def __init__(self, data, design_matrix, mask=None, formula=None, model=def_model, method=None, niter=def_niter):
# Convert input data and design into sequences
if not hasattr(data, "__iter__"):
data = [data]
if not hasattr(design_matrix, "__iter__"):
design_matrix = [design_matrix]
# configure spatial properties
# the 'sampling' direction is assumed to be the last
# TODO: check that all input images have the same shape and
# that it's consistent with the mask
nomask = mask == None
if nomask:
self.xyz = None
self.axis = len(data[0].get_shape()) - 1
else:
self.xyz = np.where(mask.get_data() > 0)
self.axis = 1
self.spatial_shape = data[0].get_shape()[0:-1]
self.affine = data[0].get_affine()
self.glm = []
for i in range(len(data)):
if not isinstance(design_matrix[i], np.ndarray):
raise ValueError("Invalid design matrix")
if nomask:
Y = data[i].get_data()
else:
Y = data[i].get_data()[self.xyz]
X = design_matrix[i]
self.glm.append(glm(Y, X, axis=self.axis, formula=formula, model=model, method=method, niter=niter))
def _run_interface(self, runtime):
beta_nii = nb.load(self.inputs.beta)
if isdefined(self.inputs.mask):
mask = nb.load(self.inputs.mask).get_data() > 0
else:
mask = np.ones(beta_nii.shape[:3]) == 1
glm = GLM.glm()
nii = nb.load(self.inputs.beta)
glm.beta = beta_nii.get_data().copy()[mask,:].T
glm.nvbeta = self.inputs.nvbeta
glm.s2 = nb.load(self.inputs.s2).get_data().copy()[mask]
glm.dof = self.inputs.dof
glm._axis = self.inputs.axis
glm._constants = self.inputs.constants
reg_names = self.inputs.reg_names
self._stat_maps = []
self._p_maps = []
self._z_maps = []
for contrast_def in self.inputs.contrasts:
name = contrast_def[0]
type = contrast_def[1]
contrast = np.zeros(len(reg_names))
for i, reg_name in enumerate(reg_names):
if reg_name in contrast_def[2]:
idx = contrast_def[2].index(reg_name)
contrast[i] = contrast_def[3][idx]
est_contrast = glm.contrast(contrast)
stat_map = np.zeros(mask.shape)
stat_map[mask] = est_contrast.stat().T
stat_map_file = os.path.abspath(name + "_stat_map.nii")
nb.save(nb.Nifti1Image(stat_map, nii.get_affine()), stat_map_file)
self._stat_maps.append(stat_map_file)
p_map = np.zeros(mask.shape)
p_map[mask] = est_contrast.pvalue().T
p_map_file = os.path.abspath(name + "_p_map.nii")
nb.save(nb.Nifti1Image(p_map, nii.get_affine()), p_map_file)
self._p_maps.append(p_map_file)
z_map = np.zeros(mask.shape)
z_map[mask] = est_contrast.zscore().T
z_map_file = os.path.abspath(name + "_z_map.nii")
nb.save(nb.Nifti1Image(z_map, nii.get_affine()), z_map_file)
self._z_maps.append(z_map_file)
return runtime
def GLMFit(file, designMatrix, output_glm, outputCon,
fit="Kalman_AR1", mask_url=None):
"""
Call the GLM Fit function with apropriate arguments
Parameters
----------
file, string or list of strings,
path of the fMRI data file(s)
designmatrix, string, path of the design matrix .csv file
mask_url=None string, path of the mask file
if None, no mask is applied
output_glm, string,
path of the output glm .npz dump
outputCon, string,
path of the output configobj contrast object
fit= 'Kalman_AR1', string to be chosen among
"Kalman_AR1", "Ordinary Least Squares", "Kalman"
that represents both the model and the fit method
Returns
-------
glm, a nipy.neurospin.glm.glm instance representing the GLM
fixme: mask should be optional
"""
if fit == "Kalman_AR1":
model = "ar1"
method = "kalman"
elif fit == "Ordinary Least Squares":
method = "ols"
model="spherical"
elif fit == "Kalman":
method = "kalman"
model = "spherical"
import DesignMatrix as dm
names, X = dm.load_dmtx_from_csv(designMatrix)
Y = load_image(file, mask_url)
import nipy.neurospin.glm as GLM
glm = GLM.glm()
glm.fit(Y.T, X, method=method, model=model)
glm.save(output_glm)
cobj = ConfigObj(outputCon)
cobj["DesignFilePath"] = designMatrix
cobj["mask_url"] = mask_url
cobj.write()
return glm
def linear_model_fit(data_images, mask_images, design_matrix, vector):
"""
Helper function for group data analysis using arbitrary design matrix
"""
# Prepare arrays
data, vardata, xyz, mask = prepare_arrays(data_images, None, mask_images)
# Create glm instance
G = glm(data, design_matrix)
# Compute requested contrast
c = G.contrast(vector)
# Compute z-map image
zmap = np.zeros(data_images[0].get_shape()).squeeze()
zmap[list(xyz)] = c.zscore()
zimg = Image(zmap, data_images[0].get_affine())
return zimg
# axis defines the "time direction"
y = np.random.randn(dimt, dimx*dimy*dimz)
axis = 0
"""
y = random.randn(dimx, dimt, dimy, dimz)
axis = 1
"""
X = np.array([np.ones(dimt), range(dimt)])
X = X.transpose() ## the design matrix X must have dimt lines
#mod = glm.glm(y, X, axis=axis) ## default is spherical model using OLS
mod = glm.glm(y, X, axis=axis, model='ar1')
#mod = glm.glm(y, X, formula='y~x1+(x1|x2)', axis=axis, model='mfx')
##mod.save('toto')
##mod = glm.load('toto')
# Define a t contrast
tcon = mod.contrast([1,0])
# Compute the t-stat
t = tcon.stat()
## t = tcon.stat(baseline=1) to test effects > 1
# Compute the p-value
p = tcon.pvalue()
fmri_data = surrogate_4d_dataset(shape=shape, n_scans=n_scans)[0] # if you want to save it as an image data_file = op.join(swd, 'fmri_data.nii') save(fmri_data, data_file) ######################################## # Perform a GLM analysis ######################################## # GLM fit Y = fmri_data.get_data() model = "ar1" method = "kalman" glm = GLM.glm() mp.pcolor(X) mp.show() glm.fit(Y.T, X, method=method, model=model) #explained = np.dot(X,glm.beta.reshape(X.shape[1],-1)).reshape(Y.T.shape).T #residuals = Y - explained #residuals_image = Nifti1Image(np.reshape(residuals, shape), affine) # specify the contrast [1 -1 0 ..] contrast = np.zeros(X.shape[1]) contrast[0] = 1 contrast[1] = -1 my_contrast = glm.contrast(contrast) # compute the constrast image related to it zvals = my_contrast.zscore()
import numpy as np
import pylab as p
from nipy.neurospin import glm
# Load data
data = np.loadtxt('ad_data2.txt')
age = data[:,0]
# Linear regression
X = np.asarray([age, np.ones(len(age))]).T
Y = data[:,1:] # csf, gm, wm
mod = glm.glm(Y, X)
# t-test: return z-scores
# Note that absolute values indicate significance and
# signs whether effects are positive or negative
c = mod.contrast([1,0])
z = c.zscore()
pval = c.pvalue()
# Display
fit = np.dot(X, mod.beta)
p.plot(age, Y[:,0], 'b+')
p.plot(age, fit[:,0], 'b')
p.plot(age, Y[:,1], 'r+')
p.plot(age, fit[:,1], 'r')
p.plot(age, Y[:,2], 'g+')
p.plot(age, fit[:,2], 'g')
fmri_data = surrogate_4d_dataset(shape=shape, n_scans=n_scans)[0] # if you want to save it as an image data_file = op.join(swd,'fmri_data.nii') save(fmri_data, data_file) ######################################## # Perform a GLM analysis ######################################## # GLM fit Y = fmri_data.get_data() model = "ar1" method = "kalman" glm = GLM.glm() mp.pcolor(X) mp.show() glm.fit(Y.T, X, method=method, model=model) #explained = np.dot(X,glm.beta.reshape(X.shape[1],-1)).reshape(Y.T.shape).T #residuals = Y - explained #residuals_image = Nifti1Image(np.reshape(residuals, shape), affine) # specify the contrast [1 -1 0 ..] contrast = np.zeros(X.shape[1]) contrast[0] = 1 contrast[1] = -1 my_contrast = glm.contrast(contrast) # compute the constrast image related to it
## Models
#X = np.asarray([baseline, conditions, saturation]).T
#X = np.asarray([conditions, saturation, subject_factor]).T
#formula='y~1+x1+x2+(1|x3)+(x1|x3)+(x2|x3)'
#contrasts = ([0,1,0], [0,0,1])
X = np.asarray([conditions, subject_factor]).T
formula = 'y~x1+(1|x2)'
contrasts = ([1,0], )
# Test: reduce data
y = Y.reshape([Y.shape[0],Y.shape[1],ndata])
y = y[0:3,0:3,:]
# Standard t-stat
print('Starting fitting...')
tic = time()
m = glm(y, X, axis=2, formula=formula, model='mfx')
dt = time()-tic
print(' duration = %d sec' % dt)
m.save('dump')
# Linear contrast
for con in contrasts:
c = m.contrast(con)
t = c.stat()
display(t, title='Linear contrast')
def _run_interface(self, runtime):
session_info = self.inputs.session_info
functional_runs = self.inputs.session_info[0]['scans']
if isinstance(functional_runs, str):
functional_runs = [functional_runs]
nii = nb.load(functional_runs[0])
data = nii.get_data()
if isdefined(self.inputs.mask):
mask = nb.load(self.inputs.mask).get_data() > 0
else:
mask = np.ones(nii.shape[:3]) == 1
timeseries = data.copy()[mask,:]
del data
for functional_run in functional_runs[1:]:
nii = nb.load(functional_run)
data = nii.get_data()
npdata = data.copy()
del data
timeseries = np.concatenate((timeseries,npdata[mask,:]), axis=1)
del npdata
nscans = timeseries.shape[1]
if 'hpf' in session_info[0].keys():
hpf = session_info[0]['hpf']
drift_model=self.inputs.drift_model
else:
hpf=0
drift_model = "Blank"
reg_names = []
for reg in session_info[0]['regress']:
reg_names.append(reg['name'])
reg_vals = np.zeros((nscans,len(reg_names)))
for i in range(len(reg_names)):
reg_vals[:,i] = np.array(session_info[0]['regress'][i]['val']).reshape(1,-1)
frametimes= np.linspace(0, (nscans-1)*self.inputs.TR, nscans)
conditions = []
onsets = []
duration = []
for i,cond in enumerate(session_info[0]['cond']):
onsets += cond['onset']
conditions += [cond['name']]*len(cond['onset'])
if len(cond['duration']) == 1:
duration += cond['duration']*len(cond['onset'])
paradigm = dm.BlockParadigm(con_id=conditions, onset=onsets, duration=duration)
design_matrix, self._reg_names = dm.dmtx_light(frametimes, paradigm, drift_model=drift_model, hfcut=hpf,
hrf_model=self.inputs.hrf_model,
add_regs=reg_vals,
add_reg_names=reg_names
)
if self.inputs.normalize_design_matrix:
for i in range(len(self._reg_names)-1):
design_matrix[:,i] = (design_matrix[:,i]-design_matrix[:,i].mean())/design_matrix[:,i].std()
if self.inputs.plot_design_matrix:
pylab.pcolor(design_matrix)
pylab.savefig("design_matrix.pdf")
pylab.close()
pylab.clf()
glm = GLM.glm()
glm.fit(timeseries.T, design_matrix, method=self.inputs.method, model=self.inputs.model)
self._beta_file = os.path.abspath("beta.nii")
beta = np.zeros(mask.shape + (glm.beta.shape[0],))
beta[mask,:] = glm.beta.T
nb.save(nb.Nifti1Image(beta, nii.get_affine()), self._beta_file)
self._s2_file = os.path.abspath("s2.nii")
s2 = np.zeros(mask.shape)
s2[mask] = glm.s2
nb.save(nb.Nifti1Image(s2, nii.get_affine()), self._s2_file)
if self.inputs.save_residuals:
explained = np.dot(design_matrix,glm.beta)
residuals = np.zeros(mask.shape + (nscans,))
residuals[mask,:] = timeseries - explained.T
self._residuals_file = os.path.abspath("residuals.nii")
nb.save(nb.Nifti1Image(residuals, nii.get_affine()), self._residuals_file)
self._nvbeta = glm.nvbeta
self._dof = glm.dof
self._constants = glm._constants
self._axis = glm._axis
if self.inputs.model == "ar1":
self._a_file = os.path.abspath("a.nii")
a = np.zeros(mask.shape)
a[mask] = glm.a.squeeze()
nb.save(nb.Nifti1Image(a, nii.get_affine()), self._a_file)
self._model = glm.model
self._method = glm.method
return runtime