def SubjectShuffleSplit(dataset,
groups,
n_iter=100,
test_size=.3,
random_state=42):
""" Specific ShuffleSplit (train on all subject images,
but test only on one image of the remaining subjects)"""
idx = set_group_indices(dataset.dx_group)
groups_idx = np.hstack([idx[group] for group in groups])
subjects = np.asarray(dataset.subjects)
subjects = subjects[groups_idx]
subjects_unique = np.unique(subjects)
n = len(subjects_unique)
ss = ShuffleSplit(n,
n_iter=n_iter,
test_size=test_size,
random_state=random_state)
subj_ss = []
for train, test in ss:
train_set = np.array([], dtype=int)
for subj in subjects_unique[train]:
subj_ind = np.where(subjects == subj)
train_set = np.concatenate((train_set, subj_ind[0]))
test_set = np.array([], dtype=int)
for subj in subjects_unique[test]:
subj_ind = np.where(subjects == subj)
test_set = np.concatenate((test_set, subj_ind[0]))
subj_ss.append([train_set, test_set])
return subj_ss
def StratifiedSubjectShuffleSplit(dataset,
groups,
n_iter=100,
test_size=.3,
random_state=42):
""" Stratified ShuffleSplit on subjects
(train and test size may change depending on the number of acquistions)"""
idx = set_group_indices(dataset.dx_group)
groups_idx = np.hstack([idx[group] for group in groups])
subjects = np.asarray(dataset.subjects)
subjects = subjects[groups_idx]
dx = np.asarray(dataset.dx_group)
dx = dx[groups_idx]
# extract unique subject ids and dx
subjects_unique_values, \
subjects_unique_indices = np.unique(subjects, return_index=True)
# extract indices for the needed groups
dx_unique_values = dx[subjects_unique_indices]
y = dx_unique_values
# generate folds stratified on dx
sss = StratifiedShuffleSplit(y,
n_iter=n_iter,
test_size=test_size,
random_state=random_state)
ssss = []
for tr, ts in sss:
# get training subjects
subjects_tr = subjects_unique_values[tr]
# get testing subjects
subjects_ts = subjects_unique_values[ts]
# get all subject indices
train = []
test = []
for subj in subjects_tr:
train.extend(np.where(subjects == subj)[0])
for subj in subjects_ts:
test.extend(np.where(subjects == subj)[0])
# append ssss
ssss.append([train, test])
return ssss
def __init__(self,
imgs,
mask,
atlas,
dx_group,
rois=False,
n_iter=100,
memory=CACHE_DIR,
memory_level=2,
n_jobs=-1,
random_state=42):
""" initialization
"""
self.imgs = np.array(imgs)
self.mask = mask
self.atlas = atlas
self.rois = rois
self.n_jobs = n_jobs
self.n_iter = n_iter
self.memory = memory
self.memory_level = memory_level
self.idx = set_group_indices(dx_group)
self.random_state = random_state
metric=metric,
mask=mask,
detrend=True,
low_pass=.1,
high_pass=.01,
t_r=3.,
resampling_target='data',
smoothing_fwhm=6.,
memory=CACHE_DIR,
memory_level=2,
n_jobs=20)
fc = conn.fit(dataset.func)
np.savez_compressed('longitudinal_dartel_fc', data=fc)
idx = set_group_indices(dataset.dx_group)
groups = [['AD', 'MCI'], ['AD', 'Normal'], ['MCI', 'Normal']]
for g in groups:
t, p = stats.ttest_ind(fc[idx[g[0]]], fc[idx[g[1]]])
tv = t
tv[np.where(p > .01)] = 0
n_rois = conn.rois['n_rois']
centroids = conn.rois['rois_centroids']
ind = np.tril_indices(n_rois, k=-1)
m = np.zeros((n_rois, n_rois))
m[ind] = tv
m = .5 * (m + m.T)
plot_connectome(m, centroids, title='_'.join(g))