aroma. The following examples shows how to use the simple strategy and overwrite the motion default to basic.

from nilearn.interfaces.fmriprep import load_confounds_strategy

# use default parameters
confounds, sample_mask = load_confounds_strategy(
    fmri_filenames, denoise_strategy="simple", motion="basic"
)
time_series = masker.fit_transform(
    fmri_filenames, confounds=confounds, sample_mask=sample_mask
)

correlation_matrix = correlation_measure.fit_transform([time_series])[0]

np.fill_diagonal(correlation_matrix, 0)

plotting.plot_matrix(
    correlation_matrix,
    figure=(10, 8),
    labels=labels[1:],
    vmax=0.8,
    vmin=-0.8,
    title="simple",
    reorder=True,
)

# add optional parameter global signal
confounds, sample_mask = load_confounds_strategy(
    fmri_filenames,
    denoise_strategy="simple",
    motion="basic",
    global_signal="basic",
)
time_series = masker.fit_transform(
    fmri_filenames, confounds=confounds, sample_mask=sample_mask
)

correlation_matrix = correlation_measure.fit_transform([time_series])[0]

np.fill_diagonal(correlation_matrix, 0)

plotting.plot_matrix(
    correlation_matrix,
    figure=(10, 8),
    labels=labels[1:],
    vmax=0.8,
    vmin=-0.8,
    title="simple with global signal",
    reorder=True,
)

plotting.show()

References

[1]

Gaël Varoquaux and R. Cameron Craddock. Learning and comparing functional connectomes across subjects. NeuroImage, 80:405–415, 2013. Mapping the Connectome. doi:10.1016/j.neuroimage.2013.04.007.

[2]

Rastko Ciric, Daniel H. Wolf, Jonathan D. Power, David R. Roalf, Graham L. Baum, Kosha Ruparel, Russell T. Shinohara, Mark A. Elliott, Simon B. Eickhoff, Christos Davatzikos, Ruben C. Gur, Raquel E. Gur, Danielle S. Bassett, and Theodore D. Satterthwaite. Benchmarking of participant-level confound regression strategies for the control of motion artifact in studies of functional connectivity. NeuroImage, 154(1):174–187, 2017. doi:10.1016/j.neuroimage.2017.03.020.

[3]

Linden Parkes, Ben Fulcher, Murat Yücel, and Alex Fornito. An evaluation of the efficacy, reliability, and sensitivity of motion correction strategies for resting-state functional MRI. NeuroImage, 171:415–436, May 2018. doi:10.1016/j.neuroimage.2017.12.073.

Estimated memory usage: 0 MB