In [ ]:
%matplotlib inline

Searchlight analysis of face vs house recognition

Searchlight analysis requires fitting a classifier a large amount of times. As a result, it is an intrinsically slow method. In order to speed up computing, in this example, Searchlight is run only on one slice on the fMRI (see the generated figures).

Load Haxby dataset

In [ ]:
import pandas as pd
from nilearn import datasets
from nilearn.image import new_img_like, load_img, get_data

# We fetch 2nd subject from haxby datasets (which is default)
haxby_dataset = datasets.fetch_haxby()

# print basic information on the dataset
print('Anatomical nifti image (3D) is located at: %s' % haxby_dataset.mask)
print('Functional nifti image (4D) is located at: %s' % haxby_dataset.func[0])

fmri_filename = haxby_dataset.func[0]
labels = pd.read_csv(haxby_dataset.session_target[0], sep=" ")
y = labels['labels']
session = labels['chunks']

Restrict to faces and houses

In [ ]:
from nilearn.image import index_img
condition_mask = y.isin(['face', 'house'])

fmri_img = index_img(fmri_filename, condition_mask)
y, session = y[condition_mask], session[condition_mask]

Prepare masks

  • mask_img is the original mask
  • process_mask_img is a subset of mask_img, it contains the voxels that should be processed (we only keep the slice z = 26 and the back of the brain to speed up computation)
In [ ]:
import numpy as np

mask_img = load_img(haxby_dataset.mask)

# .astype() makes a copy.
process_mask = get_data(mask_img).astype(
picked_slice = 29
process_mask[..., (picked_slice + 1):] = 0
process_mask[..., :picked_slice] = 0
process_mask[:, 30:] = 0
process_mask_img = new_img_like(mask_img, process_mask)

Searchlight computation

In [ ]:
# Make processing parallel
# /!\ As each thread will print its progress, n_jobs > 1 could mess up the
#     information output.
n_jobs = 1

# Define the cross-validation scheme used for validation.
# Here we use a KFold cross-validation on the session, which corresponds to
# splitting the samples in 4 folds and make 4 runs using each fold as a test
# set once and the others as learning sets
from sklearn.model_selection import KFold
cv = KFold(n_splits=4)

import nilearn.decoding
# The radius is the one of the Searchlight sphere that will scan the volume
searchlight = nilearn.decoding.SearchLight(
    radius=5.6, n_jobs=n_jobs,
    verbose=1, cv=cv), y)

F-scores computation

In [ ]:
from nilearn.input_data import NiftiMasker

# For decoding, standardizing is often very important
nifti_masker = NiftiMasker(mask_img=mask_img, sessions=session,
                           standardize=True, memory='nilearn_cache',
fmri_masked = nifti_masker.fit_transform(fmri_img)

from sklearn.feature_selection import f_classif
f_values, p_values = f_classif(fmri_masked, y)
p_values = -np.log10(p_values)
p_values[p_values > 10] = 10
p_unmasked = get_data(nifti_masker.inverse_transform(p_values))


Use the fmri mean image as a surrogate of anatomical data

In [ ]:
from nilearn import image
mean_fmri = image.mean_img(fmri_img)

from nilearn.plotting import plot_stat_map, plot_img, show
searchlight_img = new_img_like(mean_fmri, searchlight.scores_)

# Because scores are not a zero-center test statistics, we cannot use
# plot_stat_map
plot_img(searchlight_img, bg_img=mean_fmri,
         title="Searchlight", display_mode="z", cut_coords=[-9],
         vmin=.42, cmap='hot', threshold=.2, black_bg=True)

# F_score results
p_ma =, mask=np.logical_not(process_mask))
f_score_img = new_img_like(mean_fmri, p_ma)
plot_stat_map(f_score_img, mean_fmri,
              title="F-scores", display_mode="z",