def select_cluster_level_threshold(
effects, var, xyz, nbsamp=1024, pval=0.001, method='rfx', threshold=0.,
group_size=-1, verbose=0, graph=None, volume=None):
"""
Calibrating the first-level threshold through permutations
Parameters
----------
effects, array of shape (nvox,nsubj)
input effects
var, array of shape (nvox,nsubj)
input variance
xyz, array of shape(nvox,3)
comon grid coorinates
nbsamp=1024, number of random draws
pval=0.001,frequentist threshold
method='crfx', string 'crfx' or 'cmfx'
thresold=0., cluster-forming threshold
group_size=-1, int,
when positive the subjects are drawn in the input population
without replacement if group_size<nsubj
with replacement otherwise
verbose=0, verbosity mode
graph: graph instance, optional
represents the dataset topology (used to form clusters)
volume: array, optional
yields for each voxle, node the surface/volume (1 by default)
Returns
-------
kc, float, the firest-level threshold
Note
----
the p-value is systematically corrected
"""
nvox = effects.shape[0]
nsubj = effects.shape[1]
k = np.zeros(nbsamp)
if group_size==-1:
group_size=nsubj
if volume==None:
volume = np.ones(nvox)
for irs in range(nbsamp):
# data splitting
group = np.arange(nsubj)
if group_size<nsubj:
# draw randomly group_size subjects among nsubj
group = np.argsort(np.random.rand(nsubj))[:group_size]
if group_size>nsubj:
group = (nsubj*np.random.rand(group_size)).astype(np.int)
x = effects[:,group]
if method=='mfx':
vx = var[:,group]
sswap = np.reshape(2*(np.random.rand(group_size)>0.5)-1, (1,group_size))
# analysis
if method=='rfx':
#y = fos.stat(x,id='student',axis=1, Magics=np.array([irs]))
y = mes.t_stat(sswap*x)
elif method=='mfx':
#y = fos.stat_mfx(x, vx, id='student_mfx', axis=1, Magics=np.array([irs]))
y = mes.mfx_t_stat(sswap*x, vx)
else: raise ValueError, 'unknown method'
# stat extraction
y = np.reshape(y,np.size(y))
xyzl = xyz[y>threshold,:]
n1 = xyzl.shape[0]
if n1>0:
if graph==None:
lg = fg.WeightedGraph(n1)
lg.from_3d_grid(xyzl.astype(np.int))
else:
lg = graph.subgraph(y>threshold)
u = lg.cc()
lvolume = volume[y>threshold]
cluster_vol = np.array([np.sum(lvolume[u==uu]) for uu in np.unique(u)])
mk = cluster_vol.max()
else:
mk=0
k[irs] = mk
if verbose:
print irs, n1, mk, y.min(), y.max()
# derive the thresholds
aux = np.sort(k)
kc = aux[int((1-pval)*nbsamp)]
return kc
#------------------------------------------------------------------------ thresholds = ConfigObj(threshold_path) zc_rfx = float(thresholds[cid]['zc_rfx']) zc_mfx = float(thresholds[cid]['zc_mfx']) z1_rfx = float(thresholds[cid]['z1_rfx']) z1_mfx = float(thresholds[cid]['z1_mfx']) z2_rfx = float(thresholds[cid]['z2_rfx']) z2_mfx = float(thresholds[cid]['z2_mfx']) #------------------------------------------------------------------------ # rfx #------------------------------------------------------------------------ #y = fos.stat(Functional, id='student', axis=1) y = mes.t_stat(Functional) #------------------------------------------------------------------------ # voxel-level rfx #------------------------------------------------------------------------ Label = np.zeros(grp_mask.get_shape()) Label[mask] = np.squeeze(y*(y>zc_rfx)) wim = Nifti1Image(Label, grp_mask.get_affine()) save(wim, op.join(swd,"vrfx_%s.nii"%cid)) print "Number of active voxels: %04d"% np.sum(y>zc_rfx) #------------------------------------------------------------------------ # cluster-level rfx #------------------------------------------------------------------------
def select_voxel_level_threshold( effects, var, nbsamp=1024, pval=0.001,
method='rfx', corrected=False,
group_size=-1, verbose=0):
"""
Calibrating the first-level threshold through permutations
Parameters
----------
effects, array of shape (nvox,nsubj)
input effects
var, array of shape (nvox,nsubj)
input variance
nbsamp=1024, number of random draws
pval=0.001, frequentist threshold
method='crfx', string 'crfx' or 'cmfx'
corrected=False, bool correaction for mcp or not
group_size=-1, int,
when positive the subjects are drawn in the input population
without replaecment if group_size<nsubj
with replacement otherwise
verbose=0, verbosity mode
Returns
-------
zc, float, the firest-level threshold
"""
nvox = effects.shape[0]
nsubj = effects.shape[1]
zmax = np.zeros(nbsamp)
if group_size==-1:
group_size=nsubj
for irs in range(nbsamp):
# data splitting
group = np.arange(nsubj)
if group_size<nsubj:
# draw randomly group_size subjects among nsubj
group = np.argsort(np.random.rand(nsubj))[:group_size]
if group_size>nsubj:
group = (nsubj*np.random.rand(group_size)).astype(np.int)
x = effects[:,group]
if (method=='mfx') or (method=='ffx') or (method=='cjt'):
vx = var[:,group]
sswap = np.reshape(2*(np.random.rand(group_size)>0.5)-1, (1,group_size))
# analysis
if method=='rfx':
#y = fos.stat(sswap*x, id='student', axis=1)
y = mes.t_stat(sswap*x)
elif method=='mfx':
#y = fos.stat_mfx(sswap*x, vx, id='student_mfx', axis=1)
y = mes.mfx_t_stat(sswap*x, vx)
elif method=='ffx':
y = fttest(sswap*x,vx)
elif method=='cjt':
sswap = np.reshape(2 * (np.random.rand(group_size) > 0.5) -1, (1, group_size))
y = conjunction(sswap*x, vx, group_size/2)
else: raise ValueError, 'unknown method'
# statistic extraction
if corrected:
zmax[irs] = y.max()
else:
aux = np.sort(y,0)
zmax[irs] = aux[int((1-pval)*nvox)]
if verbose:
print irs, aux.min(), zmax[irs], aux.max()
# derive the thresholds
if corrected:
aux = np.sort(zmax)
zc = aux[int((1-pval)*nbsamp)]
else:
zc = zmax[zmax!=0].mean()# fix
return zc
