# encoding: utf-8
"""Module implementing representational similarity analysis (RSA) at the source level.
Kriegeskorte, N., Mur, M., & Bandettini, P. A. (2008). Representational
similarity analysis - connecting the branches of systems neuroscience.
Frontiers in Systems Neuroscience, 2(November), 4.
https://doi.org/10.3389/neuro.06.004.2008
Authors
-------
Marijn van Vliet <marijn.vanvliet@aalto.fi>
Annika Hultén <annika.hulten@aalto.fi>
Ossi Lehtonen <ossi.lehtonen@aalto.fi>
"""
from copy import deepcopy
from warnings import warn
import numpy as np
import mne
from mne.utils import logger, verbose
from scipy.linalg import block_diag
import nibabel as nib
from .rdm import _n_items_from_rdm, rdm_array
from .rsa import rsa_array
from .searchlight import searchlight
from .sensor_level import _tmin_tmax_to_indices, _construct_tmin
[docs]@verbose
def rsa_stcs(
stcs,
rdm_model,
src,
spatial_radius=None,
temporal_radius=None,
stc_rdm_metric="correlation",
stc_rdm_params=dict(),
rsa_metric="spearman",
ignore_nan=False,
y=None,
n_folds=1,
sel_vertices=None,
sel_vertices_by_index=None,
tmin=None,
tmax=None,
n_jobs=1,
verbose=False,
):
"""Perform RSA in a searchlight pattern on MNE-Python source estimates.
The output is a source estimate where the "signal" at each source point is
the RSA, computed for a patch surrounding the source point. Source estimate
objects can be either defined along a cortical surface (``SourceEstimate``
objects) or volumetric (``VolSourceEstimate`` objects). For surface source
estimates, distances between vertices are measured in 2D space, namely as
the length of the path along the surface from one vertex to another. For
volume source estimates, distances are measured in 3D space as a straight
line from one voxel to another.
Parameters
----------
stcs : list of mne.SourceEstimate | list of mne.VolSourceEstimate
For each item, a source estimate for the brain activity.
rdm_model : ndarray, shape (n, n) | (n * (n - 1) // 2,) | list of ndarray
The model RDM, see :func:`compute_rdm`. For efficiency, you can give it
in condensed form, meaning only the upper triangle of the matrix as a
vector. See :func:`scipy.spatial.distance.squareform`. To perform RSA
against multiple models at the same time, supply a list of model RDMs.
Use :func:`compute_rdm` to compute RDMs.
src : instance of mne.SourceSpaces
The source space used by the source estimates specified in the `stcs`
parameter.
spatial_radius : float | None
The spatial radius of the searchlight patch in meters. All source
points within this radius will belong to the searchlight patch. Set to
None to only perform the searchlight over time, flattening across
sensors. Defaults to None.
temporal_radius : float | None
The temporal radius of the searchlight patch in seconds. Set to None to
only perform the searchlight over sensors, flattening across time.
Defaults to None.
stc_rdm_metric : str
The metric to use to compute the RDM for the source estimates. This can
be any metric supported by the scipy.distance.pdist function. See also
the ``stc_rdm_params`` parameter to specify and additional parameter
for the distance function. Defaults to 'correlation'.
stc_rdm_params : dict
Extra arguments for the distance metric used to compute the RDMs.
Refer to :mod:`scipy.spatial.distance` for a list of all other metrics
and their arguments. Defaults to an empty dictionary.
rsa_metric : str
The RSA metric to use to compare the RDMs. Valid options are:
* 'spearman' for Spearman's correlation (the default)
* 'pearson' for Pearson's correlation
* 'kendall-tau-a' for Kendall's Tau (alpha variant)
* 'partial' for partial Pearson correlations
* 'partial-spearman' for partial Spearman correlations
* 'regression' for linear regression weights
Defaults to 'spearman'.
ignore_nan : bool
Whether to treat NaN's as missing values and ignore them when computing
the distance metric. Defaults to ``False``.
.. versionadded:: 0.8
y : ndarray of int, shape (n_items,) | None
For each source estimate, a number indicating the item to which it
belongs. When ``None``, each source estimate is assumed to belong to a
different item. Defaults to ``None``.
n_folds : int | sklearn.model_selection.BaseCrollValidator | None
Number of cross-validation folds to use when computing the distance
metric. Folds are created based on the ``y`` parameter. Specify
``None`` to use the maximum number of folds possible, given the data.
Alternatively, you can pass a Scikit-Learn cross validator object (e.g.
``sklearn.model_selection.KFold``) to assert fine-grained control over
how folds are created.
Defaults to 1 (no cross-validation).
sel_vertices : mne.Label | list of mne.Label | list of ndarray | None
When set, searchlight patches will only be generated for the subset of
ROI labels, or vertices/voxels with the given vertex numbers. When
giving vertex numbers, supply a list of numpy arrays with for each
hemisphere, the selected vertex numbers. For volume source spaces,
supple a list with only a single element, with that element being the
ndarray with vertex numbers.
See also ``sel_vertices_by_index`` for an alternative manner of
selecting vertices.
sel_vertices_by_index : ndarray of int, shape (n_vertices,)
When set, searchlight patches will only be generated for the subset of
vertices with the given indices in the ``stc.data`` array.
See also ``sel_vertices`` for an alternative manner of selecting vertices.
tmin : float | None
When set, searchlight patches will only be generated from subsequent
time points starting from this time point. This value is given in
seconds. Defaults to ``None``, in which case patches are generated
starting from the first time point.
tmax : float | None
When set, searchlight patches will only be generated up to and
including this time point. This value is given in seconds. Defaults to
``None``, in which case patches are generated up to and including the
last time point.
n_jobs : int
The number of processes (=number of CPU cores) to use. Specify -1 to
use all available cores. Defaults to 1.
verbose : bool
Whether to display a progress bar. In order for this to work, you need
the tqdm python module installed. Defaults to False.
Returns
-------
stc : SourceEstimate | VolSourceEstimate | list of SourceEstimate | list of VolSourceEstimate
The correlation values for each searchlight patch. When spatial_radius
is set to None, there will only be one vertex. When temporal_radius is
set to None, there will only be one time point. When multiple models
have been supplied, a list will be returned containing the RSA results
for each model.
See Also
--------
compute_rdm
""" # noqa E501
# Check for compatibility of the source estimates and the model features
one_model = type(rdm_model) is np.ndarray
if one_model:
rdm_model = [rdm_model]
# Check for compatibility of the stcs and the model features
for rdm in rdm_model:
n_items = _n_items_from_rdm(rdm)
if len(stcs) != n_items and y is None:
raise ValueError(
"The number of source estimates (%d) should be equal to the "
"number of items in `rdm_model` (%d). Alternatively, use "
"the `y` parameter to assign source estimates to items."
% (len(stcs), n_items)
)
if y is not None and len(np.unique(y)) != n_items:
raise ValueError(
"The number of items in `rdm_model` (%d) does not match "
"the number of items encoded in the `y` matrix (%d)."
% (n_items, len(np.unique(y)))
)
src = _check_stcs_compatibility(stcs, src)
if spatial_radius is not None:
dist = _get_distance_matrix(src, dist_lim=spatial_radius, n_jobs=n_jobs)
else:
dist = None
if temporal_radius is not None:
# Convert the temporal radius to samples
temporal_radius = int(temporal_radius // stcs[0].tstep)
if temporal_radius < 1:
raise ValueError("Temporal radius is less than one sample.")
samples_from, samples_to = _tmin_tmax_to_indices(stcs[0].times, tmin, tmax)
if sel_vertices_by_index is not None:
sel_series = sel_vertices_by_index
elif sel_vertices is None:
sel_series = np.arange(len(stcs[0].data))
else:
sel_series = vertex_selection_to_indices(stcs[0].vertices, sel_vertices)
logger.info(
f"Performing RSA between SourceEstimates and {len(rdm_model)} model RDM(s)"
)
if spatial_radius is not None:
logger.info(f" Spatial radius: {spatial_radius} meters")
if sel_vertices is not None:
logger.info(f" Using {len(sel_series)} vertices")
else:
logger.info(
f" Using {sum(len(v) for v in stcs[0].vertices)} vertices"
)
if temporal_radius is not None:
logger.info(f" Temporal radius: {temporal_radius} samples")
if tmin is not None or tmax is not None:
logger.info(f" Time interval: {tmin}-{tmax} seconds")
# Perform the RSA
X = np.array([stc.data for stc in stcs])
patches = searchlight(
X.shape,
dist=dist,
spatial_radius=spatial_radius,
temporal_radius=temporal_radius,
sel_series=sel_series,
samples_from=samples_from,
samples_to=samples_to,
)
logger.info(f" Number of searchlight patches: {len(patches)}")
data = rsa_array(
X,
rdm_model,
patches,
data_rdm_metric=stc_rdm_metric,
data_rdm_params=stc_rdm_params,
rsa_metric=rsa_metric,
ignore_nan=ignore_nan,
y=y,
n_folds=n_folds,
n_jobs=n_jobs,
verbose=verbose,
)
# Pack the result in a SourceEstimate object
if spatial_radius is not None:
vertices = vertex_indices_to_numbers(stcs[0].vertices, sel_series)
else:
if src.kind == "volume":
vertices = [np.array([1])]
else:
vertices = [np.array([1]), np.array([])]
data = data[np.newaxis, ...]
tmin = _construct_tmin(stcs[0].times, samples_from, samples_to, temporal_radius)
tstep = stcs[0].tstep
if one_model:
if src.kind == "volume":
return mne.VolSourceEstimate(
data, vertices, tmin, tstep, subject=stcs[0].subject
)
else:
return mne.SourceEstimate(
data, vertices, tmin, tstep, subject=stcs[0].subject
)
else:
if src.kind == "volume":
return [
mne.VolSourceEstimate(
data[..., i], vertices, tmin, tstep, subject=stcs[0].subject
)
for i in range(data.shape[-1])
]
else:
return [
mne.SourceEstimate(
data[..., i], vertices, tmin, tstep, subject=stcs[0].subject
)
for i in range(data.shape[-1])
]
[docs]@verbose
def rdm_stcs(
stcs,
src,
spatial_radius=None,
temporal_radius=None,
dist_metric="sqeuclidean",
dist_params=dict(),
y=None,
n_folds=None,
sel_vertices=None,
sel_vertices_by_index=None,
tmin=None,
tmax=None,
n_jobs=1,
verbose=False,
):
"""Generate RDMs in a searchlight pattern on MNE-Python source estimates.
RDMs are computed using a patch surrounding each source point. Source
estimate objects can be either defined along a cortical surface
(``SourceEstimate`` objects) or volumetric (``VolSourceEstimate`` objects).
For surface source estimates, distances between vertices are measured in 2D
space, namely as the length of the path along the surface from one vertex
to another. For volume source estimates, distances are measured in 3D space
as a straight line from one voxel to another.
Parameters
----------
stcs : list of mne.SourceEstimate | list of mne.VolSourceEstimate
For each item, a source estimate for the brain activity.
src : instance of mne.SourceSpaces
The source space used by the source estimates specified in the `stcs`
parameter.
spatial_radius : float | None
The spatial radius of the searchlight patch in meters. All source
points within this radius will belong to the searchlight patch. Set to
None to only perform the searchlight over time, flattening across
sensors. Defaults to None.
temporal_radius : float | None
The temporal radius of the searchlight patch in seconds. Set to None to
only perform the searchlight over sensors, flattening across time.
Defaults to None.
dist_metric : str
The metric to use to compute the RDM for the source estimates. This can
be any metric supported by the scipy.distance.pdist function. See also
the ``stc_rdm_params`` parameter to specify and additional parameter
for the distance function. Defaults to 'sqeuclidean'.
dist_params : dict
Extra arguments for the distance metric used to compute the RDMs.
Refer to :mod:`scipy.spatial.distance` for a list of all other metrics
and their arguments. Defaults to an empty dictionary.
y : ndarray of int, shape (n_items,) | None
For each source estimate, a number indicating the item to which it
belongs. When ``None``, each source estimate is assumed to belong to a
different item. Defaults to ``None``.
n_folds : int | sklearn.model_selection.BaseCrollValidator | None
Number of cross-validation folds to use when computing the distance
metric. Folds are created based on the ``y`` parameter. Specify
``None`` to use the maximum number of folds possible, given the data.
Alternatively, you can pass a Scikit-Learn cross validator object (e.g.
``sklearn.model_selection.KFold``) to assert fine-grained control over
how folds are created.
Defaults to 1 (no cross-validation).
sel_vertices : mne.Label | list of mne.Label | list of ndarray | None
When set, searchlight patches will only be generated for the subset of
ROI labels, or vertices/voxels with the given vertex numbers. When
giving vertex numbers, supply a list of numpy arrays with for each
hemisphere, the selected vertex numbers. For volume source spaces,
supple a list with only a single element, with that element being the
ndarray with vertex numbers.
See also ``sel_vertices_by_index`` for an alternative manner of
selecting vertices.
sel_vertices_by_index : ndarray of int, shape (n_vertices,)
When set, searchlight patches will only be generated for the subset of
vertices with the given indices in the ``stc.data`` array.
See also ``sel_vertices`` for an alternative manner of selecting vertices.
tmin : float | None
When set, searchlight patches will only be generated from subsequent
time points starting from this time point. This value is given in
seconds. Defaults to ``None``, in which case patches are generated
starting from the first time point.
tmax : float | None
When set, searchlight patches will only be generated up to and
including this time point. This value is given in seconds. Defaults to
``None``, in which case patches are generated up to and including the
last time point.
n_jobs : int
The number of processes (=number of CPU cores) to use for the
source-to-source distance computation. Specify -1 to use all available
cores. Defaults to 1.
verbose : bool
Whether to display a progress bar. In order for this to work, you need
the tqdm python module installed. Defaults to False.
Yields
------
rdm : ndarray, shape (n_items, n_items)
A RDM for each searchlight patch.
"""
src = _check_stcs_compatibility(stcs, src)
if spatial_radius is not None:
dist = _get_distance_matrix(src, dist_lim=spatial_radius, n_jobs=n_jobs)
else:
dist = None
# Convert the temporal radius to samples
if temporal_radius is not None:
temporal_radius = int(temporal_radius // stcs[0].tstep)
if temporal_radius < 1:
raise ValueError("Temporal radius is less than one sample.")
samples_from, samples_to = _tmin_tmax_to_indices(stcs[0].times, tmin, tmax)
if sel_vertices_by_index is not None:
sel_series = sel_vertices_by_index
elif sel_vertices is None:
sel_series = np.arange(len(stcs[0].data))
else:
sel_series = vertex_selection_to_indices(stcs[0].vertices, sel_vertices)
X = np.array([stc.data for stc in stcs])
patches = searchlight(
X.shape,
dist=dist,
spatial_radius=spatial_radius,
temporal_radius=temporal_radius,
sel_series=sel_series,
samples_from=samples_from,
samples_to=samples_to,
)
yield from rdm_array(
X,
patches,
dist_metric=dist_metric,
dist_params=dist_params,
y=y,
n_folds=n_folds,
n_jobs=n_jobs,
)
@verbose
def rsa_stcs_rois(
stcs,
rdm_model,
src,
rois,
temporal_radius=0.1,
stc_rdm_metric="correlation",
stc_rdm_params=dict(),
rsa_metric="spearman",
ignore_nan=False,
y=None,
n_folds=1,
tmin=None,
tmax=None,
n_jobs=1,
verbose=False,
):
"""Perform RSA for a list of ROIs using MNE-Python source estimates.
The output is a source estimate where the "signal" at each source point is
the RSA, computed for a patch surrounding the source point. Source estimate
objects can be either defined along a cortical surface (``SourceEstimate``
objects) or volumetric (``VolSourceEstimate`` objects). For surface source
estimates, distances between vertices are measured in 2D space, namely as
the length of the path along the surface from one vertex to another. For
volume source estimates, distances are measured in 3D space as a straight
line from one voxel to another.
Parameters
----------
stcs : list of mne.SourceEstimate | list of mne.VolSourceEstimate
For each item, a source estimate for the brain activity.
rdm_model : ndarray, shape (n, n) | (n * (n - 1) // 2,) | list of ndarray
The model RDM, see :func:`compute_rdm`. For efficiency, you can give it
in condensed form, meaning only the upper triangle of the matrix as a
vector. See :func:`scipy.spatial.distance.squareform`. To perform RSA
against multiple models at the same time, supply a list of model RDMs.
Use :func:`compute_rdm` to compute RDMs.
src : instance of mne.SourceSpaces
The source space used by the source estimates specified in the `stcs`
parameter.
rois : list of mne.Label
The spatial regions of interest (ROIs) to compute the RSA for. This
needs to be specified as a list of ``mne.Label`` objects, such as
returned by ``mne.read_annotations``.
temporal_radius : float | None
The temporal radius of the searchlight patch in seconds. Set to None to
only perform the searchlight over sensors, flattening across time.
Defaults to 0.1.
stc_rdm_metric : str
The metric to use to compute the RDM for the source estimates. This can
be any metric supported by the scipy.distance.pdist function. See also
the ``stc_rdm_params`` parameter to specify and additional parameter
for the distance function. Defaults to 'correlation'.
stc_rdm_params : dict
Extra arguments for the distance metric used to compute the RDMs.
Refer to :mod:`scipy.spatial.distance` for a list of all other metrics
and their arguments. Defaults to an empty dictionary.
rsa_metric : str
The RSA metric to use to compare the RDMs. Valid options are:
* 'spearman' for Spearman's correlation (the default)
* 'pearson' for Pearson's correlation
* 'kendall-tau-a' for Kendall's Tau (alpha variant)
* 'partial' for partial Pearson correlations
* 'partial-spearman' for partial Spearman correlations
* 'regression' for linear regression weights
Defaults to 'spearman'.
ignore_nan : bool
Whether to treat NaN's as missing values and ignore them when computing
the distance metric. Defaults to ``False``.
.. versionadded:: 0.8
y : ndarray of int, shape (n_items,) | None
For each source estimate, a number indicating the item to which it
belongs. When ``None``, each source estimate is assumed to belong to a
different item. Defaults to ``None``.
n_folds : int | sklearn.model_selection.BaseCrollValidator | None
Number of cross-validation folds to use when computing the distance
metric. Folds are created based on the ``y`` parameter. Specify
``None`` to use the maximum number of folds possible, given the data.
Alternatively, you can pass a Scikit-Learn cross validator object (e.g.
``sklearn.model_selection.KFold``) to assert fine-grained control over
how folds are created.
Defaults to 1 (no cross-validation).
tmin : float | None
When set, searchlight patches will only be generated from subsequent
time points starting from this time point. This value is given in
seconds. Defaults to ``None``, in which case patches are generated
starting from the first time point.
tmax : float | None
When set, searchlight patches will only be generated up to and
including this time point. This value is given in seconds. Defaults to
``None``, in which case patches are generated up to and including the
last time point.
n_jobs : int
The number of processes (=number of CPU cores) to use. Specify -1 to
use all available cores. Defaults to 1.
verbose : bool
Whether to display a progress bar. In order for this to work, you need
the tqdm python module installed. Defaults to False.
Returns
-------
data : ndarray, shape (n_rois, n_times) | list of ndarray
The correlation values for each ROI. When temporal_radius is set to
None, there will be time dimension. When multiple models have been
supplied, a list will be returned containing the RSA results for each
model.
stc : SourceEstimate | list of SourceEstimate
The correlation values for each ROI, backfilled into a full
SourceEstimate object. Each vertex belonging to the same ROI will have
the same values. When temporal_radius is set to None, there will only
be one time point. When multiple models have been supplied, a list will
be returned containing the RSA results for each model.
See Also
--------
compute_rdm
""" # noqa E501
# Check for compatibility of the source estimates and the model features
one_model = type(rdm_model) is np.ndarray
if one_model:
rdm_model = [rdm_model]
# Check for compatibility of the stcs and the model features
for rdm in rdm_model:
n_items = _n_items_from_rdm(rdm)
if len(stcs) != n_items and y is None:
raise ValueError(
"The number of source estimates (%d) should be equal to the "
"number of items in `rdm_model` (%d). Alternatively, use "
"the `y` parameter to assign source estimates to items."
% (len(stcs), n_items)
)
if y is not None and len(np.unique(y)) != n_items:
raise ValueError(
"The number of items in `rdm_model` (%d) does not match "
"the number of items encoded in the `y` matrix (%d)."
% (n_items, len(np.unique(y)))
)
src = _check_stcs_compatibility(stcs, src)
if temporal_radius is not None:
# Convert the temporal radius to samples
temporal_radius = int(temporal_radius // stcs[0].tstep)
if temporal_radius < 1:
raise ValueError("Temporal radius is less than one sample.")
samples_from, samples_to = _tmin_tmax_to_indices(stcs[0].times, tmin, tmax)
# Convert the labels to data indices
roi_inds = [vertex_selection_to_indices(stcs[0].vertices, roi) for roi in rois]
# Perform the RSA
X = np.array([stc.data for stc in stcs])
patches = searchlight(
X.shape,
spatial_radius=roi_inds,
temporal_radius=temporal_radius,
samples_from=samples_from,
samples_to=samples_to,
)
data = rsa_array(
X,
rdm_model,
patches,
data_rdm_metric=stc_rdm_metric,
data_rdm_params=stc_rdm_params,
rsa_metric=rsa_metric,
ignore_nan=ignore_nan,
y=y,
n_folds=n_folds,
n_jobs=n_jobs,
verbose=verbose,
)
# Pack the result in SourceEstimate objects
subject = stcs[0].subject
tmin = _construct_tmin(stcs[0].times, samples_from, samples_to, temporal_radius)
tstep = stcs[0].tstep
if one_model:
stc = backfill_stc_from_rois(
data, rois, src, tmin=tmin, tstep=tstep, subject=subject
)
else:
stc = [
backfill_stc_from_rois(
data[..., i], rois, src, tmin=tmin, tstep=tstep, subject=subject
)
for i in range(data.shape[-1])
]
return data, stc
[docs]@verbose
def rsa_nifti(
image,
rdm_model,
spatial_radius=0.01,
image_rdm_metric="correlation",
image_rdm_params=dict(),
rsa_metric="spearman",
ignore_nan=False,
y=None,
n_folds=1,
roi_mask=None,
brain_mask=None,
n_jobs=1,
verbose=False,
):
"""Perform RSA in a searchlight pattern on Nibabel Nifti-like images.
The output is a 3D Nifti image where the data at each voxel is is
the RSA, computed for a patch surrounding the voxel.
.. versionadded:: 0.4
Parameters
----------
image : 4D Nifti-like image
The Nitfi image data. The 4th dimension must contain the images
for each item.
rdm_model : ndarray, shape (n, n) | (n * (n - 1) // 2,) | list of ndarray
The model RDM, see :func:`compute_rdm`. For efficiency, you can give it
in condensed form, meaning only the upper triangle of the matrix as a
vector. See :func:`scipy.spatial.distance.squareform`. To perform RSA
against multiple models at the same time, supply a list of model RDMs.
Use :func:`compute_rdm` to compute RDMs.
spatial_radius : float
The spatial radius of the searchlight patch in meters. All source
points within this radius will belong to the searchlight patch.
Defaults to 0.01.
image_rdm_metric : str
The metric to use to compute the RDM for the data. This can be
any metric supported by the scipy.distance.pdist function. See also the
``image_rdm_params`` parameter to specify and additional parameter for
the distance function. Defaults to 'correlation'.
image_rdm_params : dict
Extra arguments for the distance metric used to compute the RDMs.
Refer to :mod:`scipy.spatial.distance` for a list of all other metrics
and their arguments. Defaults to an empty dictionary.
rsa_metric : str
The RSA metric to use to compare the RDMs. Valid options are:
* 'spearman' for Spearman's correlation (the default)
* 'pearson' for Pearson's correlation
* 'kendall-tau-a' for Kendall's Tau (alpha variant)
* 'partial' for partial Pearson correlations
* 'partial-spearman' for partial Spearman correlations
* 'regression' for linear regression weights
Defaults to 'spearman'.
ignore_nan : bool
Whether to treat NaN's as missing values and ignore them when computing
the distance metric. Defaults to ``False``.
.. versionadded:: 0.8
y : ndarray of int, shape (n_items,) | None
For each source estimate, a number indicating the item to which it
belongs. When ``None``, each source estimate is assumed to belong to a
different item. Defaults to ``None``.
n_folds : int | sklearn.model_selection.BaseCrollValidator | None
Number of cross-validation folds to use when computing the distance
metric. Folds are created based on the ``y`` parameter. Specify
``None`` to use the maximum number of folds possible, given the data.
Alternatively, you can pass a Scikit-Learn cross validator object (e.g.
``sklearn.model_selection.KFold``) to assert fine-grained control over
how folds are created.
Defaults to 1 (no cross-validation).
roi_mask : 3D Nifti-like image | None
When set, searchlight patches will only be generated for the subset of
voxels with non-zero values in the given mask. This is useful for
restricting the analysis to a region of interest (ROI). Note that while
the center of the patches are all within the ROI, the patch itself may
extend beyond the ROI boundaries.
Defaults to ``None``, in which case patches for all voxels are
generated.
brain_mask : 3D Nifti-like image | None
When set, searchlight patches are restricted to only contain voxels
with non-zero values in the given mask. This is useful for make sure
only information from inside the brain is used. In contrast to the
`roi_mask`, searchlight patches will not use data outside of this mask.
Defaults to ``None``, in which case all voxels are included in the
analysis.
n_jobs : int
The number of processes (=number of CPU cores) to use. Specify -1 to
use all available cores. Defaults to 1.
verbose : bool
Whether to display a progress bar. In order for this to work, you need
the tqdm python module installed. Defaults to False.
Returns
-------
rsa_results : 3D Nifti1Image | list of 3D Nifti1Image
The correlation values for each searchlight patch. When multiple models
have been supplied, a list will be returned containing the RSA results
for each model.
See Also
--------
compute_rdm
"""
# Check for compatibility of the source estimates and the model features
one_model = type(rdm_model) is np.ndarray
if one_model:
rdm_model = [rdm_model]
if (
not isinstance(image, tuple(nib.imageclasses.all_image_classes))
or image.ndim != 4
):
raise ValueError("The image data must be 4-dimensional Nifti-like " "images")
# Check for compatibility of the BOLD images and the model features
for rdm in rdm_model:
n_items = _n_items_from_rdm(rdm)
if image.shape[3] != n_items and y is None:
raise ValueError(
"The number of images (%d) should be equal to the "
"number of items in `rdm_model` (%d). Alternatively, use "
"the `y` parameter to assign evokeds to items."
% (image.shape[3], n_items)
)
if y is not None and len(np.unique(y)) != n_items:
raise ValueError(
"The number of items in `rdm_model` (%d) does not match "
"the number of items encoded in the `y` matrix (%d)."
% (n_items, len(np.unique(y)))
)
# Get data as (n_items x n_voxels)
X = image.get_fdata().reshape(-1, image.shape[3]).T
# Find voxel positions
voxels = np.array(list(np.ndindex(image.shape[:-1])))
voxel_loc = voxels @ image.affine[:3, :3]
voxel_loc /= 1000 # convert position from mm to meters
# Apply masks
result_mask = np.ones(image.shape[:3], dtype=bool)
if brain_mask is not None:
if brain_mask.ndim != 3 or brain_mask.shape != image.shape[:3]:
raise ValueError(
"Brain mask must be a 3-dimensional Nifi-like "
"image with the same dimensions as the data "
"image"
)
brain_mask = brain_mask.get_fdata() != 0
result_mask &= brain_mask
brain_mask = brain_mask.ravel()
X = X[:, brain_mask]
voxel_loc = voxel_loc[brain_mask]
if roi_mask is not None:
if roi_mask.ndim != 3 or roi_mask.shape != image.shape[:3]:
raise ValueError(
"ROI mask must be a 3-dimensional Nifi-like "
"image with the same dimensions as the data "
"image"
)
roi_mask = roi_mask.get_fdata() != 0
result_mask &= roi_mask
roi_mask = roi_mask.ravel()
if brain_mask is not None:
roi_mask = roi_mask[brain_mask]
roi_mask = np.flatnonzero(roi_mask)
# Compute distances between voxels
logger.info("Computing distances...")
from sklearn.neighbors import NearestNeighbors
nn = NearestNeighbors(radius=spatial_radius, n_jobs=n_jobs).fit(voxel_loc)
dist = nn.radius_neighbors_graph(mode="distance")
# Perform the RSA
patches = searchlight(
X.shape,
dist=dist,
spatial_radius=spatial_radius,
temporal_radius=None,
sel_series=roi_mask,
)
rsa_result = rsa_array(
X,
rdm_model,
patches,
data_rdm_metric=image_rdm_metric,
data_rdm_params=image_rdm_params,
rsa_metric=rsa_metric,
ignore_nan=ignore_nan,
y=y,
n_folds=n_folds,
n_jobs=n_jobs,
verbose=verbose,
)
if one_model:
data = np.zeros(image.shape[:3])
data[result_mask] = rsa_result
return nib.Nifti1Image(data, image.affine, image.header)
else:
results = []
for i in range(rsa_result.shape[-1]):
data = np.zeros(image.shape[:3])
data[result_mask] = rsa_result[:, i]
results.append(nib.Nifti1Image(data, image.affine, image.header))
return results
[docs]@verbose
def rdm_nifti(
image,
spatial_radius=0.01,
dist_metric="correlation",
dist_params=dict(),
y=None,
n_folds=1,
roi_mask=None,
brain_mask=None,
n_jobs=1,
verbose=False,
):
"""Generate RDMs in a searchlight pattern on Nibabel Nifty-like images.
RDMs are computed using a patch surrounding each voxel.
.. versionadded:: 0.4
Parameters
----------
image : 4D Nifti-like image
The Nitfi image data. The 4th dimension must contain the images
for each item.
spatial_radius : float
The spatial radius of the searchlight patch in meters. All source
points within this radius will belong to the searchlight patch.
Defaults to 0.01.
dist_metric : str
The metric to use to compute the RDM for the data. This can be
any metric supported by the scipy.distance.pdist function. See also the
``dist_params`` parameter to specify and additional parameter for
the distance function. Defaults to 'correlation'.
dist_params : dict
Extra arguments for the distance metric used to compute the RDMs.
Refer to :mod:`scipy.spatial.distance` for a list of all other metrics
and their arguments. Defaults to an empty dictionary.
y : ndarray of int, shape (n_items,) | None
For each source estimate, a number indicating the item to which it
belongs. When ``None``, each source estimate is assumed to belong to a
different item. Defaults to ``None``.
n_folds : int | sklearn.model_selection.BaseCrollValidator | None
Number of cross-validation folds to use when computing the distance
metric. Folds are created based on the ``y`` parameter. Specify
``None`` to use the maximum number of folds possible, given the data.
Alternatively, you can pass a Scikit-Learn cross validator object (e.g.
``sklearn.model_selection.KFold``) to assert fine-grained control over
how folds are created.
Defaults to 1 (no cross-validation).
roi_mask : 3D Nifti-like image | None
When set, searchlight patches will only be generated for the subset of
voxels with non-zero values in the given mask. This is useful for
restricting the analysis to a region of interest (ROI). Note that while
the center of the patches are all within the ROI, the patch itself may
extend beyond the ROI boundaries.
Defaults to ``None``, in which case patches for all voxels are
generated.
brain_mask : 3D Nifti-like image | None
When set, searchlight patches are restricted to only contain voxels
with non-zero values in the given mask. This is useful for make sure
only information from inside the brain is used. In contrast to the
`roi_mask`, searchlight patches will not use data outside of this mask.
Defaults to ``None``, in which case all voxels are included in the
analysis.
n_jobs : int
The number of processes (=number of CPU cores) to use. Specify -1 to
use all available cores. Defaults to 1.
verbose : bool
Whether to display a progress bar. In order for this to work, you need
the tqdm python module installed. Defaults to False.
Yields
------
rdm : ndarray, shape (n_items, n_items)
A RDM for each searchlight patch.
"""
if (
not isinstance(image, tuple(nib.imageclasses.all_image_classes))
or image.ndim != 4
):
raise ValueError("The image data must be 4-dimensional Nifti-like " "images")
# Get data as (n_items x n_voxels)
X = image.get_fdata().reshape(-1, image.shape[3]).T
# Find voxel positions
voxels = np.array(list(np.ndindex(image.shape[:-1])))
voxel_loc = voxels @ image.affine[:3, :3]
voxel_loc /= 1000 # convert position from mm to meters
# Apply masks
result_mask = np.ones(image.shape[:3], dtype=bool)
if brain_mask is not None:
if brain_mask.ndim != 3 or brain_mask.shape != image.shape[:3]:
raise ValueError(
"Brain mask must be a 3-dimensional Nifi-like "
"image with the same dimensions as the data "
"image"
)
brain_mask = brain_mask.get_fdata() != 0
result_mask &= brain_mask
brain_mask = brain_mask.ravel()
X = X[:, brain_mask]
voxel_loc = voxel_loc[brain_mask]
if roi_mask is not None:
if roi_mask.ndim != 3 or roi_mask.shape != image.shape[:3]:
raise ValueError(
"ROI mask must be a 3-dimensional Nifi-like "
"image with the same dimensions as the data "
"image"
)
roi_mask = roi_mask.get_fdata() != 0
result_mask &= roi_mask
roi_mask = roi_mask.ravel()
if brain_mask is not None:
roi_mask = roi_mask[brain_mask]
roi_mask = np.flatnonzero(roi_mask)
# Compute distances between voxels
logger.info("Computing distances...")
from sklearn.neighbors import NearestNeighbors
nn = NearestNeighbors(radius=spatial_radius, n_jobs=n_jobs).fit(voxel_loc)
dist = nn.radius_neighbors_graph(mode="distance")
# Compute RDMs
patches = searchlight(
X.shape,
dist=dist,
spatial_radius=spatial_radius,
temporal_radius=None,
sel_series=roi_mask,
)
yield from rdm_array(
X,
patches,
dist_metric=dist_metric,
dist_params=dist_params,
y=y,
n_folds=n_folds,
n_jobs=n_jobs,
)
def _check_stcs_compatibility(stcs, src):
"""Check for compatibility of the source estimates and source space."""
if src.kind == "volume" and not isinstance(stcs[0], mne.VolSourceEstimate):
raise ValueError(
"Volume source estimates provided, but not a volume source space "
f"(src.kind={src.kind})."
)
if src.kind == "volume":
if np.any(stcs[0].vertices != src[0]["vertno"]):
src = _restrict_src_to_vertices(src, stcs[0].vertices)
for stc in stcs:
if np.any(stc.vertices != src[0]["vertno"]):
raise ValueError("Not all source estimates have the same vertices.")
else:
for src_hemi, stc_hemi_vertno in zip(src, stcs[0].vertices):
if np.any(stc_hemi_vertno != src_hemi["vertno"]):
src = _restrict_src_to_vertices(src, stcs[0].vertices)
for stc in stcs:
for src_hemi, stc_hemi_vertno in zip(src, stcs[0].vertices):
if np.any(stc_hemi_vertno != src_hemi["vertno"]):
raise ValueError("Not all source estimates have the same vertices.")
times = stcs[0].times
for stc in stcs:
if np.any(stc.times != times):
raise ValueError("Not all source estimates have the same time points.")
return src
def _get_distance_matrix(src, dist_lim, n_jobs=1):
"""Get vertex-to-vertex distance matrix from source space.
During inverse computation, the source space was downsampled (i.e. using
ico4). Construct vertex-to-vertex distance matrices using only the
vertices that are defined in the source solution.
Parameters
----------
src : mne.SourceSpaces
The source space to get the distance matrix for.
dist_lim : float
Maximum distance required. We don't care about distances beyond this
maximum.
n_jobs : int
Number of CPU cores to use if distance computation is necessary.
Defaults to 1.
Returns
-------
dist : ndarray (n_vertices, n_vertices)
The vertex-to-vertex distance matrix.
"""
dist = []
# Check if distances have been pre-computed in the given source space. Give
# a warning if the pre-computed distances may have had a too limited
# dist_lim setting.
needs_distance_computation = False
for hemi in src:
if "dist" not in hemi or hemi["dist"] is None:
needs_distance_computation = True
else:
if hemi["dist_limit"] != np.inf and hemi["dist_limit"][0] < dist_lim:
warn(
f"Source space has pre-computed distances, but all "
f"distances are smaller than the searchlight radius "
f"({dist_lim}). You may want to consider recomputing "
f"the source space distances using the "
f"mne.add_source_space_distances function."
)
if needs_distance_computation:
if dist_lim is None:
dist_lim = np.inf
if src.kind == "volume":
src = _add_volume_source_space_distances(src, dist_lim)
else:
src = mne.add_source_space_distances(src, dist_lim, n_jobs=n_jobs)
for hemi in src:
inuse = np.flatnonzero(hemi["inuse"])
dist.append(hemi["dist"][np.ix_(inuse, inuse)].toarray())
# Collect the distances in a single matrix
dist = block_diag(*dist)
dist[dist == 0] = np.inf # Across hemisphere distance is infinity
dist.flat[:: dist.shape[0] + 1] = 0 # Distance to yourself is zero
return dist
def _add_volume_source_space_distances(src, dist_limit):
"""Compute the distance between voxels in a volume source space.
Operates in-place!
Code is mostly taken from `mne.add_source_space_distances`.
Parameters
----------
src : instance of mne.SourceSpaces
The volume source space to compute the voxel-wise distances for.
dist_limit : float
The maximum distance (in meters) to consider. Voxels that are further
apart than this distance will have a distance of infinity. Use this to
reduce computation time.
Returns
-------
src : instance of mne.SourceSpaces
The volume source space, now with the 'dist' and 'dist_limit' fields
set.
"""
# Lazy import to not have to load the huge scipy module every time mne_rsa
# get's loaded.
from scipy.sparse import csr_matrix
assert src.kind == "volume"
n_sources = src[0]["np"]
neighbors = np.array(src[0]["neighbor_vert"])
rows, cols = np.nonzero(neighbors != -1)
cols = neighbors[(rows, cols)]
dist = np.linalg.norm(src[0]["rr"][rows, :] - src[0]["rr"][cols, :], axis=1)
con_matrix = csr_matrix((dist, (rows, cols)), shape=(n_sources, n_sources))
dist = mne.source_space._source_space._do_src_distances(
con_matrix, src[0]["vertno"], np.arange(src[0]["nuse"]), dist_limit
)[0]
d = dist.ravel() # already float32
idx = d > 0
d = d[idx]
i, j = np.meshgrid(src[0]["vertno"], src[0]["vertno"])
i = i.ravel()[idx]
j = j.ravel()[idx]
src[0]["dist"] = csr_matrix((d, (i, j)), shape=(n_sources, n_sources))
src[0]["dist_limit"] = np.array([dist_limit], "float32")
return src
def make_mri_con_matrix(img):
from scipy.sparse import csr_matrix
# Create 3 x 3 x 3 cube of (ijk) indices, centered around (0, 0, 0)
cube = np.array(list(np.ndindex(3, 3, 3))) - [1, 1, 1]
# Remove center of the cube
cube = np.delete(cube, len(cube) // 2, axis=0)
# Compute distance from all points in the cube to the center
dist = np.linalg.norm(cube @ img.affine[:3, :3], axis=1)
# Copy the cube, centering it around each voxel
voxels = np.array(list(np.ndindex(*img.shape[:3])))
neighbours = voxels[:, :, np.newaxis] + cube.T[np.newaxis, :, :]
assert neighbours.shape == (len(voxels), 3, len(cube))
# Convert ijk coordinates to voxel numbers
neighbours = np.ravel_multi_index(
(neighbours[:, 0, :], neighbours[:, 1, :], neighbours[:, 2, :]),
img.shape[:3],
mode="clip",
)
rows = np.repeat(np.arange(neighbours.shape[0]), neighbours.shape[1])
cols = neighbours.ravel()
dist = np.tile(dist, neighbours.shape[0])
con_matrix = csr_matrix((dist, (rows, cols)), shape=(len(voxels), len(voxels)))
return con_matrix
def backfill_stc_from_rois(values, rois, src, tmin=0, tstep=1, subject=None):
"""Backfill the ROI values into a full mne.SourceEstimate object.
Each vertex belonging to the same region of interest (ROI) will have the
sample value.
Parameters
----------
values : ndarray, shape (n_rois, ...)
For each ROI, either a single value or a timecourse of values.
rois : list of mne.Label
The spatial regions of interest (ROIs) to compute the RSA for. This
needs to be specified as a list of ``mne.Label`` objects, such as
returned by ``mne.read_annotations``.
src : instance of mne.SourceSpaces
The source space used by the source estimates specified in the `stcs`
parameter.
tmin : float
Time corrsponding to the first sample.
tstep : float
Difference in time between two samples.
subject : str | None
The name of the FreeSurfer subject.
Returns
-------
stc : mne.SourceEstimate
The backfilled source estimate object.
"""
values = np.asarray(values)
if values.ndim == 1:
n_samples = 1
else:
n_samples = values.shape[1]
data = np.zeros((src[0]["nuse"] + src[1]["nuse"], n_samples))
verts_lh = src[0]["vertno"]
verts_rh = src[1]["vertno"]
for roi, rsa_timecourse in zip(rois, values):
roi = roi.copy().restrict(src)
if roi.hemi == "lh":
roi_ind = np.searchsorted(verts_lh, roi.vertices)
else:
roi_ind = np.searchsorted(verts_rh, roi.vertices)
roi_ind += src[0]["nuse"]
for ind in roi_ind:
data[ind] = rsa_timecourse
return mne.SourceEstimate(
data, vertices=[verts_lh, verts_rh], tmin=tmin, tstep=tstep, subject=subject
)
def vertex_selection_to_indices(vertno, sel_vertices):
"""Unify across different ways of selecting vertices."""
if isinstance(sel_vertices, mne.Label):
sel_vertices = [sel_vertices]
if not isinstance(sel_vertices, list):
raise ValueError(
"Invalid type for sel_vertices. It should be an mne.Label, a list of "
f"mne.Label's or a list of numpy arrays, but {type(sel_vertices)} was "
"provided."
)
if len(sel_vertices) == 0:
raise ValueError("Empty list provided for sel_vertices.")
# Deal with mne.Label objects
if isinstance(sel_vertices[0], mne.Label):
labels = sel_vertices
sel_vertices = [[] for _ in range(len(vertno))]
for label in labels:
if label.hemi == "lh":
sel_vertices[0].extend(label.get_vertices_used(vertno[0]).tolist())
else:
if len(vertno) == 1:
raise ValueError(
"one of the Label's provided for sel_vertices defines vertices "
"on the right hemisphere. However, the provided "
"SourceEstimate's only have vertices in the left hemisphere."
)
sel_vertices[1].extend(label.get_vertices_used(vertno[1]).tolist())
# At this point, sel_vertices should be an array of vertex numbers.
# Convert them to indices for the .data array.
if len(sel_vertices) != len(vertno):
raise ValueError(
f"sel_vertices should be a list with {len(vertno)} elements: one for "
"each hemisphere (volume source spaces have one hemisphere.)"
)
sel_series = []
for hemi, (hemi_vertno, hemi_sel_vertno) in enumerate(zip(vertno, sel_vertices)):
if len(hemi_sel_vertno) == 0:
continue
sel_inds = np.searchsorted(hemi_vertno, hemi_sel_vertno)
if not np.array_equal(hemi_vertno[sel_inds], hemi_sel_vertno):
raise ValueError("Some selected vertices are not present in the data.")
if hemi > 0:
sel_inds += len(vertno[hemi - 1])
sel_series.append(sel_inds)
return np.unique(np.hstack(sel_series))
def vertex_indices_to_numbers(vertno, vert_ind):
"""Convert vertex indices to vertex numbers."""
min_vert_ind = 0
sel_vert_no = [[] for _ in vertno]
for hemi, hemi_vertno in enumerate(vertno):
if hemi > 0:
min_vert_ind += len(vertno[hemi - 1])
max_vert_ind = min_vert_ind + len(hemi_vertno)
hemi_vert_ind = vert_ind[(vert_ind >= min_vert_ind) & (vert_ind < max_vert_ind)]
hemi_vert_ind -= min_vert_ind
sel_vert_no[hemi] = hemi_vertno[hemi_vert_ind]
return sel_vert_no
@verbose
def _restrict_src_to_vertices(src, vertno, verbose=None):
"""Restrict a source space to the given vertices.
Parameters
----------
src: instance of SourceSpaces
The source space to be restricted.
vertno : tuple of lists (vertno_lh, vertno_rh)
For each hemisphere, the vertex numbers to keep.
verbose : bool | str | int | None
If not None, override default verbose level (see :func:`mne.verbose`
and :ref:`Logging documentation <tut_logging>` for more).
Returns
-------
src_out : instance of SourceSpaces
The restricted source space.
"""
src_out = deepcopy(src)
vert_no_lh, vert_no_rh = vertno
if not (
np.all(np.in1d(vert_no_lh, src[0]["vertno"]))
and np.all(np.in1d(vert_no_rh, src[1]["vertno"]))
):
raise ValueError("One or more vertices were not present in the source space.")
logger.info(
"Restricting source space to {n_keep} out of {n_total} vertices.".format(
n_keep=len(vert_no_lh) + len(vert_no_rh),
n_total=src[0]["nuse"] + src[1]["nuse"],
)
)
for hemi, verts in zip(src_out, (vert_no_lh, vert_no_rh)):
# Ensure vertices are in sequential order
verts = np.sort(verts)
# Restrict the source space
hemi["vertno"] = verts
hemi["nuse"] = len(verts)
hemi["inuse"] = hemi["inuse"].copy()
hemi["inuse"].fill(0)
if hemi["nuse"] > 0: # Don't use empty array as index
hemi["inuse"][verts] = 1
hemi["use_tris"] = np.array([[]], int)
hemi["nuse_tri"] = np.array([0])
return src_out