def make_bsa_2d(betas, theta=3., dmax=5., ths=0, thq=0.5, smin=0,
nbeta=[0],method='simple'):
""" Function for performing bayesian structural analysis on a set of images.
"""
ref_dim = np.shape(betas[0])
nbsubj = betas.shape[0]
xyz = np.array(np.where(betas[:1])).T
nvox = np.size(xyz, 0)
# create the field strcture that encodes image topology
Fbeta = ff.Field(nvox)
Fbeta.from_3d_grid(xyz.astype(np.int), 18)
# Get coordinates in mm
tal = xyz.astype(np.float)
# get the functional information
lbeta = np.array([np.ravel(betas[k]) for k in range(nbsubj)]).T
# the voxel volume is 1.0
g0 = 1.0/(1.0*nvox)
bdensity = 1
affine = np.eye(4)
shape = (1,ref_dim[0],ref_dim[1])
if method=='ipmi':
group_map, AF, BF, likelihood = \
bsa.compute_BSA_ipmi(Fbeta, lbeta, tal, dmax,xyz, affine,
shape, thq,
smin, ths, theta, g0, bdensity)
if method=='simple':
group_map, AF, BF, likelihood = \
bsa.compute_BSA_simple(Fbeta, lbeta, tal, dmax,xyz,
affine, shape, thq, smin, ths,
theta, g0)
if method=='dev':
group_map, AF, BF, likelihood = \
bsa.compute_BSA_dev(Fbeta, lbeta, tal, dmax, xyz, affine,
shape, thq,
smin, ths, theta, g0, bdensity)
if method=='sbf':
pval = 0.2
group_map, AF, BF = sbf.Compute_Amers (Fbeta, lbeta, xyz, affine,
shape,
tal, dmax, theta, ths ,pval)
return AF, BF
def make_bsa_2d(betas, theta=3., dmax=5., ths=0, thq=0.5, smin=0,
nbeta=[0], method='simple'):
"""
Function for performing bayesian structural analysis on a set of images.
Fixme: 'quick' is not tested
"""
ref_dim = np.shape(betas[0])
nbsubj = betas.shape[0]
xyz = np.array(np.where(betas[:1])).T
nvox = np.size(xyz, 0)
# create the field strcture that encodes image topology
Fbeta = ff.Field(nvox)
Fbeta.from_3d_grid(xyz.astype(np.int), 18)
# get the functional information
lbeta = np.array([np.ravel(betas[k]) for k in range(nbsubj)]).T
# the voxel volume is 1.0
g0 = 1.0/(1.0*nvox)
bdensity = 1
dom = domain_from_array(np.ones(ref_dim))
if method=='simple':
group_map, AF, BF, likelihood = \
bsa.compute_BSA_simple(dom, lbeta, dmax, thq, smin, ths,
theta, g0, bdensity)
if method=='ipmi':
group_map, AF, BF, likelihood = \
bsa.compute_BSA_ipmi(dom, lbeta, dmax, thq, smin, ths,
theta, g0, bdensity)
if method=='sbf':
pval = 0.2
group_map, AF, BF = sbf.Compute_Amers (
dom, lbeta, dmax, theta, ths, pval)
return AF, BF
def make_bsa_2d(betas, theta=3., dmax=5., ths=0, thq=0.5, smin=0,
method='simple',verbose = 0):
"""
Function for performing bayesian structural analysis
on a set of images.
Parameters
----------
betas, array of shape (nsubj, dimx, dimy) the data used
Note that it is assumed to be a t- or z-variate
theta=3., float,
first level threshold of betas
dmax=5., float, expected between subject variability
ths=0, float,
null hypothesis for the prevalence statistic
thq=0.5, float,
p-value of the null rejection
smin=0, int,
threshold on the nu_mber of contiguous voxels
to make regions meaningful structures
method= 'simple', string,
estimation method used ; to be chosen among
'simple', 'dev', 'loo', 'ipmi'
verbose=0, verbosity mode
Returns
-------
AF the landmark_regions instance describing the result
BF: list of hroi instances describing the individual data
"""
ref_dim = np.shape(betas[0])
nsubj = betas.shape[0]
xyz = np.array(np.where(betas[:1])).T.astype(np.int)
nvox = np.size(xyz, 0)
# create the field strcture that encodes image topology
Fbeta = ff.Field(nvox)
Fbeta.from_3d_grid(xyz, 18)
# Get coordinates in mm
coord = xyz.astype(np.float)
# get the functional information
lbeta = np.array([np.ravel(betas[k]) for k in range(nsubj)]).T
# the voxel volume is 1.0
g0 = 1.0/(1.0*nvox)*1./np.sqrt(2*np.pi*dmax**2)
affine = np.eye(4)
shape = (1, ref_dim[0], ref_dim[1])
lmax=0
bdensity = 1
if method=='ipmi':
group_map, AF, BF, likelihood = \
bsa.compute_BSA_ipmi(Fbeta, lbeta, coord, dmax, xyz,
affine, shape, thq,
smin, ths, theta, g0, bdensity)
if method=='simple':
group_map, AF, BF, likelihood = \
bsa.compute_BSA_simple(Fbeta, lbeta, coord, dmax, xyz,
affine, shape, thq, smin, ths,
theta, g0)
if method=='loo':
mll, ll0 = bsa.compute_BSA_loo(Fbeta, lbeta, coord, dmax, xyz,
affine, shape, thq, smin, ths,
theta, g0)
return mll, ll0
if method=='dev':
group_map, AF, BF, likelihood = \
bsa.compute_BSA_dev(Fbeta, lbeta, coord, dmax, xyz,
affine, shape, thq,
smin, ths, theta, g0, bdensity)
if method=='simple_quick':
likelihood = np.zeros(ref_dim)
group_map, AF, BF, coclustering = \
bsa.compute_BSA_simple_quick(Fbeta, lbeta, coord, dmax, xyz,
affine, shape, thq, smin, ths,
theta, g0)
if method=='sbf':
likelihood = np.zeros(ref_dim)
group_map, AF, BF = sbf.Compute_Amers (Fbeta, lbeta, xyz, affine, shape,
coord, dmax=dmax, thr=theta,
ths=ths , pval=thq)
if method not in['loo', 'dev','simple','ipmi','simple_quick','sbf']:
raise ValueError,'method is not ocrreactly defined'
if verbose==0:
return AF,BF
if AF != None:
lmax = AF.k+2
AF.show()
group_map.shape = ref_dim
mp.figure()
mp.subplot(1,3,1)
mp.imshow(group_map, interpolation='nearest', vmin=-1, vmax=lmax)
mp.title('Blob separation map')
mp.colorbar()
if AF != None:
group_map = AF.map_label(coord,0.95,dmax)
group_map.shape = ref_dim
mp.subplot(1,3,2)
mp.imshow(group_map, interpolation='nearest', vmin=-1, vmax=lmax)
mp.title('group-level position 95% \n confidence regions')
mp.colorbar()
mp.subplot(1,3,3)
likelihood.shape = ref_dim
mp.imshow(likelihood, interpolation='nearest')
mp.title('Spatial density under h1')
mp.colorbar()
mp.figure()
if nsubj==10:
for s in range(nsubj):
mp.subplot(2, 5, s+1)
lw = -np.ones(ref_dim)
if BF[s]!=None:
nls = BF[s].get_roi_feature('label')
nls[nls==-1] = np.size(AF)+2
for k in range(BF[s].k):
xyzk = BF[s].xyz[k].T
lw[xyzk[1],xyzk[2]] = nls[k]
mp.imshow(lw, interpolation='nearest', vmin=-1, vmax=lmax)
mp.axis('off')
mp.figure()
if nsubj==10:
for s in range(nsubj):
mp.subplot(2,5,s+1)
mp.imshow(betas[s],interpolation='nearest',vmin=betas.min(),
vmax=betas.max())
mp.axis('off')
return AF, BF
def make_bsa_image_with_output_paths(mask_images, betas, denspath, crpath,
theta=3., dmax= 5., ths=0, thq=0.5, smin=0,
method='simple'):
"""
Deprecated : will be removed soon
idem make_bsa_image but paths of the output are set explictly.
Moreover the segmented regions are written in one single image
"""
# Sanity check
if len(mask_images)!=len(betas):
print len(mask_images),len(betas)
raise ValueError,"the number of masks and activation images\
should be the same"
nsubj = len(mask_images)
# Read the referential information
nim = load(mask_images[0])
ref_dim = nim.get_shape()
affine = nim.get_affine()
# Read the masks and compute the "intersection"
mask = intersect_masks(mask_images)
xyz = np.array(np.where(mask)).T
nvox = xyz.shape[0]
# create the field strcture that encodes image topology
Fbeta = ff.Field(nvox)
Fbeta.from_3d_grid(xyz.astype(np.int),18)
# Get coordinates in mm
xyz = np.hstack((xyz,np.ones((nvox,1))))
coord = np.dot(xyz,affine.T)[:,:3]
xyz = xyz.astype(np.int)
# read the functional images
lbeta = []
for s in range(nsubj):
rbeta = load(betas[s])
beta = rbeta.get_data()
beta = beta[mask]
lbeta.append(beta)
lbeta = np.array(lbeta).T
lbeta = np.reshape(lbeta,(nvox,nsubj))
# launch the method
g0 = 1.0/(np.absolute(np.linalg.det(affine))*nvox)
bdensity = 1
crmap = np.zeros(nvox)
p = np.zeros(nvox)
AF = None
BF = [None for s in range(nsubj)]
if method=='ipmi':
crmap,AF,BF,p = bsa.compute_BSA_ipmi(Fbeta, lbeta, coord, dmax,
xyz[:,:3], affine, ref_dim, thq, smin, ths,
theta, g0, bdensity)
if method=='dev':
crmap,AF,BF,p = bsa.compute_BSA_dev (Fbeta, lbeta, coord,
dmax, xyz[:,:3], affine, ref_dim,
thq, smin,ths, theta, g0, bdensity,verbose=1)
if method=='simple':
crmap,AF,BF,p = bsa.compute_BSA_simple (Fbeta, lbeta, coord, dmax,
xyz[:,:3], affine, ref_dim,
thq, smin, ths, theta, g0, verbose=0)
if method=='simple_quick':
crmap,AF,BF,co_clust = bsa.compute_BSA_simple_quick (Fbeta, lbeta, coord, dmax,
xyz[:,:3], affine, ref_dim,
thq, smin, ths, theta, g0, verbose=0)
density = np.zeros(nvox)
crmap = AF.map_label(coord,0.95,dmax)
if method=='loo':
crmap,AF,BF,p = bsa.compute_BSA_loo (Fbeta, lbeta, coord, dmax,
xyz[:,:3], affine, ref_dim,
thq, smin,ths, theta, g0, verbose=0)
# Write the results
Label = -2*np.ones(ref_dim,'int16')
Label[mask] = crmap.astype('i')
wim = Nifti1Image (Label, affine)
wim.get_header()['descrip'] = 'group Level labels from bsa procedure'
save(wim, crpath)
density = np.zeros(ref_dim)
density[mask] = p
wim = Nifti1Image (density, affine)
wim.get_header()['descrip'] = 'group-level spatial density of active regions'
save(wim, denspath)
if AF==None:
default_idx = 0
else:
default_idx = AF.k+2
# write everything in one image
wdim = (ref_dim[0], ref_dim[1], ref_dim[2], nsubj+1)
Label = -2*np.ones(wdim,'int16')
Label[mask,0] = crmap.astype(np.int)
for s in range(nsubj):
Label[mask,s+1]=-1
if BF[s]!=None:
nls = BF[s].get_roi_feature('label')
nls[nls==-1] = default_idx
for k in range(BF[s].k):
xyzk = BF[s].xyz[k].T
Label[xyzk[0],xyzk[1],xyzk[2],s+1] = nls[k]
wim = Nifti1Image (Label, affine)
wim.get_header()['descrip'] = 'group Level and individual labels\
from bsa procedure'
save(wim, crpath)
return AF,BF, maxc
def make_bsa_image(mask_images, betas, theta=3., dmax= 5., ths=0, thq=0.5,
smin=0, swd="/tmp/", method='simple', subj_id=None,
nbeta='default', densPath=None, crPath=None, verbose=0):
"""
main function for performing bsa on a set of images.
It creates the some output images in the given directory
Parameters
------------
mask_images: A list of image paths that yield binary images,
one for each subject
the number os subjects, nsubj, is taken as len(mask_images)
betas: A list of image paths that yields the activation images,
one for each subject
theta=3., threshold used to ignore all the image data that si below
dmax=5., prior width of the spatial model;
corresponds to multi-subject uncertainty
ths=0: threshold on the representativity measure of the obtained
regions
thq=0.5: p-value of the representativity test:
test = p(representativity>ths)>thq
smin=0: minimal size (in voxels) of the extracted blobs
smaller blobs are merged into larger ones
swd='/tmp': writedir
method='simple': applied region detection method; to be chose among
'simple', 'dev','ipmi'
subj_id=None: list of strings, identifiers of the subjects.
by default it is range(nsubj)
nbeta='default', string, identifier of the contrast
densPath=None, string, path of the output density image
if False, no image is written
if None, the path is computed from swd, nbeta
crPath=None, string, path of the (4D) output label image
if False, no ime is written
if None, many images are written,
with paths computed from swd, subj_id and nbeta
Returns
-------
AF: an nipy.neurospin.spatial_models.structural_bfls.landmark_regions
instance that describes the structures found at the group level
None is returned if nothing has been found significant
at the group level
BF : a list of nipy.neurospin.spatial_models.hroi.Nroi instances
(one per subject) that describe the individual coounterpart of AF
if method=='loo', the output is different:
mll, float, the average likelihood of the data under H1 after cross validation
ll0, float the log-likelihood of the data under the global null
fixme: unique mask should be allowed
"""
# Sanity check
if len(mask_images)!=len(betas):
raise ValueError,"the number of masks and activation images\
should be the same"
nsubj = len(mask_images)
if subj_id==None:
subj_id = [str[i] for i in range(nsubj)]
# Read the referential information
nim = load(mask_images[0])
ref_dim = nim.get_shape()
affine = nim.get_affine()
# Read the masks and compute the "intersection"
mask = intersect_masks(mask_images)
xyz = np.array(np.where(mask)).T
nvox = xyz.shape[0]
# create the field strcture that encodes image topology
Fbeta = ff.Field(nvox)
Fbeta.from_3d_grid(xyz.astype(np.int),18)
# Get coordinates in mm
xyz = np.hstack((xyz,np.ones((nvox,1))))
coord = np.dot(xyz,affine.T)[:,:3]
xyz = xyz.astype(np.int)
# read the functional images
lbeta = []
for s in range(nsubj):
rbeta = load(betas[s])
beta = rbeta.get_data()
beta = beta[mask]
lbeta.append(beta)
lbeta = np.array(lbeta).T
# launch the method
g0 = 1.0/(np.absolute(np.linalg.det(affine))*nvox)
bdensity = 1
crmap = np.zeros(nvox)
p = np.zeros(nvox)
AF = None
BF = [None for s in range(nsubj)]
if method=='ipmi':
crmap,AF,BF,p = bsa.compute_BSA_ipmi(Fbeta, lbeta, coord, dmax,
xyz[:,:3], affine, ref_dim, thq, smin, ths,
theta, g0, bdensity, verbose=verbose)
if method=='dev':
crmap,AF,BF,p = bsa.compute_BSA_dev (Fbeta, lbeta, coord,
dmax, xyz[:,:3], affine, ref_dim,
thq, smin,ths, theta, g0, bdensity, verbose=verbose)
if method=='simple':
crmap,AF,BF,p = bsa.compute_BSA_simple (Fbeta, lbeta, coord, dmax,
xyz[:,:3], affine, ref_dim,
thq, smin, ths, theta, g0, verbose=verbose)
if method=='simple_quick':
crmap,AF,BF,co_clust = bsa.compute_BSA_simple_quick(Fbeta, lbeta, coord, dmax,
xyz[:,:3], affine, ref_dim,
thq, smin, ths, theta, g0, verbose=verbose)
density = np.zeros(nvox)
crmap = AF.map_label(coord,0.95,dmax)
if method=='loo':
mll, ll0 = bsa.compute_BSA_loo (Fbeta, lbeta, coord, dmax,
xyz[:,:3], affine, ref_dim,
thq, smin,ths, theta, g0, verbose=verbose)
return mll, ll0
# Write the results as images
# the spatial density image
if densPath != False:
density = np.zeros(ref_dim)
density[mask] = p
wim = Nifti1Image (density, affine)
wim.get_header()['descrip'] = 'group-level spatial density of active regions'
if densPath==None:
densPath = op.join(swd,"density_%s.nii"%nbeta)
save(wim, densPath)
if crPath==False:
return AF, BF
if AF==None:
default_idx = 0
else:
default_idx = AF.k+2
if crPath==None:
# write a 3D image for group-level labels
crPath = op.join(swd,"CR_%s.nii"%nbeta)
Label = -2*np.ones(ref_dim,'int16')
Label[mask] = crmap
wim = Nifti1Image (Label, affine)
wim.get_header()['descrip'] = 'group Level labels from bsa procedure'
save(wim, crPath)
#write 3d images for the subjects
for s in range(nsubj):
LabelImage = op.join(swd,"AR_s%s_%s.nii"%(subj_id[s],nbeta))
Label = -2*np.ones(ref_dim,'int16')
Label[mask]=-1
if BF[s]!=None:
nls = BF[s].get_roi_feature('label')
nls[nls==-1] = default_idx
for k in range(BF[s].k):
xyzk = BF[s].xyz[k].T
Label[xyzk[0],xyzk[1],xyzk[2]] = nls[k]
wim = Nifti1Image (Label, affine)
wim.get_header()['descrip'] = 'Individual label image from bsa procedure'
save(wim, LabelImage)
else:
# write everything in a single 4D image
wdim = (ref_dim[0], ref_dim[1], ref_dim[2], nsubj+1)
Label = -2*np.ones(wdim,'int16')
Label[mask,0] = crmap
for s in range(nsubj):
Label[mask,s+1]=-1
if BF[s]!=None:
nls = BF[s].get_roi_feature('label')
nls[nls==-1] = default_idx
for k in range(BF[s].k):
xyzk = BF[s].xyz[k].T
Label[xyzk[0],xyzk[1],xyzk[2],s+1] = nls[k]
wim = Nifti1Image (Label, affine)
wim.get_header()['descrip'] = 'group Level and individual labels\
from bsa procedure'
save(wim, crPath)
return AF,BF
def make_bsa_image(
mask_images, betas, theta=3., dmax= 5., ths=0, thq=0.5, smin=0, swd=None,
method='simple', subj_id=None, nbeta='default', densPath=None,
crPath=None, verbose=0, reshuffle=False):
"""
main function for performing bsa on a set of images.
It creates the some output images in the given directory
Parameters
------------
mask_images: A list of image paths that yield binary images,
one for each subject
the number os subjects, nsubj, is taken as len(mask_images)
betas: A list of image paths that yields the activation images,
one for each subject
theta=3., threshold used to ignore all the image data that si below
dmax=5., prior width of the spatial model;
corresponds to multi-subject uncertainty
ths=0: threshold on the representativity measure of the obtained
regions
thq=0.5: p-value of the representativity test:
test = p(representativity>ths)>thq
smin=0: minimal size (in voxels) of the extracted blobs
smaller blobs are merged into larger ones
swd: string, optional
if not None, output directory
method='simple': applied region detection method; to be chose among
'simple', 'ipmi'
subj_id=None: list of strings, identifiers of the subjects.
by default it is range(nsubj)
nbeta='default', string, identifier of the contrast
densPath=None, string, path of the output density image
if False, no image is written
if None, the path is computed from swd, nbeta
crPath=None, string, path of the (4D) output label image
if False, no ime is written
if None, many images are written,
with paths computed from swd, subj_id and nbeta
reshuffle: bool, optional
if true, randomly swap the sign of the data
Returns
-------
AF: an nipy.neurospin.spatial_models.structural_bfls.landmark_regions
instance that describes the structures found at the group level
None is returned if nothing has been found significant
at the group level
BF : a list of nipy.neurospin.spatial_models.hroi.Nroi instances
(one per subject) that describe the individual coounterpart of AF
if method=='loo', the output is different:
mll, float, the average likelihood of the data under the model
after cross validation
ll0, float the log-likelihood of the data under the global null
fixme: unique mask should be allowed
"""
# Sanity check
if len(mask_images)!=len(betas):
raise ValueError,"the number of masks and activation images\
should be the same"
nsubj = len(mask_images)
if subj_id==None:
subj_id = [str(i) for i in range(nsubj)]
# Read the referential information
nim = load(mask_images[0])
ref_dim = nim.get_shape()[:3]
affine = nim.get_affine()
# Read the masks and compute the "intersection"
mask = np.reshape(intersect_masks(mask_images), ref_dim)
# encode it as a domain
dom = domain_from_image(Nifti1Image(mask, affine), nn=18)
nvox = dom.size
# read the functional images
lbeta = []
for s in range(nsubj):
rbeta = load(betas[s])
beta = np.reshape(rbeta.get_data(), ref_dim)
lbeta.append(beta[mask])
lbeta = np.array(lbeta).T
if reshuffle:
rswap = 2*(np.random.randn(nsubj)>0.5)-1
lbeta = np.dot(lbeta, np.diag(rswap))
# launch the method
crmap = np.zeros(nvox)
p = np.zeros(nvox)
AF = None
BF = [None for s in range(nsubj)]
if method=='ipmi':
crmap,AF,BF,p = bsa.compute_BSA_ipmi(dom, lbeta, dmax, thq, smin,
ths, theta, verbose=verbose)
if method=='simple':
crmap,AF,BF,p = bsa.compute_BSA_simple(
dom, lbeta, dmax, thq, smin, ths, theta, verbose=verbose)
if method=='quick':
crmap, AF, BF, co_clust = bsa.compute_BSA_quick(
dom, lbeta, dmax, thq, smin, ths, theta, verbose=verbose)
density = np.zeros(nvox)
crmap = AF.map_label(dom.coord, 0.95, dmax)
if method=='loo':
mll, ll0 = bsa.compute_BSA_loo (
dom, lbeta, dmax, thq, smin, ths, theta, verbose=verbose)
return mll, ll0
# Write the results as images
# the spatial density image
if densPath != False:
density = np.zeros(ref_dim)
density[mask] = p
wim = Nifti1Image (density, affine)
wim.get_header()['descrip'] = 'group-level spatial density \
of active regions'
if densPath==None:
densPath = op.join(swd,"density_%s.nii"%nbeta)
save(wim, densPath)
if crPath==False:
return AF, BF
default_idx = AF.k+2
if crPath==None and swd==None:
return AF, BF
if crPath==None:
# write a 3D image for group-level labels
crPath = op.join(swd, "CR_%s.nii"%nbeta)
Label = -2*np.ones(ref_dim, 'int16')
Label[mask] = crmap
wim = Nifti1Image (Label, affine)
wim.get_header()['descrip'] = 'group Level labels from bsa procedure'
save(wim, crPath)
# write a prevalence image
crPath = op.join(swd, "prevalence_%s.nii"%nbeta)
# prev = AF.prevalence_density()
prev = np.zeros(ref_dim)
prev[mask] = AF.prevalence_density()
wim = Nifti1Image (prev, affine)
wim.get_header()['descrip'] = 'Weighted prevalence image'
save(wim, crPath)
#write 3d images for the subjects
for s in range(nsubj):
LabelImage = op.join(swd,"AR_s%s_%s.nii"%(subj_id[s],nbeta))
Label = -2*np.ones(ref_dim, 'int16')
Label[mask] = -1
if BF[s] != None:
nls = BF[s].get_roi_feature('label')
nls[nls==-1] = default_idx
lab = BF[s].label
lab[lab>-1] = nls[lab[lab>-1]]
Label[mask] = lab
wim = Nifti1Image (Label, affine)
wim.get_header()['descrip'] = \
'Individual label image from bsa procedure'
save(wim, LabelImage)
else:
# write everything in a single 4D image
wdim = (ref_dim[0], ref_dim[1], ref_dim[2], nsubj+1)
Label = -2*np.ones(wdim,'int16')
Label[mask, 0] = crmap
for s in range(nsubj):
Label[mask,s+1]=-1
if BF[s]!=None:
nls = BF[s].get_roi_feature('label')
nls[nls==-1] = default_idx
lab = BF[s].label
lab[lab>-1] = nls[lab[lab>-1]]
Label[mask, s+1] = lab
#for k in range(BF[s].k):
# Label[mask, s+1][BF[s].label==k] = nls[k]
wim = Nifti1Image (Label, affine)
wim.get_header()['descrip'] = 'group Level and individual labels\
from bsa procedure'
save(wim, crPath)
return AF,BF