Deriving spatial maps from group fMRI data using ICA and Dictionary Learning

Various approaches exist to derive spatial maps or networks from group fmr data. The methods extract distributed brain regions that exhibit similar BOLD fluctuations over time. Decomposition methods allow for generation of many independent maps simultaneously without the need to provide a priori information (e.g. seeds or priors.)

This example will apply two popular decomposition methods, ICA and Dictionary learning, to fMRI data measured while children and young adults watch movies. The resulting maps will be visualized using atlas plotting tools.

CanICA is an ICA method for group-level analysis of fMRI data. Compared to other strategies, it brings a well-controlled group model, as well as a thresholding algorithm controlling for specificity and sensitivity with an explicit model of the signal.

The reference paper is Varoquaux et al.[1].

Load brain development fMRI dataset

from nilearn.datasets import fetch_development_fmri

rest_dataset = fetch_development_fmri(n_subjects=30)
func_filenames = rest_dataset.func  # list of 4D nifti files for each subject

# print basic information on the dataset
print(f"First functional nifti image (4D) is at: {rest_dataset.func[0]}")

[fetch_development_fmri] Dataset found in
/home/runner/nilearn_data/development_fmri
[fetch_development_fmri] Dataset found in
/home/runner/nilearn_data/development_fmri/development_fmri
[fetch_development_fmri] Dataset found in
/home/runner/nilearn_data/development_fmri/development_fmri
First functional nifti image (4D) is at: /home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar128_task-pixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz

Apply CanICA on the data

We use “whole-brain-template” as a strategy to compute the mask, as this leads to slightly faster and more reproducible results. However, the images need to be in MNI template space.

import warnings

from sklearn.exceptions import ConvergenceWarning

from nilearn.decomposition import CanICA

canica = CanICA(
    n_components=20,
    memory="nilearn_cache",
    memory_level=1,
    verbose=1,
    random_state=0,
    standardize="zscore_sample",
    mask_strategy="whole-brain-template",
    n_jobs=2,
)
with warnings.catch_warnings():
    # silence warnings about ICA not converging
    # Consider increasing tolerance or the maximum number of iterations.
    warnings.filterwarnings(action="ignore", category=ConvergenceWarning)
    canica.fit(func_filenames)

# Retrieve the independent components in brain space. Directly
# accessible through attribute `components_img_`.
canica_components_img = canica.components_img_
# components_img is a Nifti Image object, and can be saved to a file with
# the following lines:
from pathlib import Path

output_dir = Path.cwd() / "results" / "plot_compare_decomposition"
output_dir.mkdir(exist_ok=True, parents=True)
print(f"Output will be saved to: {output_dir}")
canica_components_img.to_filename(output_dir / "canica_resting_state.nii.gz")

[CanICA.fit] Loading data from
['/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar128_task-
pixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar126_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar125_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar124_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar123_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar127_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar024_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar023_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar022_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar021_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar020_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar019_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar018_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar017_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar016_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar001_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar013_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar012_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar011_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar010_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar009_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar008_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar007_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar006_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar005_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar004_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar003_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar002_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar014_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar015_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz']
[CanICA.fit] Computing mask
[CanICA.fit] Resamping mask
[CanICA.fit] Finished fit
[CanICA.fit] Loading data
[CanICA.fit] Computing image from signals
[Parallel(n_jobs=2)]: Using backend LokyBackend with 2 concurrent workers.
[Parallel(n_jobs=2)]: Done  10 out of  10 | elapsed:    3.9s finished
[CanICA.fit] Computing image from signals
Output will be saved to: /home/runner/work/nilearn/nilearn/examples/03_connectivity/results/plot_compare_decomposition

To visualize we plot the outline of all components on one figure

from nilearn.plotting import plot_prob_atlas

# Plot all ICA components together
plot_prob_atlas(canica_components_img, title="All ICA components")

/home/runner/work/nilearn/nilearn/.tox/doc/lib/python3.10/site-packages/numpy/ma/core.py:2826: UserWarning: Warning: converting a masked element to nan.
  _data = np.array(data, dtype=dtype, copy=copy,

<nilearn.plotting.displays._slicers.OrthoSlicer object at 0x7f96f04a2920>

Finally, we plot the map for each ICA component separately

from nilearn.image import iter_img
from nilearn.plotting import plot_stat_map, show

for i, cur_img in enumerate(iter_img(canica_components_img)):
    plot_stat_map(
        cur_img,
        display_mode="z",
        title=f"IC {int(i)}",
        cut_coords=1,
        vmax=0.05,
        vmin=-0.05,
        colorbar=False,
    )


show()

• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 

Compare CanICA to dictionary learning

Dictionary learning is a sparsity based decomposition method for extracting spatial maps. It extracts maps that are naturally sparse and usually cleaner than ICA. Here, we will compare networks built with CanICA to networks built with Dictionary learning.

For more detailse see Mensch et al.[2].

Create a dictionary learning estimator

from nilearn.decomposition import DictLearning

dict_learning = DictLearning(
    n_components=20,
    memory="nilearn_cache",
    memory_level=1,
    verbose=1,
    random_state=0,
    n_epochs=1,
    mask_strategy="whole-brain-template",
    standardize="zscore_sample",
    n_jobs=2,
)

print("[Example] Fitting dictionary learning model")
dict_learning.fit(func_filenames)
print("[Example] Saving results")
# Grab extracted components umasked back to Nifti image.
# Note: For older versions, less than 0.4.1. components_img_
# is not implemented. See Note section above for details.
dictlearning_components_img = dict_learning.components_img_
dictlearning_components_img.to_filename(
    output_dir / "dictionary_learning_resting_state.nii.gz"
)

[Example] Fitting dictionary learning model
[DictLearning.fit] Loading data from
['/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar128_task-
pixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar126_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar125_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar124_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar123_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar127_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar024_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar023_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar022_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar021_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar020_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar019_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar018_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar017_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar016_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar001_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar013_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar012_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar011_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar010_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar009_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar008_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar007_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar006_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar005_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar004_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar003_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar002_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar014_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz',
'/home/runner/nilearn_data/development_fmri/development_fmri/sub-pixar015_task-p
ixar_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz']
[DictLearning.fit] Computing mask
[DictLearning.fit] Resamping mask
[DictLearning.fit] Finished fit
[DictLearning.fit] Loading data
[DictLearning.fit] Learning initial components
[Parallel(n_jobs=2)]: Using backend LokyBackend with 2 concurrent workers.
[Parallel(n_jobs=2)]: Done   1 out of   1 | elapsed:    0.4s finished
[DictLearning.fit] Computing initial loadings
________________________________________________________________________________
[Memory] Calling nilearn.decomposition.dict_learning._compute_loadings...
_compute_loadings(array([[ 0.003026, ..., -0.000948],
       ...,
       [ 0.001213, ...,  0.014412]]),
array([[-0.280625, ...,  0.825802],
       ...,
       [-0.997198, ..., -0.015035]]))
_________________________________________________compute_loadings - 0.2s, 0.0min
[DictLearning.fit]  Learning dictionary
________________________________________________________________________________
[Memory] Calling sklearn.decomposition._dict_learning.dict_learning_online...
dict_learning_online(array([[-0.280625, ..., -0.997198],
       ...,
       [ 0.825802, ..., -0.015035]]),
20, alpha=10, batch_size=20, method='cd', dict_init=array([[-0.168973, ...,  0.081441],
       ...,
       [ 0.00847 , ..., -0.496571]]), verbose=0, random_state=0, return_code=True, shuffle=True, n_jobs=1, max_iter=1090)
_____________________________________________dict_learning_online - 1.4s, 0.0min
[DictLearning.fit] Computing image from signals
[Example] Saving results

Visualize the results

First plot all DictLearning components together

plot_prob_atlas(
    dictlearning_components_img, title="All DictLearning components"
)

/home/runner/work/nilearn/nilearn/.tox/doc/lib/python3.10/site-packages/numpy/ma/core.py:2826: UserWarning: Warning: converting a masked element to nan.
  _data = np.array(data, dtype=dtype, copy=copy,

<nilearn.plotting.displays._slicers.OrthoSlicer object at 0x7f96f4d86b00>

One plot of each component

for i, cur_img in enumerate(iter_img(dictlearning_components_img)):
    plot_stat_map(
        cur_img,
        display_mode="z",
        title=f"Comp {int(i)}",
        cut_coords=1,
        vmax=0.1,
        vmin=-0.1,
        colorbar=False,
    )

• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 
• 

Estimate explained variance per component and plot using matplotlib

The fitted object dict_learning can be used to calculate the score per component

scores = dict_learning.score(func_filenames, per_component=True)

# Plot the scores
import numpy as np
from matplotlib import pyplot as plt
from matplotlib.ticker import FormatStrFormatter

plt.figure(figsize=(4, 4), constrained_layout=True)

positions = np.arange(len(scores))
plt.barh(positions, scores)
plt.ylabel("Component #", size=12)
plt.xlabel("Explained variance", size=12)
plt.yticks(np.arange(20))
plt.gca().xaxis.set_major_formatter(FormatStrFormatter("%.3f"))

show()

________________________________________________________________________________
[Memory] Calling nilearn.decomposition._base._explained_variance...
_explained_variance(array([[-2.806378e-01, ...,  8.257976e-01],
       ...,
       [-1.977336e-15, ..., -3.897215e-16]]),
array([[ 0., ...,  0.],
       ...,
       [-0., ...,  0.]]), per_component=True)
/home/runner/work/nilearn/nilearn/.tox/doc/lib/python3.10/site-packages/nilearn/decomposition/_base.py:666: UserWarning: Persisting input arguments took 1.31s to run.
If this happens often in your code, it can cause performance problems
(results will be correct in all cases).
The reason for this is probably some large input arguments for a wrapped
 function (e.g. large strings).
THIS IS A JOBLIB ISSUE. If you can, kindly provide the joblib's team with an
 example so that they can fix the problem.
  return self._cache(_explained_variance)(
______________________________________________explained_variance - 14.1s, 0.2min

Note

To see how to extract subject-level timeseries’ from regions created using Dictionary learning, see example Regions extraction using dictionary learning and functional connectomes.

References

[1]

Gael Varoquaux, Sepideh Sadaghiani, Philippe Pinel, Andreas Kleinschmidt, Jean-Baptiste Poline, and Bertrand Thirion. A group model for stable multi-subject ica on fmri datasets. NeuroImage, 51(1):288–299, 2010. URL: https://pubmed.ncbi.nlm.nih.gov/20153834, doi:10.1016/j.neuroimage.2010.02.010.

[2]

Arthur Mensch, Gael Varoquaux, and Bertrand Thirion. Compressed online dictionary learning for fast resting-state fmri decomposition. In 2016 IEEE 13th International Symposium on Biomedical Imaging (ISBI), volume, 1282–1285. 2016. doi:10.1109/ISBI.2016.7493501.

Estimated memory usage: 6127 MB