def perm_prfx(domain,
graphs,
features,
nb_parcel,
ldata,
initial_mask=None,
nb_perm=100,
niter=5,
dmax=10.,
lamb=100.0,
chunksize=1.e5,
verbose=1):
"""
caveat: assumes that the functional dimension is 1
"""
from ..utils.reproducibility_measures import ttest
# permutations for the assesment of the results
prfx0 = []
adim = domain.coord.shape[1]
nb_subj = len(ldata)
for q in range(nb_perm):
feature = []
sldata = []
for s in range(nb_subj):
lf = features[s].copy()
swap = (rand() > 0.5) * 2 - 1
lf[:, 0:-adim] = swap * lf[:, 0:-adim]
sldata.append(swap * ldata[s])
feature.append(lf)
# optimization part
all_labels, proto_anat = _optim_hparcel(feature,
domain,
graphs,
nb_parcel,
lamb,
dmax,
niter,
initial_mask,
chunksize=chunksize)
labels = -np.ones((domain.size, nb_subj)).astype(np.int)
for s in range(nb_subj):
labels[initial_mask[:, s] > -1, s] = all_labels[s]
# compute the group-level labels
template_labels = voronoi(domain.coord, proto_anat)
# create the parcellation
pcl = MultiSubjectParcellation(domain,
individual_labels=labels,
template_labels=template_labels)
pdata = pcl.make_feature('functional',
np.rollaxis(np.array(ldata), 1, 0))
prfx = ttest(np.squeeze(pdata))
if verbose:
print q, prfx.max(0)
prfx0.append(prfx.max(0))
return prfx0
def perm_prfx(domain, graphs, features, nb_parcel, ldata, initial_mask=None,
nb_perm=100, niter=5, dmax=10., lamb=100.0, chunksize=1.e5,
verbose=1):
"""
caveat: assumes that the functional dimension is 1
"""
from ..utils.reproducibility_measures import ttest
# permutations for the assesment of the results
prfx0 = []
adim = domain.coord.shape[1]
nb_subj = len(ldata)
for q in range(nb_perm):
feature = []
sldata = []
for s in range(nb_subj):
lf = features[s].copy()
swap = (rand() > 0.5) * 2 - 1
lf[:, 0:-adim] = swap * lf[:, 0:-adim]
sldata.append(swap * ldata[s])
feature.append(lf)
# optimization part
all_labels, proto_anat = _optim_hparcel(
feature, domain, graphs, nb_parcel, lamb, dmax, niter,
initial_mask, chunksize=chunksize)
labels = - np.ones((domain.size, nb_subj)).astype(np.int)
for s in range(nb_subj):
labels[initial_mask[:, s] > -1, s] = all_labels[s]
# compute the group-level labels
template_labels = voronoi(domain.coord, proto_anat)
# create the parcellation
pcl = MultiSubjectParcellation(domain, individual_labels=labels,
template_labels=template_labels)
pdata = pcl.make_feature('functional',
np.rollaxis(np.array(ldata), 1, 0))
prfx = ttest(np.squeeze(pdata))
if verbose:
print q, prfx.max(0)
prfx0.append(prfx.max(0))
return prfx0
def hparcel(domain, ldata, nb_parcel, nb_perm=0, niter=5, mu=10., dmax=10.,
lamb=100.0, chunksize=1.e5, verbose=0, initial_mask=None):
"""
Function that performs the parcellation by optimizing the
inter-subject similarity while retaining the connectedness
within subject and some consistency across subjects.
Parameters
----------
domain: discrete_domain.DiscreteDomain instance,
yields all the spatial information on the parcelled domain
ldata: list of (n_subj) arrays of shape (domain.size, dim)
the feature data used to inform the parcellation
nb_parcel: int,
the number of parcels
nb_perm: int, optional,
the number of times the parcellation and prfx
computation is performed on sign-swaped data
niter: int, optional,
number of iterations to obtain the convergence of the method
information in the clustering algorithm
mu: float, optional,
relative weight of anatomical information
dmax: float optional,
radius of allowed deformations
lamb: float optional
parameter to control the relative importance of space vs function
chunksize; int, optional
number of points used in internal sub-sampling
verbose: bool, optional,
verbosity mode
initial_mask: array of shape (domain.size, nb_subj), optional
initial subject-depedent masking of the domain
Results
-------
Pa: the resulting parcellation structure appended with the labelling
"""
# a various parameters
nbvox = domain.size
nb_subj = len(ldata)
if initial_mask is None:
initial_mask = np.ones((nbvox, nb_subj), np.int)
graphs = []
feature = []
for s in range(nb_subj):
# build subject-specific models of the data
lnvox = np.sum(initial_mask[:, s] > - 1)
lac = domain.coord[initial_mask[:, s] > - 1]
beta = np.reshape(ldata[s], (lnvox, ldata[s].shape[1]))
lf = np.hstack((beta, mu * lac / (1.e-15 + np.std(domain.coord, 0))))
feature.append(lf)
g = wgraph_from_coo_matrix(domain.topology)
g.remove_trivial_edges()
graphs.append(g)
# main function
all_labels, proto_anat = _optim_hparcel(
feature, domain, graphs, nb_parcel, lamb, dmax, niter, initial_mask,
chunksize=chunksize, verbose=verbose)
# write the individual labelling
labels = - np.ones((nbvox, nb_subj)).astype(np.int)
for s in range(nb_subj):
labels[initial_mask[:, s] > -1, s] = all_labels[s]
# compute the group-level labels
template_labels = voronoi(domain.coord, proto_anat)
# create the parcellation
pcl = MultiSubjectParcellation(domain, individual_labels=labels,
template_labels=template_labels,
nb_parcel=nb_parcel)
pcl.make_feature('functional', np.rollaxis(np.array(ldata), 1, 0))
if nb_perm > 0:
prfx0 = perm_prfx(domain, graphs, feature, nb_parcel, ldata,
initial_mask, nb_perm, niter, dmax, lamb, chunksize)
return pcl, prfx0
else:
return pcl
def hparcel(domain,
ldata,
nb_parcel,
nb_perm=0,
niter=5,
mu=10.,
dmax=10.,
lamb=100.0,
chunksize=1.e5,
verbose=0,
initial_mask=None):
"""
Function that performs the parcellation by optimizing the
inter-subject similarity while retaining the connectedness
within subject and some consistency across subjects.
Parameters
----------
domain: discrete_domain.DiscreteDomain instance,
yields all the spatial information on the parcelled domain
ldata: list of (n_subj) arrays of shape (domain.size, dim)
the feature data used to inform the parcellation
nb_parcel: int,
the number of parcels
nb_perm: int, optional,
the number of times the parcellation and prfx
computation is performed on sign-swaped data
niter: int, optional,
number of iterations to obtain the convergence of the method
information in the clustering algorithm
mu: float, optional,
relative weight of anatomical information
dmax: float optional,
radius of allowed deformations
lamb: float optional
parameter to control the relative importance of space vs function
chunksize; int, optional
number of points used in internal sub-sampling
verbose: bool, optional,
verbosity mode
initial_mask: array of shape (domain.size, nb_subj), optional
initial subject-depedent masking of the domain
Returns
-------
Pa: the resulting parcellation structure appended with the labelling
"""
# a various parameters
nbvox = domain.size
nb_subj = len(ldata)
if initial_mask is None:
initial_mask = np.ones((nbvox, nb_subj), np.int)
graphs = []
feature = []
for s in range(nb_subj):
# build subject-specific models of the data
lnvox = np.sum(initial_mask[:, s] > -1)
lac = domain.coord[initial_mask[:, s] > -1]
beta = np.reshape(ldata[s], (lnvox, ldata[s].shape[1]))
lf = np.hstack((beta, mu * lac / (1.e-15 + np.std(domain.coord, 0))))
feature.append(lf)
g = wgraph_from_coo_matrix(domain.topology)
g.remove_trivial_edges()
graphs.append(g)
# main function
all_labels, proto_anat = _optim_hparcel(feature,
domain,
graphs,
nb_parcel,
lamb,
dmax,
niter,
initial_mask,
chunksize=chunksize,
verbose=verbose)
# write the individual labelling
labels = -np.ones((nbvox, nb_subj)).astype(np.int)
for s in range(nb_subj):
labels[initial_mask[:, s] > -1, s] = all_labels[s]
# compute the group-level labels
template_labels = voronoi(domain.coord, proto_anat)
# create the parcellation
pcl = MultiSubjectParcellation(domain,
individual_labels=labels,
template_labels=template_labels,
nb_parcel=nb_parcel)
pcl.make_feature('functional', np.rollaxis(np.array(ldata), 1, 0))
if nb_perm > 0:
prfx0 = perm_prfx(domain, graphs, feature, nb_parcel, ldata,
initial_mask, nb_perm, niter, dmax, lamb, chunksize)
return pcl, prfx0
else:
return pcl
