def test_brain_data(tmpdir):
# Add 3mm to list to test that resolution as well
for resolution in ['2mm']:
MNI_Template["resolution"] = resolution
sim = Simulator()
r = 10
sigma = 1
y = [0, 1]
n_reps = 3
output_dir = str(tmpdir)
dat = sim.create_data(y, sigma, reps=n_reps, output_dir=output_dir)
if MNI_Template["resolution"] == '2mm':
shape_3d = (91, 109, 91)
shape_2d = (6, 238955)
elif MNI_Template["resolution"] == '3mm':
shape_3d = (60, 72, 60)
shape_2d = (6, 71020)
y = pd.read_csv(os.path.join(str(tmpdir.join('y.csv'))),header=None, index_col=None)
holdout = pd.read_csv(os.path.join(str(tmpdir.join('rep_id.csv'))),header=None,index_col=None)
# Test load list of 4D images
file_list = [str(tmpdir.join('data.nii.gz')), str(tmpdir.join('data.nii.gz'))]
dat = Brain_Data(file_list)
dat = Brain_Data([nb.load(x) for x in file_list])
# Test load list
dat = Brain_Data(data=str(tmpdir.join('data.nii.gz')), Y=y)
# Test concatenate
out = Brain_Data([x for x in dat])
assert isinstance(out, Brain_Data)
assert len(out)==len(dat)
# Test to_nifti
d = dat.to_nifti()
assert d.shape[0:3] == shape_3d
# Test load nibabel
assert Brain_Data(d)
# Test shape
assert dat.shape() == shape_2d
# Test Mean
assert dat.mean().shape()[0] == shape_2d[1]
# Test Std
assert dat.std().shape()[0] == shape_2d[1]
# Test add
new = dat + dat
assert new.shape() == shape_2d
# Test subtract
new = dat - dat
assert new.shape() == shape_2d
# Test multiply
new = dat * dat
assert new.shape() == shape_2d
# Test Indexing
index = [0, 3, 1]
assert len(dat[index]) == len(index)
index = range(4)
assert len(dat[index]) == len(index)
index = dat.Y == 1
assert len(dat[index.values.flatten()]) == index.values.sum()
assert len(dat[index]) == index.values.sum()
assert len(dat[:3]) == 3
# Test Iterator
x = [x for x in dat]
assert len(x) == len(dat)
assert len(x[0].data.shape) == 1
# # Test T-test
out = dat.ttest()
assert out['t'].shape()[0] == shape_2d[1]
# # # Test T-test - permutation method
# out = dat.ttest(threshold_dict={'permutation':'tfce','n_permutations':50,'n_jobs':1})
# assert out['t'].shape()[0]==shape_2d[1]
# Test Regress
dat.X = pd.DataFrame({'Intercept':np.ones(len(dat.Y)),
'X1':np.array(dat.Y).flatten()}, index=None)
# Standard OLS
out = dat.regress()
assert type(out['beta'].data) == np.ndarray
assert type(out['t'].data) == np.ndarray
assert type(out['p'].data) == np.ndarray
assert type(out['residual'].data) == np.ndarray
assert type(out['df'].data) == np.ndarray
assert out['beta'].shape() == (2, shape_2d[1])
assert out['t'][1].shape()[0] == shape_2d[1]
# Robust OLS
out = dat.regress(mode='robust')
assert type(out['beta'].data) == np.ndarray
assert type(out['t'].data) == np.ndarray
assert type(out['p'].data) == np.ndarray
assert type(out['residual'].data) == np.ndarray
assert type(out['df'].data) == np.ndarray
assert out['beta'].shape() == (2, shape_2d[1])
assert out['t'][1].shape()[0] == shape_2d[1]
# Test threshold
i=1
tt = threshold(out['t'][i], out['p'][i], .05)
assert isinstance(tt, Brain_Data)
# Test write
dat.write(os.path.join(str(tmpdir.join('test_write.nii'))))
assert Brain_Data(os.path.join(str(tmpdir.join('test_write.nii'))))
# Test append
assert dat.append(dat).shape()[0] == shape_2d[0]*2
# Test distance
distance = dat.distance(method='euclidean')
assert isinstance(distance, Adjacency)
assert distance.square_shape()[0] == shape_2d[0]
# Test predict
stats = dat.predict(algorithm='svm',
cv_dict={'type': 'kfolds', 'n_folds': 2},
plot=False, **{'kernel':"linear"})
# Support Vector Regression, with 5 fold cross-validation with Platt Scaling
# This will output probabilities of each class
stats = dat.predict(algorithm='svm',
cv_dict=None, plot=False,
**{'kernel':'linear', 'probability':True})
assert isinstance(stats['weight_map'], Brain_Data)
# Logistic classificiation, with 2 fold cross-validation.
stats = dat.predict(algorithm='logistic',
cv_dict={'type': 'kfolds', 'n_folds': 2},
plot=False)
assert isinstance(stats['weight_map'], Brain_Data)
# Ridge classificiation,
stats = dat.predict(algorithm='ridgeClassifier', cv_dict=None, plot=False)
assert isinstance(stats['weight_map'], Brain_Data)
# Ridge
stats = dat.predict(algorithm='ridge',
cv_dict={'type': 'kfolds', 'n_folds': 2,
'subject_id':holdout}, plot=False, **{'alpha':.1})
# Lasso
stats = dat.predict(algorithm='lasso',
cv_dict={'type': 'kfolds', 'n_folds': 2,
'stratified':dat.Y}, plot=False, **{'alpha':.1})
# PCR
stats = dat.predict(algorithm='pcr', cv_dict=None, plot=False)
# Test Similarity
r = dat.similarity(stats['weight_map'])
assert len(r) == shape_2d[0]
r2 = dat.similarity(stats['weight_map'].to_nifti())
assert len(r2) == shape_2d[0]
r = dat.similarity(stats['weight_map'], method='dot_product')
assert len(r) == shape_2d[0]
r = dat.similarity(stats['weight_map'], method='cosine')
assert len(r) == shape_2d[0]
r = dat.similarity(dat, method='correlation')
assert r.shape == (dat.shape()[0],dat.shape()[0])
r = dat.similarity(dat, method='dot_product')
assert r.shape == (dat.shape()[0],dat.shape()[0])
r = dat.similarity(dat, method='cosine')
assert r.shape == (dat.shape()[0],dat.shape()[0])
# Test apply_mask - might move part of this to test mask suite
s1 = create_sphere([12, 10, -8], radius=10)
assert isinstance(s1, nb.Nifti1Image)
masked_dat = dat.apply_mask(s1)
assert masked_dat.shape()[1] == np.sum(s1.get_data() != 0)
# Test extract_roi
mask = create_sphere([12, 10, -8], radius=10)
assert len(dat.extract_roi(mask)) == shape_2d[0]
# Test r_to_z
z = dat.r_to_z()
assert z.shape() == dat.shape()
# Test copy
d_copy = dat.copy()
assert d_copy.shape() == dat.shape()
# Test detrend
detrend = dat.detrend()
assert detrend.shape() == dat.shape()
# Test standardize
s = dat.standardize()
assert s.shape() == dat.shape()
assert np.isclose(np.sum(s.mean().data), 0, atol=.1)
s = dat.standardize(method='zscore')
assert s.shape() == dat.shape()
assert np.isclose(np.sum(s.mean().data), 0, atol=.1)
# Test Sum
s = dat.sum()
assert s.shape() == dat[1].shape()
# Test Groupby
s1 = create_sphere([12, 10, -8], radius=10)
s2 = create_sphere([22, -2, -22], radius=10)
mask = Brain_Data([s1, s2])
d = dat.groupby(mask)
assert isinstance(d, Groupby)
# Test Aggregate
mn = dat.aggregate(mask, 'mean')
assert isinstance(mn, Brain_Data)
assert len(mn.shape()) == 1
# Test Threshold
s1 = create_sphere([12, 10, -8], radius=10)
s2 = create_sphere([22, -2, -22], radius=10)
mask = Brain_Data(s1)*5
mask = mask + Brain_Data(s2)
m1 = mask.threshold(upper=.5)
m2 = mask.threshold(upper=3)
m3 = mask.threshold(upper='98%')
m4 = Brain_Data(s1)*5 + Brain_Data(s2)*-.5
m4 = mask.threshold(upper=.5,lower=-.3)
assert np.sum(m1.data > 0) > np.sum(m2.data > 0)
assert np.sum(m1.data > 0) == np.sum(m3.data > 0)
assert np.sum(m4.data[(m4.data > -.3) & (m4.data <.5)]) == 0
assert np.sum(m4.data[(m4.data < -.3) | (m4.data >.5)]) > 0
# Test Regions
r = mask.regions(min_region_size=10)
m1 = Brain_Data(s1)
m2 = r.threshold(1, binarize=True)
# assert len(r)==2
assert len(np.unique(r.to_nifti().get_data())) == 2
diff = m2-m1
assert np.sum(diff.data) == 0
# Test Bootstrap
masked = dat.apply_mask(create_sphere(radius=10, coordinates=[0, 0, 0]))
n_samples = 3
b = masked.bootstrap('mean', n_samples=n_samples)
assert isinstance(b['Z'], Brain_Data)
b = masked.bootstrap('std', n_samples=n_samples)
assert isinstance(b['Z'], Brain_Data)
b = masked.bootstrap('predict', n_samples=n_samples, plot=False)
assert isinstance(b['Z'], Brain_Data)
b = masked.bootstrap('predict', n_samples=n_samples,
plot=False, cv_dict={'type':'kfolds','n_folds':3})
assert isinstance(b['Z'], Brain_Data)
b = masked.bootstrap('predict', n_samples=n_samples,
save_weights=True, plot=False)
assert len(b['samples'])==n_samples
# Test decompose
n_components = 3
stats = dat.decompose(algorithm='pca', axis='voxels',
n_components=n_components)
assert n_components == len(stats['components'])
assert stats['weights'].shape == (len(dat), n_components)
stats = dat.decompose(algorithm='ica', axis='voxels',
n_components=n_components)
assert n_components == len(stats['components'])
assert stats['weights'].shape == (len(dat), n_components)
dat.data = dat.data + 2
dat.data[dat.data<0] = 0
stats = dat.decompose(algorithm='nnmf', axis='voxels',
n_components=n_components)
assert n_components == len(stats['components'])
assert stats['weights'].shape == (len(dat), n_components)
stats = dat.decompose(algorithm='fa', axis='voxels',
n_components=n_components)
assert n_components == len(stats['components'])
assert stats['weights'].shape == (len(dat), n_components)
stats = dat.decompose(algorithm='pca', axis='images',
n_components=n_components)
assert n_components == len(stats['components'])
assert stats['weights'].shape == (len(dat), n_components)
stats = dat.decompose(algorithm='ica', axis='images',
n_components=n_components)
assert n_components == len(stats['components'])
assert stats['weights'].shape == (len(dat), n_components)
dat.data = dat.data + 2
dat.data[dat.data<0] = 0
stats = dat.decompose(algorithm='nnmf', axis='images',
n_components=n_components)
assert n_components == len(stats['components'])
assert stats['weights'].shape == (len(dat), n_components)
stats = dat.decompose(algorithm='fa', axis='images',
n_components=n_components)
assert n_components == len(stats['components'])
assert stats['weights'].shape == (len(dat), n_components)
# Test Hyperalignment Method
sim = Simulator()
y = [0, 1]
n_reps = 10
s1 = create_sphere([0, 0, 0], radius=3)
d1 = sim.create_data(y, 1, reps=n_reps, output_dir=None).apply_mask(s1)
d2 = sim.create_data(y, 2, reps=n_reps, output_dir=None).apply_mask(s1)
d3 = sim.create_data(y, 3, reps=n_reps, output_dir=None).apply_mask(s1)
# Test procrustes using align
data = [d1, d2, d3]
out = align(data, method='procrustes')
assert len(data) == len(out['transformed'])
assert len(data) == len(out['transformation_matrix'])
assert data[0].shape() == out['common_model'].shape()
transformed = np.dot(d1.data, out['transformation_matrix'][0])
centered = d1.data - np.mean(d1.data, 0)
transformed = (np.dot(centered/np.linalg.norm(centered), out['transformation_matrix'][0])*out['scale'][0])
np.testing.assert_almost_equal(0, np.sum(out['transformed'][0].data - transformed), decimal=5)
# Test deterministic brain_data
bout = d1.align(out['common_model'], method='deterministic_srm')
assert d1.shape() == bout['transformed'].shape()
assert d1.shape() == bout['common_model'].shape()
assert d1.shape()[1] == bout['transformation_matrix'].shape[0]
btransformed = np.dot(d1.data, bout['transformation_matrix'])
np.testing.assert_almost_equal(0, np.sum(bout['transformed'].data - btransformed))
# Test deterministic brain_data
bout = d1.align(out['common_model'], method='probabilistic_srm')
assert d1.shape() == bout['transformed'].shape()
assert d1.shape() == bout['common_model'].shape()
assert d1.shape()[1] == bout['transformation_matrix'].shape[0]
btransformed = np.dot(d1.data, bout['transformation_matrix'])
np.testing.assert_almost_equal(0, np.sum(bout['transformed'].data-btransformed))
# Test procrustes brain_data
bout = d1.align(out['common_model'], method='procrustes')
assert d1.shape() == bout['transformed'].shape()
assert d1.shape() == bout['common_model'].shape()
assert d1.shape()[1] == bout['transformation_matrix'].shape[0]
centered = d1.data - np.mean(d1.data, 0)
btransformed = (np.dot(centered/np.linalg.norm(centered), bout['transformation_matrix'])*bout['scale'])
np.testing.assert_almost_equal(0, np.sum(bout['transformed'].data-btransformed), decimal=5)
np.testing.assert_almost_equal(0, np.sum(out['transformed'][0].data - bout['transformed'].data))
# Test hyperalignment on Brain_Data over time (axis=1)
sim = Simulator()
y = [0, 1]
n_reps = 10
s1 = create_sphere([0, 0, 0], radius=5)
d1 = sim.create_data(y, 1, reps=n_reps, output_dir=None).apply_mask(s1)
d2 = sim.create_data(y, 2, reps=n_reps, output_dir=None).apply_mask(s1)
d3 = sim.create_data(y, 3, reps=n_reps, output_dir=None).apply_mask(s1)
data = [d1, d2, d3]
out = align(data, method='procrustes', axis=1)
assert len(data) == len(out['transformed'])
assert len(data) == len(out['transformation_matrix'])
assert data[0].shape() == out['common_model'].shape()
centered = data[0].data.T-np.mean(data[0].data.T, 0)
transformed = (np.dot(centered/np.linalg.norm(centered), out['transformation_matrix'][0])*out['scale'][0])
np.testing.assert_almost_equal(0,np.sum(out['transformed'][0].data-transformed.T), decimal=5)
bout = d1.align(out['common_model'], method='deterministic_srm', axis=1)
assert d1.shape() == bout['transformed'].shape()
assert d1.shape() == bout['common_model'].shape()
assert d1.shape()[0] == bout['transformation_matrix'].shape[0]
btransformed = np.dot(d1.data.T, bout['transformation_matrix'])
np.testing.assert_almost_equal(0, np.sum(bout['transformed'].data-btransformed.T))
bout = d1.align(out['common_model'], method='probabilistic_srm', axis=1)
assert d1.shape() == bout['transformed'].shape()
assert d1.shape() == bout['common_model'].shape()
assert d1.shape()[0] == bout['transformation_matrix'].shape[0]
btransformed = np.dot(d1.data.T, bout['transformation_matrix'])
np.testing.assert_almost_equal(0, np.sum(bout['transformed'].data-btransformed.T))
bout = d1.align(out['common_model'], method='procrustes', axis=1)
assert d1.shape() == bout['transformed'].shape()
assert d1.shape() == bout['common_model'].shape()
assert d1.shape()[0] == bout['transformation_matrix'].shape[0]
centered = d1.data.T-np.mean(d1.data.T, 0)
btransformed = (np.dot(centered/np.linalg.norm(centered), bout['transformation_matrix'])*bout['scale'])
np.testing.assert_almost_equal(0, np.sum(bout['transformed'].data-btransformed.T), decimal=5)
np.testing.assert_almost_equal(0, np.sum(out['transformed'][0].data-bout['transformed'].data))
First, we will run a high pass filter to remove any low frequency scanner drift. We will pick a fairly arbitrary filter size of 0.0078hz (1/128s). We will also run spatial smoothing with a 6mm FWHM gaussian kernel to increase a signal to noise ratio at each voxel. These steps are very easy to run using nltools after the data has been loaded. data = data.filter(sampling_freq=1/2.4, high_pass=1/128) data = data.smooth(6) ## Independent Component Analysis (ICA) Ok, we are finally ready to run an ICA analysis on our data. ICA attempts to perform blind source separation by decomposing a multivariate signal into additive subcomponents that are maximally independent. We will be using the `decompose()` method on our `Brain_Data` instance. This runs the [FastICA](https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.fastica.html) algorithm implemented by scikit-learn. You can choose whether you want to run spatial ICA by setting `axis='voxels` or temporal ICA by setting `axis='images'`. We also recommend running the whitening flat `whiten=True`. By default `decompose` will estimate the maximum components that are possible given the data. We recommend using a completely arbitrary heuristic of 20-30 components. tr = 2.4 output = data.decompose(algorithm='ica', n_components=30, axis='images', whiten=True) ## Viewing Components We will use the interactive `component_viewer` from nltools to explore the results of the analysis. This viewer uses ipywidgets to select the `Component` to view and also the threshold. You can manually enter a component number to view or scroll up and down. Components have been standardized, this allows us to threshold the brain in terms of standard deviations. For example, the default threshold of 2.0, means that any voxel that loads on the component greater or less than 2 standard deviations will be overlaid on the standard brain. You can play with different thresholds to be more or less inclusive - a threshold of 0 will overlay all of the voxels. If you play with any of the numbers, make sure you press tab to update the plot. The second plot is the time course of the voxels that load on the component. The x-axis is in TRs, which for this dataset is 2.4 sec. The third plot is the powerspectrum of the timecourse. There is not a large range of possible values as we can only observe signals at the nyquist frequency, which is half of our sampling frequency of 1/2.4s (approximately 0.21hz) to a lower bound of 0.0078hz based on our high pass filter. There might be systematic oscillatory signals. Remember, that signals that oscillate a faster frequency than the nyquist frequency will be aliased. This includes physiological artifacts such as respiration and cardiac signals. It is important to note that ICA cannot resolve the sign of the component. So make sure you consider signals that are positive as well as negative. component_viewer(output, tr=2.4)
