Representational similarity analysis (RSA) on fMRI dataΒΆ

In this example we are going to take a look at representational similarity analysis (RSA). This term was coined by Kriegeskorte et al. (2008) and refers to a technique where data samples are converted into a self-referential distance space, in order to aid comparison across domains. The premise is that whenever no appropriate transformation is known to directly compare two types of data directly (1st-level), it is still useful to compare similarities computed in individual domains (2nd-level). For example, this analysis technique has been used to identify inter-species commonalities in brain response pattern variations during stimulation with visual objects (single-cell recordings in monkeys compared to human fMRI, Kriegeskorte et al., 2008), and to relate brain response pattern similarities to predictions of computational models of perception (Connolly et al., 2012).

import numpy as np
import pylab as pl
from os.path import join as pjoin
from mvpa2 import cfg

In this example we use a dataset from Haxby et al. (2001) were participants watched pictures of eight different visual objects, while fMRI was recorded. The following snippet load a portion of this dataset (single subject) from regions on the ventral and occipital surface of the brain.

# load dataset -- ventral and occipital ROIs
from mvpa2.datasets.sources.native import load_tutorial_data
datapath = pjoin(cfg.get('location', 'tutorial data'), 'haxby2001')
ds = load_tutorial_data(roi=(15, 16, 23, 24, 36, 38, 39, 40, 48))

We only do minimal pre-processing: linear trend removal and Z-scoring all voxel time-series with respect to the mean and standard deviation of the “rest” condition.

# only minial detrending
from mvpa2.mappers.detrend import poly_detrend
poly_detrend(ds, polyord=1, chunks_attr='chunks')
# z-scoring with respect to the 'rest' condition
from mvpa2.mappers.zscore import zscore
zscore(ds, chunks_attr='chunks', param_est=('targets', 'rest'))
# now remove 'rest' samples
ds = ds[ != 'rest']

RSA is all about so-called dissimilarity matrices: square, symmetric matrices with a zero diagonal that encode the (dis)similarity between all pairs of data samples or conditions in a dataset. We compose a little helper function to plot such matrices, including a color-scale and proper labeling of matrix rows and columns.

# little helper function to plot dissimilarity matrices
# since we are using correlation-distance, we use colorbar range of [0,2]
def plot_mtx(mtx, labels, title):
    pl.imshow(mtx, interpolation='nearest')
    pl.xticks(range(len(mtx)), labels, rotation=-45)
    pl.yticks(range(len(mtx)), labels)
    pl.clim((0, 2))

As a start, we want to inspect the dissimilarity structure of the stimulation conditions in the entire ROI. For this purpose, we average all samples of each conditions into a single exemplar, using an FxMapper() instance.

# compute a dataset with the mean samples for all conditions
from mvpa2.mappers.fx import mean_group_sample
mtgs = mean_group_sample(['targets'])
mtds = mtgs(ds)

After these preparations we can use the PDist() measure to compute the desired distance matrix – by default using correlation distance as a metric. The square argument will cause a ful square matrix to be produced, instead of a leaner upper-triangular matrix in vector form.

# basic ROI RSA -- dissimilarity matrix for the entire ROI
from mvpa2.measures import rsa
dsm = rsa.PDist(square=True)
res = dsm(mtds)
plot_mtx(res,, 'ROI pattern correlation distances')

Inspecting the figure we can see that there is not much structure in the matrix, except for the face and the house condition being slightly more dissimilar than others.

Now, let’s take a look at the variation of similarity structure through the brain. We can plug the PDist() measure into a searchlight to quickly scan the brain and harvest this information.

# same as above, but done in a searchlight fashion
from mvpa2.measures.searchlight import sphere_searchlight
dsm = rsa.PDist(square=False)
sl = sphere_searchlight(dsm, 2)
slres = sl(mtds)

The result is a compact distance matrix in vector form for each searchlight location. We can now try to score each matrix. Let’s find the distance matrix with the largest overall distances across all stimulation conditions, i.e. the location in the brain where brain response patterns are most dissimilar.

# score each searchlight sphere result wrt global pattern dissimilarity
distinctiveness = np.sum(np.abs(slres), axis=0)
print 'Most dissimilar patterns around', \
# take a look at the this dissimilarity structure
from scipy.spatial.distance import squareform
plot_mtx(squareform(slres.samples[:, distinctiveness.argmax()]),,
         'Maximum distinctive searchlight pattern correlation distances')

That looks confusing. But how do we know that this is not just noise (it probably is)? One way would be to look at how stable a distance matrix is, when computed for different portions of a dataset.

To perform this analysis, we use another FxMapper() instance that averages all data into a single sample per stimulation conditions, per chunk. A chunk in this context indicates a complete fMRI recording run.

# more interesting: let's look at the stability of similarity structures
# across experiment runs
# mean condition samples, as before, but now individually for each run
mtcgs = mean_group_sample(['targets', 'chunks'])
mtcds = mtcgs(ds)

With this dataset we can use PDistConsistency() to compute the similarity of dissimilarity matrices computes from different chunks. And, of course, it can be done in a searchlight.

# searchlight consistency measure -- how correlated are the structures
# across runs
dscm = rsa.PDistConsistency()
sl_cons = sphere_searchlight(dscm, 2)
slres_cons = sl_cons(mtcds)

Now we can determine the brain location with the most stable dissimilarity matrix.

# mean correlation
mean_consistency = np.mean(slres_cons, axis=0)
print 'Most stable dissimilarity patterns around', \
# Look at this pattern
plot_mtx(squareform(slres.samples[:, mean_consistency.argmax()]),,
         'Most consistent searchlight pattern correlation distances')

It is all in the face!

It would be interesting to know where in the brain dissimilarity structures can be found that are similar to this one. PDistTargetSimilarity() can be used to discover this kind of information with any kind of target dissimilarity structure. We need to transpose the result for aggregation into a searchlight map, as PDistTargetSimilarity can return more features than just the correlation coefficient.

# let's see where in the brain we find dissimilarity structures that are
# similar to our most stable one
tdsm = rsa.PDistTargetSimilarity(
            slres.samples[:, mean_consistency.argmax()])
# using a searchlight
from mvpa2.base.learner import ChainLearner
from mvpa2.mappers.shape import TransposeMapper
sl_tdsm = sphere_searchlight(ChainLearner([tdsm, TransposeMapper()]), 2)
slres_tdsm = sl_tdsm(mtds)

Lastly, we can map this result back onto the 3D voxel grid, and overlay it onto the brain anatomy.

# plot the spatial distribution using NiPy
vol = ds.a.mapper.reverse1(slres_tdsm.samples[0])
import nibabel as nb
anat = nb.load(pjoin(datapath, 'sub001', 'anatomy', 'highres001.nii.gz'))

from nipy.labs.viz_tools.activation_maps import plot_map
pl.figure(figsize=(15, 4))
sp = pl.subplot(121)
pl.title('Distribution of target similarity structure correlation')
slices = plot_map(
            cut_coords=np.array((12, -42, -20)),
img = pl.gca().get_images()[1]
cax = pl.axes([.05, .05, .05, .9])
pl.colorbar(img, cax=cax)

sp = pl.subplot(122)
pl.ylabel("Number of voxels")
pl.xlabel("Target similarity structure correlation")

See also

The full source code of this example is included in the PyMVPA source distribution (doc/examples/