python tool for fMRI searchlight mapping
python setup.py develop
python setup.py test
import nibabel as nib
import numpy as np
from searchlight import RSASearchLight
from searchlight.utils import corr_rdms, makeimagestack
from scipy.spatial.distance import squareform
from scipy.stats import rankdata
import matplotlib.pyplot as plt
# we will perform a searchlight with a radius of 3
radius = 3
# get a binary mask of the subject's brain
mask = nib.load('mask.nii').get_data()
# get a x y z by conditions volume of fmri patterns
betas = nib.load('betas.nii').get_data()
x, y, z, n_conditions = betas.shape
assert mask.shape==betas.shape[0:3], 'dimensions of mask and betas must match.'
# initialise the searchlight.
SL = RSASearchLight(
mask, # pass the binary mask
radius=3, # radius of 3
threshold=.7, # threshold of .7 (how much voxels of a sphere need to be inside the brain)
njobs=2, # this will distribute the searchlight mapping on 2 cores.
verbose=True # this will make use of tqdm to display time spent and left
)
SL.fit_rsa(
betas, # pass in the betas
wantreshape=False # the resulting SL.RDM is shape n_centers, n_pairs
)
# SL.RDM is a numpy array with n_centers, n_comps (upper triangular vector)
# SL.RDM (if wantreshape was set to True when we fit, would be a x , y , z by n_pairs array)
index = 100
plt.imshow(squareform(SL.RDM[index, :])) # plots the RDM of the 100th center
# let's make model inference
n_pairs = n_conditions*(n_conditions-1) // 2
model_rdm = np.random.rand(1, n_pairs)
rdm_corr_to_model = corr_rdms(SL.RDM, model_rdm)
brain_vol = np.zeros((x*y*z))
brain_vol[SL.centerIndices] = rdm_corr_to_model
brain_vol = np.reshape(brain_vol, [x, y, z])
# then you can use matplotlib imshow and makeimagestack
# for a quick visualisation.
plt.imshow(makeimagestack(brain_vol))
plt.colorbar()
# in the above example, we computed a pearson correlation between the brain and the model.
# to implement a spearman correlation we rank the two datasets first.
model_rdm_rank = np.apply_along_axis(rankdata, 1, model_rdm)
brain_RDMs_rank = np.apply_along_axis(rankdata, 1, SL.RDM)
# and then correlate
rdm_corr_spearman_to_model = corr_rdms(brain_RDMs_rank, model_rdm_rank)
brain_vol_sp = np.zeros((x*y*z))
brain_vol_sp[SL.centerIndices] = rdm_corr_spearman_to_model
brain_vol_sp = np.reshape(brain_vol_sp, [x, y, z])
# then you can use matplotlib imshow and makeimagestack
# for a quick visualisation.
plt.imshow(makeimagestack(brain_vol_sp))
plt.colorbar()- this searchlight package will ultimately sit on pyrsa.
- e.g. one could import a distance metric from pyrsa.rdm.calc and pass it to SL.fit_rsa as a fitting method.
- write an nibabel backed write_output module.