def make_mask_map_4d(data, infile, outfile):
""" Make mask map with 4d dimeions
data: values for levels in infile. Shape = [4th dimension, regions]
infile: input file to replace levels with values
outfile: output file name
"""
from neurosynth.base.mask import Masker
from neurosynth.base import imageutils
from nibabel import nifti1
data = np.array(data)
# Load image with masker
masker = Masker(infile)
img = imageutils.load_imgs(infile, masker)
header = masker.get_header()
shape = header.get_data_shape()[0:3] + (data.shape[0],)
header.set_data_shape(shape)
result = []
for t_dim, t_val in enumerate(data):
result.append(img.copy())
for num, value in enumerate(t_val):
np.place(result[t_dim], img == num + 1, [value])
result = np.hstack(result)
header.set_data_dtype(result.dtype) # Avoids loss of precision
img = nifti1.Nifti1Image(masker.unmask(result).squeeze(), None, header)
img.to_filename(outfile)
def make_mask_map_4d(data, infile, outfile):
""" Make mask map with 4d dimeions
data: values for levels in infile. Shape = [4th dimension, regions]
infile: input file to replace levels with values
outfile: output file name
"""
from neurosynth.base.mask import Masker
from neurosynth.base import imageutils
from nibabel import nifti1
data = np.array(data)
# Load image with masker
masker = Masker(infile)
img = imageutils.load_imgs(infile, masker)
header = masker.get_header()
shape = header.get_data_shape()[0:3] + (data.shape[0], )
header.set_data_shape(shape)
result = []
for t_dim, t_val in enumerate(data):
result.append(img.copy())
for num, value in enumerate(t_val):
np.place(result[t_dim], img == num + 1, [value])
result = np.hstack(result)
header.set_data_dtype(result.dtype) # Avoids loss of precision
img = nifti1.Nifti1Image(masker.unmask(result).squeeze(), None, header)
img.to_filename(outfile)
def __init__(self,
dataset=None,
method='pearson',
features=None,
mask=None,
image_type='pFgA_z',
threshold=0.001):
""" Initialize a new Decoder instance.
Args:
dataset: An optional Dataset instance containing features to use in
decoding.
method: The decoding method to use (optional). By default, Pearson
correlation.
features: Optional list of features to use in decoding. If None,
use all features found in dataset. If features is a list of
strings, use only the subset of features in the Dataset that
are named in the list. If features is a list of filenames,
ignore the dataset entirely and use only the features passed as
image files in decoding.
mask: An optional mask to apply to features and input images. If
None, will use the one in the current Dataset.
image_type: An optional string indicating the type of image to use
when constructing feature-based images. See
meta.analyze_features() for details. By default, uses reverse
inference z-score images.
threshold: If decoding from a Dataset instance, this is the feature
threshold to use to generate the feature maps used in the
decoding.
"""
self.dataset = dataset
if dataset is not None:
self.masker = self.dataset.masker
if features is None:
features = dataset.get_feature_names()
if mask is not None:
self.masker.add(mask)
elif mask is not None:
self.masker = Masker(mask)
else:
self.masker = None
self.method = method.lower()
if self.method == 'roi':
self.feature_names = features
else:
self.load_features(features,
image_type=image_type,
threshold=threshold)
def make_mask_map(data, infile, outfile, index=None):
from neurosynth.base.mask import Masker
from neurosynth.base import imageutils
# Load image with masker
masker = Masker(infile)
img = imageutils.load_imgs(infile, masker)
img = np.round(img)
data = list(data)
if index is None:
index = np.unique(img)
rev_index = None
else:
all_reg = np.arange(0, img.max())
rev_index = all_reg[np.invert(np.in1d(all_reg, index))]
min_val = img.min()
for num, value in enumerate(data):
ix = index[num]
np.place(img, img == ix, [value])
if rev_index is not None:
for value in rev_index:
np.place(img, img == value + min_val, 0)
img = img.astype('float32')
imageutils.save_img(img, outfile, masker)
class TestMasker(unittest.TestCase):
def setUp(self):
""" Create a new Dataset and add features. """
maskfile = get_resource_path() + 'MNI152_T1_2mm_brain.nii.gz'
self.masker = Masker(maskfile)
def test_add_and_remove_masks(self):
self.masker.add(get_test_data_path() + 'sgacc_mask.nii.gz')
self.masker.add({'motor': get_test_data_path() + 'medial_motor.nii.gz'})
self.assertEqual(len(self.masker.layers), 2)
self.assertEqual(len(self.masker.stack), 2)
self.assertEqual(set(self.masker.layers.keys()), set(['layer_0', 'motor']))
self.assertEqual(np.sum(self.masker.layers['motor']), 1419)
self.masker.remove('motor')
self.assertEqual(len(self.masker.layers), 1)
self.assertEqual(len(self.masker.stack), 1)
self.masker.add(get_test_data_path() + 'medial_motor.nii.gz')
self.masker.remove(-1)
self.assertTrue('layer_0' in self.masker.layers.keys())
self.assertEqual(len(self.masker.layers), 1)
def __init__(self, dataset=None, method='pearson', features=None,
mask=None, image_type='pFgA_z', threshold=0.001):
""" Initialize a new Decoder instance.
Args:
dataset: An optional Dataset instance containing features to use in
decoding.
method: The decoding method to use (optional). By default, Pearson
correlation.
features: Optional list of features to use in decoding. If None,
use all features found in dataset. If features is a list of
strings, use only the subset of features in the Dataset that
are named in the list. If features is a list of filenames,
ignore the dataset entirely and use only the features passed as
image files in decoding.
mask: An optional mask to apply to features and input images. If
None, will use the one in the current Dataset.
image_type: An optional string indicating the type of image to use
when constructing feature-based images. See
meta.analyze_features() for details. By default, uses reverse
inference z-score images.
threshold: If decoding from a Dataset instance, this is the feature
threshold to use to generate the feature maps used in the
decoding.
"""
self.dataset = dataset
if dataset is not None:
self.masker = self.dataset.masker
if features is None:
features = dataset.get_feature_names()
if mask is not None:
self.masker.add(mask)
elif mask is not None:
self.masker = Masker(mask)
else:
self.masker = None
self.method = method.lower()
if self.method == 'roi':
self.feature_names = features
else:
self.load_features(features, image_type=image_type,
threshold=threshold)
# This script takes a parcellation that may not have continous numbers,
# removes parcels below some size and reorders in order
# while also outputting the region and community number
from neurosynth.base.mask import Masker
from neurosynth.base import imageutils
import numpy as np
import csv
min_vox = 300
file = '../masks/Andy/aal_MNI_V4.nii'
outfile = '../masks/Andy/aal_MNI_V4_' + str(min_vox) + '.nii'
# Load image with masker
masker = Masker(file)
img = imageutils.load_imgs(file, masker)
# How many levels in the original image
print "Original shape:"
print np.bincount([int(vox) for vox in img]).shape
# Get how many voxels per level and calc those that pass min_vox
count = np.bincount(img.astype('int').squeeze())
non_0_ix = np.where(count >= min_vox)[0]
zero_ix = np.where(count < min_vox)[0]
# Remove those not in a good community
bad = list(set(zero_ix))
# Remove
class Decoder:
def __init__(self, dataset=None, method='pearson', features=None,
mask=None, image_type='specificity_z', threshold=0.001):
""" Initialize a new Decoder instance.
Args:
dataset: An optional Dataset instance containing features to use in
decoding.
method: The decoding method to use (optional). By default, Pearson
correlation.
features: Optional list of features to use in decoding. If None,
use all features found in dataset. If features is a list of
strings, use only the subset of features in the Dataset that
are named in the list. If features is a list of filenames,
ignore the dataset entirely and use only the features passed as
image files in decoding.
mask: An optional mask to apply to features and input images. If
None, will use the one in the current Dataset.
image_type: An optional string indicating the type of image to use
when constructing feature-based images. See
meta.analyze_features() for details. By default, uses reverse
inference z-score images.
threshold: If decoding from a Dataset instance, this is the feature
threshold to use to generate the feature maps used in the
decoding.
"""
self.dataset = dataset
if dataset is not None:
self.masker = self.dataset.masker
if features is None:
features = dataset.get_feature_names()
if mask is not None:
self.masker.add(mask)
elif mask is not None:
self.masker = Masker(mask)
else:
self.masker = None
self.method = method.lower()
if self.method == 'roi':
self.feature_names = features
else:
self.load_features(features, image_type=image_type,
threshold=threshold)
def decode(self, images, save=None, round=4, names=None, **kwargs):
""" Decodes a set of images.
Args:
images: The images to decode. Can be:
- A single String specifying the filename of the image to decode
- A list of filenames
- A single NumPy array containing the image data
save: Optional filename to save results to. If None (default), returns
all results as an array.
round: Optional integer indicating number of decimals to round result
to. Defaults to 4.
names: Optional list of names corresponding to the images in filenames.
If passed, must be of same length and in same order as filenames.
By default, the columns in the output will be named using the image
filenames.
Returns:
An n_features x n_files numpy array, where each feature is a row and
each image is a column. The meaning of the values depends on the
decoding method used. """
if isinstance(images, string_types):
images = [images]
if isinstance(images, list):
imgs_to_decode = imageutils.load_imgs(images, self.masker)
else:
imgs_to_decode = images
methods = {
'pearson': self._pearson_correlation,
'dot': self._dot_product,
'roi': self._roi_association
}
result = np.around(
methods[self.method](imgs_to_decode, **kwargs), round)
# if save is not None:
if names is None:
if type(images).__module__ == np.__name__:
names = ['image_%d' % i for i in range(images.shape[1])]
elif self.method == 'roi':
names = ['cluster_%d' % i for i in range(result.shape[1])]
else:
names = images
result = pd.DataFrame(result, columns=names, index=self.feature_names)
if save is not None:
result.to_csv(save, index_label='Feature')
return result
def set_method(self, method):
""" Set decoding method. """
self.method = method
def load_features(self, features, image_type=None, from_array=False,
threshold=0.001):
""" Load features from current Dataset instance or a list of files.
Args:
features: List containing paths to, or names of, features to
extract. Each element in the list must be a string containing
either a path to an image, or the name of a feature (as named
in the current Dataset). Mixing of paths and feature names
within the list is not allowed.
image_type: Optional suffix indicating which kind of image to use
for analysis. Only used if features are taken from the Dataset;
if features is a list of filenames, image_type is ignored.
from_array: If True, the features argument is interpreted as a
string pointing to the location of a 2D ndarray on disk
containing feature data, where rows are voxels and columns are
individual features.
threshold: If features are taken from the dataset, this is the
threshold passed to the meta-analysis module to generate fresh
images.
"""
if from_array:
if isinstance(features, list):
features = features[0]
self._load_features_from_array(features)
elif path.exists(features[0]):
self._load_features_from_images(features)
else:
self._load_features_from_dataset(
features, image_type=image_type, threshold=threshold)
def _load_features_from_array(self, features):
""" Load feature data from a 2D ndarray on disk. """
self.feature_images = np.load(features)
self.feature_names = range(self.feature_images.shape[1])
def _load_features_from_dataset(self, features=None, image_type=None,
threshold=0.001):
""" Load feature image data from the current Dataset instance. See
load_features() for documentation.
"""
self.feature_names = self.dataset.feature_table.feature_names
if features is not None:
self.feature_names = [f for f in features if f in self.feature_names]
from neurosynth.analysis import meta
self.feature_images = meta.analyze_features(
self.dataset, self.feature_names, image_type=image_type,
threshold=threshold)
# Apply a mask if one was originally passed
if self.masker.layers:
in_mask = self.masker.get_mask(in_global_mask=True)
self.feature_images = self.feature_images[in_mask, :]
def _load_features_from_images(self, images, names=None):
""" Load feature image data from image files.
Args:
images: A list of image filenames.
names: An optional list of strings to use as the feature names. Must
be in the same order as the images.
"""
if names is not None and len(names) != len(images):
raise Exception(
"Lists of feature names and images must be of same length!")
self.feature_names = names if names is not None else images
self.feature_images = imageutils.load_imgs(images, self.masker)
def train_classifiers(self, features=None):
''' Train a set of classifiers '''
# for f in features:
# clf = Classifier(None)
# self.classifiers.append(clf)
pass
def _pearson_correlation(self, imgs_to_decode):
""" Decode images using Pearson's r.
Computes the correlation between each input image and each feature
image across voxels.
Args:
imgs_to_decode: An ndarray of images to decode, with voxels in rows
and images in columns.
Returns:
An n_features x n_images 2D array, with each cell representing the
pearson correlation between the i'th feature and the j'th image
across all voxels.
"""
x, y = imgs_to_decode.astype(float), self.feature_images.astype(float)
return self._xy_corr(x, y)
def _dot_product(self, imgs_to_decode):
""" Decoding using the dot product.
"""
return np.dot(imgs_to_decode.T, self.feature_images).T
def _roi_association(self, imgs_to_decode, value='z', binarize=None):
""" Computes the strength of association between activation in a mask
and presence/absence of a semantic feature. This is essentially a
generalization of the voxel-wise reverse inference z-score to the
multivoxel case.
"""
imgs_to_decode = imgs_to_decode.squeeze()
x = average_within_regions(self.dataset, imgs_to_decode).astype(float)
y = self.dataset.feature_table.data[self.feature_names].values
if binarize is not None:
y[y > binarize] = 1.
y[y < 1.] = 0.
r = self._xy_corr(x.T, y)
if value == 'r':
return r
elif value == 'z':
f_r = np.arctanh(r)
return f_r*np.sqrt(y.shape[0]-3)
def _xy_corr(self, x, y):
x, y = x - x.mean(0), y - y.mean(0)
x, y = x / np.sqrt((x ** 2).sum(0)), y / np.sqrt((y ** 2).sum(0))
return x.T.dot(y).T
def plot_polar(self, data, n_top=3, overplot=False, labels=None,
palette='husl'):
r = np.linspace(0, 10, num=100)
n_panels = data.shape[1]
if labels is None:
labels = []
for i in range(n_panels):
labels.extend(data.iloc[:, i].order(ascending=False)
.index[:n_top])
labels = np.unique(labels)
data = data.loc[labels, :]
# Use hierarchical clustering to order
from scipy.spatial.distance import pdist
from scipy.cluster.hierarchy import linkage, leaves_list
dists = pdist(data, metric='correlation')
pairs = linkage(dists)
order = leaves_list(pairs)
data = data.iloc[order, :]
labels = [labels[i] for i in order]
theta = np.linspace(0.0, 2 * np.pi, len(labels), endpoint=False)
if overplot:
fig, ax = plt.subplots(1, 1, subplot_kw=dict(polar=True))
fig.set_size_inches(10, 10)
else:
fig, axes = plt.subplots(1, n_panels, sharex=False, sharey=False,
subplot_kw=dict(polar=True))
fig.set_size_inches((6 * n_panels, 6))
# A bit silly to import seaborn just for this...
# should extract just the color_palette functionality.
import seaborn as sns
colors = sns.color_palette(palette, n_panels)
for i in range(n_panels):
if overplot:
alpha = 0.2
else:
ax = axes[i]
alpha = 0.8
ax.set_ylim(data.values.min(), data.values.max())
d = data.iloc[:, i].values
ax.fill(theta, d, color=colors[i], alpha=alpha, ec='k',
linewidth=0)
ax.fill(theta, d, alpha=1.0, ec=colors[i],
linewidth=2, fill=False)
ax.set_xticks(theta)
ax.set_xticklabels(labels, fontsize=18)
[lab.set_fontsize(18) for lab in ax.get_yticklabels()]
ax.set_title('Cluster %d' % i, fontsize=22, y=1.12)
plt.tight_layout()
return plt
def setUp(self):
""" Create a new Dataset and add features. """
maskfile = get_resource_path() + 'MNI152_T1_2mm_brain.nii.gz'
self.masker = Masker(maskfile)
class Decoder:
def __init__(self, dataset=None, method='pearson', features=None,
mask=None, image_type='pFgA_z', threshold=0.001):
""" Initialize a new Decoder instance.
Args:
dataset: An optional Dataset instance containing features to use in
decoding.
method: The decoding method to use (optional). By default, Pearson
correlation.
features: Optional list of features to use in decoding. If None,
use all features found in dataset. If features is a list of
strings, use only the subset of features in the Dataset that
are named in the list. If features is a list of filenames,
ignore the dataset entirely and use only the features passed as
image files in decoding.
mask: An optional mask to apply to features and input images. If
None, will use the one in the current Dataset.
image_type: An optional string indicating the type of image to use
when constructing feature-based images. See
meta.analyze_features() for details. By default, uses reverse
inference z-score images.
threshold: If decoding from a Dataset instance, this is the feature
threshold to use to generate the feature maps used in the
decoding.
"""
self.dataset = dataset
if dataset is not None:
self.masker = self.dataset.masker
if features is None:
features = dataset.get_feature_names()
if mask is not None:
self.masker.add(mask)
elif mask is not None:
self.masker = Masker(mask)
else:
self.masker = None
self.method = method.lower()
if self.method == 'roi':
self.feature_names = features
else:
self.load_features(features, image_type=image_type,
threshold=threshold)
def decode(self, images, save=None, round=4, names=None, **kwargs):
""" Decodes a set of images.
Args:
images: The images to decode. Can be:
- A single String specifying the filename of the image to decode
- A list of filenames
- A single NumPy array containing the image data
save: Optional filename to save results to. If None (default), returns
all results as an array.
round: Optional integer indicating number of decimals to round result
to. Defaults to 4.
names: Optional list of names corresponding to the images in filenames.
If passed, must be of same length and in same order as filenames.
By default, the columns in the output will be named using the image
filenames.
Returns:
An n_features x n_files numpy array, where each feature is a row and
each image is a column. The meaning of the values depends on the
decoding method used. """
if isinstance(images, basestring):
images = [images]
if isinstance(images, list):
imgs_to_decode = imageutils.load_imgs(images, self.masker)
else:
imgs_to_decode = images
methods = {
'pearson': self._pearson_correlation,
'dot': self._dot_product,
'roi': self._roi_association
}
result = np.around(
methods[self.method](imgs_to_decode, **kwargs), round)
# if save is not None:
if names is None:
if type(images).__module__ == np.__name__:
names = ['image_%d' % i for i in range(images.shape[1])]
elif self.method == 'roi':
names = ['cluster_%d' % i for i in range(result.shape[1])]
else:
names = images
result = pd.DataFrame(result, columns=names, index=self.feature_names)
if save is not None:
result.to_csv(save, index_label='Feature')
return result
def set_method(self, method):
""" Set decoding method. """
self.method = method
def load_features(self, features, image_type=None, from_array=False,
threshold=0.001):
""" Load features from current Dataset instance or a list of files.
Args:
features: List containing paths to, or names of, features to
extract. Each element in the list must be a string containing
either a path to an image, or the name of a feature (as named
in the current Dataset). Mixing of paths and feature names
within the list is not allowed.
image_type: Optional suffix indicating which kind of image to use
for analysis. Only used if features are taken from the Dataset;
if features is a list of filenames, image_type is ignored.
from_array: If True, the features argument is interpreted as a
string pointing to the location of a 2D ndarray on disk
containing feature data, where rows are voxels and columns are
individual features.
threshold: If features are taken from the dataset, this is the
threshold passed to the meta-analysis module to generate fresh
images.
"""
if from_array:
if isinstance(features, list):
features = features[0]
self._load_features_from_array(features)
elif path.exists(features[0]):
self._load_features_from_images(features)
else:
self._load_features_from_dataset(
features, image_type=image_type, threshold=threshold)
def _load_features_from_array(self, features):
""" Load feature data from a 2D ndarray on disk. """
self.feature_images = np.load(features)
self.feature_names = range(self.feature_images.shape[1])
def _load_features_from_dataset(self, features=None, image_type=None,
threshold=0.001):
""" Load feature image data from the current Dataset instance. See
load_features() for documentation.
"""
self.feature_names = self.dataset.feature_table.feature_names
if features is not None:
self.feature_names = filter(
lambda x: x in self.feature_names, features)
from neurosynth.analysis import meta
self.feature_images = meta.analyze_features(
self.dataset, self.feature_names, image_type=image_type,
threshold=threshold)
# Apply a mask if one was originally passed
if self.masker.layers:
in_mask = self.masker.get_mask(in_global_mask=True)
self.feature_images = self.feature_images[in_mask, :]
def _load_features_from_images(self, images, names=None):
""" Load feature image data from image files.
Args:
images: A list of image filenames.
names: An optional list of strings to use as the feature names. Must
be in the same order as the images.
"""
if names is not None and len(names) != len(images):
raise Exception(
"Lists of feature names and images must be of same length!")
self.feature_names = names if names is not None else images
self.feature_images = imageutils.load_imgs(images, self.masker)
def train_classifiers(self, features=None):
''' Train a set of classifiers '''
# for f in features:
# clf = Classifier(None)
# self.classifiers.append(clf)
pass
def _pearson_correlation(self, imgs_to_decode):
""" Decode images using Pearson's r.
Computes the correlation between each input image and each feature
image across voxels.
Args:
imgs_to_decode: An ndarray of images to decode, with voxels in rows
and images in columns.
Returns:
An n_features x n_images 2D array, with each cell representing the
pearson correlation between the i'th feature and the j'th image
across all voxels.
"""
x, y = imgs_to_decode.astype(float), self.feature_images.astype(float)
return self._xy_corr(x, y)
def _dot_product(self, imgs_to_decode):
""" Decoding using the dot product.
"""
return np.dot(imgs_to_decode.T, self.feature_images).T
def _roi_association(self, imgs_to_decode, value='z', binarize=None):
""" Computes the strength of association between activation in a mask
and presence/absence of a semantic feature. This is essentially a
generalization of the voxel-wise reverse inference z-score to the
multivoxel case.
"""
imgs_to_decode = imgs_to_decode.squeeze()
x = average_within_regions(self.dataset, imgs_to_decode).astype(float)
y = self.dataset.feature_table.data[self.feature_names].values
if binarize is not None:
y[y > binarize] = 1.
y[y < 1.] = 0.
r = self._xy_corr(x.T, y)
if value == 'r':
return r
elif value == 'z':
f_r = np.arctanh(r)
return f_r*np.sqrt(y.shape[0]-3)
def _xy_corr(self, x, y):
x, y = x - x.mean(0), y - y.mean(0)
x, y = x / np.sqrt((x ** 2).sum(0)), y / np.sqrt((y ** 2).sum(0))
return x.T.dot(y).T
def plot_polar(self, data, n_top=3, overplot=False, labels=None,
palette='husl'):
r = np.linspace(0, 10, num=100)
n_panels = data.shape[1]
if labels is None:
labels = []
for i in range(n_panels):
labels.extend(data.iloc[:, i].order(ascending=False)
.index[:n_top])
labels = np.unique(labels)
data = data.loc[labels, :]
# Use hierarchical clustering to order
from scipy.spatial.distance import pdist
from scipy.cluster.hierarchy import linkage, leaves_list
dists = pdist(data, metric='correlation')
pairs = linkage(dists)
order = leaves_list(pairs)
data = data.iloc[order, :]
labels = [labels[i] for i in order]
theta = np.linspace(0.0, 2 * np.pi, len(labels), endpoint=False)
if overplot:
fig, ax = plt.subplots(1, 1, subplot_kw=dict(polar=True))
fig.set_size_inches(10, 10)
else:
fig, axes = plt.subplots(n_panels, 1, sharex=False, sharey=False,
subplot_kw=dict(polar=True))
fig.set_size_inches((6, 6 * n_panels))
# A bit silly to import seaborn just for this...
# should extract just the color_palette functionality.
import seaborn as sns
colors = sns.color_palette(palette, n_panels)
for i in range(n_panels):
if overplot:
alpha = 0.2
else:
ax = axes[i]
alpha = 0.8
ax.set_ylim(data.values.min(), data.values.max())
d = data.iloc[:, i].values
ax.fill(
theta, d, ec='k', alpha=alpha, color=colors[i], linewidth=2)
ax.set_xticks(theta)
ax.set_xticklabels(labels, fontsize=18)
[lab.set_fontsize(18) for lab in ax.get_yticklabels()]
ax.set_title('Cluster %d' % i, fontsize=22, y=1.12)
plt.tight_layout()
return plt
class Decoder:
def __init__(self, dataset=None, method='pearson', features=None, mask=None, image_type='pFgA_z', threshold=0.001):
""" Initialize a new Decoder instance.
Args:
dataset: An optional Dataset instance containing features to use in decoding.
method: The decoding method to use (optional). By default, Pearson correlation.
features: Optional list of features to use in decoding. If None, use all
features found in dataset. If features is a list of strings, use only the
subset of features in the Dataset that are named in the list. If features
is a list of filenames, ignore the dataset entirely and use only the
features passed as image files in decoding.
mask: An optional mask to apply to features and input images. If None, will use
the one in the current Dataset.
image_type: An optional string indicating the type of image to use when constructing
feature-based images. See meta.analyze_features() for details. By default, uses
reverse inference z-score images.
threshold: If decoding from a Dataset instance, this is the feature threshold to
use to generate the feature maps used in the decoding.
"""
self.dataset = dataset
if dataset is not None:
self.masker = self.dataset.masker
if features is None:
features = dataset.get_feature_names()
if mask is not None:
self.masker.add(mask)
elif mask is not None:
self.masker = Masker(mask)
else:
self.masker = None
self.method = method.lower()
self.load_features(features, image_type=image_type, threshold=threshold)
def decode(self, images, save=None, round=4, names=None):
""" Decodes a set of images.
Args:
images: The images to decode. Can be:
- A single String specifying the filename of the image to decode
- A list of filenames
- A single NumPy array containing the image data
save: Optional filename to save results to. If None (default), returns
all results as an array.
round: Optional integer indicating number of decimals to round result
to. Defaults to 4.
names: Optional list of names corresponding to the images in filenames.
If passed, must be of same length and in same order as filenames.
By default, the columns in the output will be named using the image
filenames.
Returns:
An n_features x n_files numpy array, where each feature is a row and
each image is a column. The meaning of the values depends on the
decoding method used. """
if isinstance(images, basestring) or isinstance(images, list):
imgs_to_decode = imageutils.load_imgs(images, self.masker)
else:
imgs_to_decode = images
methods = {
'pearson': self._pearson_correlation(imgs_to_decode),
# 'nb': self._naive_bayes(imgs_to_decode),
'pattern': self._pattern_expression(imgs_to_decode)
}
result = np.around(methods[self.method], round)
if save is not None:
if names is None:
if type(images).__module__ == np.__name__:
names = ['image_%d' for i in range(images.shape[1])]
else:
names = images
rownames = np.array(
self.feature_names, dtype='|S32')[:, np.newaxis]
f = open(save, 'w')
f.write('\t'.join(['Feature'] + names) + '\n')
np.savetxt(f, np.hstack((
rownames, result)), fmt='%s', delimiter='\t')
else:
return result
def set_method(self, method):
""" Set decoding method. """
self.method = method
def load_features(self, features, image_type=None, from_array=False, threshold=0.001):
""" Load features from current Dataset instance or a list of files.
Args:
features: List containing paths to, or names of, features to extract.
Each element in the list must be a string containing either a path to an
image, or the name of a feature (as named in the current Dataset).
Mixing of paths and feature names within the list is not allowed.
image_type: Optional suffix indicating which kind of image to use for analysis.
Only used if features are taken from the Dataset; if features is a list
of filenames, image_type is ignored.
from_array: If True, the features argument is interpreted as a string pointing
to the location of a 2D ndarray on disk containing feature data, where
rows are voxels and columns are individual features.
threshold: If features are taken from the dataset, this is the threshold
passed to the meta-analysis module to generate fresh images.
"""
if from_array:
if isinstance(features, list):
features = features[0]
self._load_features_from_array(features)
elif path.exists(features[0]):
self._load_features_from_images(features)
else:
self._load_features_from_dataset(features, image_type=image_type, threshold=threshold)
def _load_features_from_array(self, features):
""" Load feature data from a 2D ndarray on disk. """
self.feature_images = np.load(features)
self.feature_names = range(self.feature_images.shape[1])
def _load_features_from_dataset(self, features=None, image_type=None, threshold=0.001):
""" Load feature image data from the current Dataset instance. See load_features()
for documentation.
"""
self.feature_names = self.dataset.feature_table.feature_names
if features is not None:
self.feature_names = filter(lambda x: x in self.feature_names, features)
from neurosynth.analysis import meta
self.feature_images = meta.analyze_features(
self.dataset, self.feature_names, image_type=image_type, threshold=threshold)
# Apply a mask if one was originally passed
if self.masker.layers:
in_mask = self.masker.get_current_mask(in_global_mask=True)
self.feature_images = self.feature_images[in_mask,:]
def _load_features_from_images(self, images, names=None):
""" Load feature image data from image files.
Args:
images: A list of image filenames.
names: An optional list of strings to use as the feature names. Must be
in the same order as the images.
"""
if names is not None and len(names) != len(images):
raise Exception( "Lists of feature names and image files must be of same length!")
self.feature_names = names if names is not None else images
self.feature_images = imageutils.load_imgs(images, self.masker)
def train_classifiers(self, features=None):
''' Train a set of classifiers '''
# for f in features:
# clf = Classifier(None)
# self.classifiers.append(clf)
pass
def _pearson_correlation(self, imgs_to_decode):
""" Decode images using Pearson's r.
Computes the correlation between each input image and each feature image across
voxels.
Args:
imgs_to_decode: An ndarray of images to decode, with voxels in rows and images
in columns.
Returns:
An n_features x n_images 2D array, with each cell representing the pearson
correlation between the i'th feature and the j'th image across all voxels.
"""
x, y = imgs_to_decode.astype(float), self.feature_images.astype(float)
x, y = x - x.mean(0), y - y.mean(0)
x, y = x / np.sqrt((x ** 2).sum(0)), y / np.sqrt((y ** 2).sum(0))
return x.T.dot(y).T
def _pattern_expression(self, imgs_to_decode):
""" Decode images using pattern expression. For explanation, see:
http://wagerlab.colorado.edu/wiki/doku.php/help/fmri_help/pattern_expression_and_connectivity
"""
return np.dot(imgs_to_decode.T, self.feature_images).T
class Decoder:
def __init__(self,
dataset=None,
method='pearson',
features=None,
mask=None,
image_type='pFgA_z',
threshold=0.001):
""" Initialize a new Decoder instance.
Args:
dataset: An optional Dataset instance containing features to use in decoding.
method: The decoding method to use (optional). By default, Pearson correlation.
features: Optional list of features to use in decoding. If None, use all
features found in dataset. If features is a list of strings, use only the
subset of features in the Dataset that are named in the list. If features
is a list of filenames, ignore the dataset entirely and use only the
features passed as image files in decoding.
mask: An optional mask to apply to features and input images. If None, will use
the one in the current Dataset.
image_type: An optional string indicating the type of image to use when constructing
feature-based images. See meta.analyze_features() for details. By default, uses
reverse inference z-score images.
threshold: If decoding from a Dataset instance, this is the feature threshold to
use to generate the feature maps used in the decoding.
"""
self.dataset = dataset
if dataset is not None:
self.masker = self.dataset.masker
if features is None:
features = dataset.get_feature_names()
if mask is not None:
self.masker.add(mask)
elif mask is not None:
self.masker = Masker(mask)
else:
self.masker = None
self.method = method.lower()
self.load_features(features,
image_type=image_type,
threshold=threshold)
def decode(self, images, save=None, round=4, names=None):
""" Decodes a set of images.
Args:
images: The images to decode. Can be:
- A single String specifying the filename of the image to decode
- A list of filenames
- A single NumPy array containing the image data
save: Optional filename to save results to. If None (default), returns
all results as an array.
round: Optional integer indicating number of decimals to round result
to. Defaults to 4.
names: Optional list of names corresponding to the images in filenames.
If passed, must be of same length and in same order as filenames.
By default, the columns in the output will be named using the image
filenames.
Returns:
An n_features x n_files numpy array, where each feature is a row and
each image is a column. The meaning of the values depends on the
decoding method used. """
if isinstance(images, basestring) or isinstance(images, list):
imgs_to_decode = imageutils.load_imgs(images, self.masker)
else:
imgs_to_decode = images
methods = {
'pearson': self._pearson_correlation(imgs_to_decode),
# 'nb': self._naive_bayes(imgs_to_decode),
'pattern': self._pattern_expression(imgs_to_decode)
}
result = np.around(methods[self.method], round)
if save is not None:
if names is None:
if type(images).__module__ == np.__name__:
names = ['image_%d' for i in range(images.shape[1])]
else:
names = images
rownames = np.array(self.feature_names, dtype='|S32')[:,
np.newaxis]
f = open(save, 'w')
f.write('\t'.join(['Feature'] + names) + '\n')
np.savetxt(f,
np.hstack((rownames, result)),
fmt='%s',
delimiter='\t')
else:
return result
def set_method(self, method):
""" Set decoding method. """
self.method = method
def load_features(self,
features,
image_type=None,
from_array=False,
threshold=0.001):
""" Load features from current Dataset instance or a list of files.
Args:
features: List containing paths to, or names of, features to extract.
Each element in the list must be a string containing either a path to an
image, or the name of a feature (as named in the current Dataset).
Mixing of paths and feature names within the list is not allowed.
image_type: Optional suffix indicating which kind of image to use for analysis.
Only used if features are taken from the Dataset; if features is a list
of filenames, image_type is ignored.
from_array: If True, the features argument is interpreted as a string pointing
to the location of a 2D ndarray on disk containing feature data, where
rows are voxels and columns are individual features.
threshold: If features are taken from the dataset, this is the threshold
passed to the meta-analysis module to generate fresh images.
"""
if from_array:
if isinstance(features, list):
features = features[0]
self._load_features_from_array(features)
elif path.exists(features[0]):
self._load_features_from_images(features)
else:
self._load_features_from_dataset(features,
image_type=image_type,
threshold=threshold)
def _load_features_from_array(self, features):
""" Load feature data from a 2D ndarray on disk. """
self.feature_images = np.load(features)
self.feature_names = range(self.feature_images.shape[1])
def _load_features_from_dataset(self,
features=None,
image_type=None,
threshold=0.001):
""" Load feature image data from the current Dataset instance. See load_features()
for documentation.
"""
self.feature_names = self.dataset.feature_table.feature_names
if features is not None:
self.feature_names = filter(lambda x: x in self.feature_names,
features)
from neurosynth.analysis import meta
self.feature_images = meta.analyze_features(self.dataset,
self.feature_names,
image_type=image_type,
threshold=threshold)
# Apply a mask if one was originally passed
if self.masker.layers:
in_mask = self.masker.get_current_mask(in_global_mask=True)
self.feature_images = self.feature_images[in_mask, :]
def _load_features_from_images(self, images, names=None):
""" Load feature image data from image files.
Args:
images: A list of image filenames.
names: An optional list of strings to use as the feature names. Must be
in the same order as the images.
"""
if names is not None and len(names) != len(images):
raise Exception(
"Lists of feature names and image files must be of same length!"
)
self.feature_names = names if names is not None else images
self.feature_images = imageutils.load_imgs(images, self.masker)
def train_classifiers(self, features=None):
''' Train a set of classifiers '''
# for f in features:
# clf = Classifier(None)
# self.classifiers.append(clf)
pass
def _pearson_correlation(self, imgs_to_decode):
""" Decode images using Pearson's r.
Computes the correlation between each input image and each feature image across
voxels.
Args:
imgs_to_decode: An ndarray of images to decode, with voxels in rows and images
in columns.
Returns:
An n_features x n_images 2D array, with each cell representing the pearson
correlation between the i'th feature and the j'th image across all voxels.
"""
x, y = imgs_to_decode.astype(float), self.feature_images.astype(float)
x, y = x - x.mean(0), y - y.mean(0)
x, y = x / np.sqrt((x**2).sum(0)), y / np.sqrt((y**2).sum(0))
return x.T.dot(y).T
def _pattern_expression(self, imgs_to_decode):
""" Decode images using pattern expression. For explanation, see:
http://wagerlab.colorado.edu/wiki/doku.php/help/fmri_help/pattern_expression_and_connectivity
"""
return np.dot(imgs_to_decode.T, self.feature_images).T
def __init__(self,
dataset,
cluster_on='coactivation',
global_mask=None,
roi_mask=None,
reference_mask=None,
features=None,
feature_threshold=0.0,
min_voxels_per_study=None,
min_studies_per_voxel=None,
dimension_reduction='ward',
n_components=500,
distance_metric='correlation',
clustering_method='ward',
output_dir='.',
prefix=None,
parcellation_kwargs={},
clustering_kwargs={}):
""" Initialize Clusterer.
Args:
dataset (Dataset): The Dataset instance to use for clustering.
cluster_on (str): The kind of data to use as the basis for voxel
clustering--i.e., what defines features passed to the clustering
algorithm. Valid options:
'studies': features are individual studies.
'coactivation': uses a precomputed distance matrix based on
coactivation of all voxels in the roi_mask across all
observations in the reference_data.
global_mask: An image defining the space to use for all analyses. If None, the
mask found in the Dataset will be used.
roi_mask: An image that determines which voxels to cluster. All non-zero voxels
will be included in the clustering analysis. When roi_mask is None, all
voxels in the global_mask (i.e., the whole brain) will be clustered. roi_mask
can be an image filename, a nibabel image, or an already-masked array with
the same dimensions as the global_mask.
reference_mask: An image defining the voxels to base the distance matrix
computation on. All non-zero voxels will be used to compute the distance
matrix. For example, if the roi_mask contains voxels in only the insula,
and reference_mask contains voxels in only the cerebellum, then voxels in
the insula will be clustered based on the similarity of their coactvation
with all and only cerebellum voxels.
features: Optional features to use for selecting a subset of the studies in the
Dataset instance. If dataset is a numpy matrix, will be ignored.
feature_threshold: float; the threshold to use for feature selection. Will be
ignored if features is None.
min_voxels_per_study: An optional integer. If provided, all voxels with fewer
than this number of studies will be removed from analysis.
min_studies_per_voxel: An optional integer. If provided, all studies with fewer
than this number of active voxels will be removed from analysis.
dimension_reduction: Either a scikit-learn object with a fit_transform method,
or the name of the parcellation method to use for reducing the dimensionality
of the reference mask. Valid options include:
None: no parcellation
'ward': spatially constrained hierarchical clustering; see Thirion
et al (2014)
'pca': principal component analysis
'grid': downsample the reference mask to an isometric grid
Defaults to 'ward'. Note that parcellation will only be used if method
is set to 'coactivation' (i.e., it will be ignored by default).
n_components: Number of components to request, if using a dimension_reduction method.
Meaning depends on parcellation algorithm.
distance_metric: Optional string providing the distance metric to use for
computation of a distance matrix. When None, no distance matrix is computed
and we assume that clustering will be done on the raw data. Valid options
are any of the strings accepted by sklearn's pairwise_distances method.
Defaults to 'correlation'. Note that for some clustering methods (e.g., k-means),
no distance matrix will be computed, and this argument will be ignored.
clustering_method: Algorithm to use for clustering. Must be one of 'ward', 'spectral',
'agglomerative', 'dbscan', 'kmeans', or 'minik'. If None, can be set
later via set_algorithm() or cluster().
output_directory: Directory to use for writing all outputs.
prefix: Optional prefix to prepend to all outputted directories/files.
parcellation_kwargs: Optional keyword arguments to pass to parcellation object.
clustering_kwargs: Optional keyword arguments to pass to clustering object.
"""
self.output_dir = output_dir
if not os.path.exists(output_dir):
os.makedirs(output_dir)
self.prefix = prefix
# Save all arguments for metadata output
self.args = {}
for a in ([
'output_dir', 'features', 'feature_threshold', 'global_mask',
'roi_mask', 'reference_mask', 'distance_metric'
] + clustering_kwargs.keys() + parcellation_kwargs.keys()):
self.args[a] = locals()[a]
self.set_algorithm(clustering_method, **clustering_kwargs)
self.dataset = dataset
self.masker = deepcopy(
dataset.masker) if global_mask is None else Masker(global_mask)
# Condition study inclusion on specific features
if features is not None:
data = self.dataset.get_ids_by_features(
features, threshold=feature_threshold, get_image_data=True)
else:
data = self.dataset.get_image_data()
# Trim data based on minimum number of voxels or studies
if min_studies_per_voxel is not None:
logger.info("Thresholding voxels based on number of studies.")
av = self.masker.unmask(data.sum(1) > min_studies_per_voxel,
output='vector')
self.masker.add(av)
if min_voxels_per_study is not None:
logger.info("Thresholding studies based on number of voxels.")
active_studies = np.where(data.sum(0) > min_voxels_per_study)[0]
data = data[:, active_studies]
self.data = data
self.set_reference_data(method=cluster_on, mask=reference_mask)
# Dimensionality reduction
if dimension_reduction is not None:
self.dimension_reduction(dimension_reduction, n_components)
# Set the voxels to cluster
if roi_mask is not None:
self.masker.add(roi_mask)
self.roi_data = data[self.masker.get_mask(), :]
# if roi_mask is not None: self.masker.remove(-1)
if distance_metric is not None:
self.create_distance_matrix(distance_metric=distance_metric)
def __init__(self,
dataset=None,
algorithm=None,
output_dir='.',
grid_scale=None,
features=None,
feature_threshold=0.0,
global_mask=None,
roi_mask=None,
distance_mask=None,
min_voxels_per_study=None,
min_studies_per_voxel=None,
distance_metric=None,
**kwargs):
""" Initialize Clusterer.
Args:
dataset: The dataset to use for clustering. Either a Dataset instance or a numpy
array with voxels in rows and features in columns.
algorithm: Optional algorithm to use for clustering. If None, an algorithm
must be passed to the cluster() method later.
output_directory: Directory to use for writing all outputs.
grid_scale: Optional integer. If provided, a 3D grid will be applied to the
image data, with values in all voxels in each grid cell being averaged
prior to clustering analysis. This is an effective means of dimension
reduction in cases where the data are otherwise too large for clustering.
features: Optional features to use for selecting a subset of the studies in the
Dataset instance. If dataset is a numpy matrix, will be ignored.
feature_threshold: float; the threshold to use for feature selection. Will be
ignored if features is None.
global_mask: An image defining the space to use for all analyses. Only necessary
if dataset is a numpy array.
roi_mask: An image that determines which voxels to cluster. All non-zero voxels
will be included in the clustering analysis. When roi_mask is None, all
voxels in the global_mask (i.e., the whole brain) will be clustered. roi_mask
can be an image filename, a nibabel image, or an already-masked array with
the same dimensions as the global_mask.
distance_mask: An image defining the voxels to base the distance matrix
computation on. All non-zero voxels will be used to compute the distance
matrix. For example, if the roi_mask contains voxels in only the insula,
and distance_mask contains voxels in only the cerebellum, then voxels in
the insula will be clustered based on the similarity of their coactvation
with all and only cerebellum voxels.
min_voxels_per_study: An optional integer. If provided, all voxels with fewer
than this number of studies will be removed from analysis.
min_studies_per_voxel: An optional integer. If provided, all studies with fewer
than this number of active voxels will be removed from analysis.
distance_metric: Optional string providing the distance metric to use for
computation of a distance matrix. When None, no distance matrix is computed
and we assume that clustering will be done on the raw data.
**kwargs: Additional keyword arguments to pass to the clustering algorithm.
"""
self.output_dir = output_dir
if algorithm is not None:
self._set_clustering_algorithm(algorithm, **kwargs)
if isinstance(dataset, Dataset):
self.dataset = dataset
if global_mask is None:
global_mask = dataset.masker
if features is not None:
data = self.dataset.get_ids_by_features(
features, threshold=feature_threshold, get_image_data=True)
else:
data = self.dataset.get_image_data()
# if min_studies_per_voxel is not None:
# logger.info("Thresholding voxels based on number of studies.")
# sum_vox = data.sum(1)
# # Save the indices for later reconstruction
# active_vox = np.where(sum_vox > min_studies_per_voxel)[0]
# n_active_vox = active_vox.shape[0]
# if min_voxels_per_study is not None:
# logger.info("Thresholding studies based on number of voxels.")
# sum_studies = data.sum(0)
# active_studies = np.where(sum_studies > min_voxels_per_study)[0]
# n_active_studies = active_studies.shape[0]
# if min_studies_per_voxel is not None:
# logger.info("Selecting voxels with more than %d studies." % min_studies_per_voxel)
# data = data[active_vox, :]
# if min_voxels_per_study is not None:
# logger.info("Selecting studies with more than %d voxels." % min_voxels_per_study)
# data = data[:, active_studies]
self.data = data
else:
self.data = dataset
if global_mask is None:
raise ValueError(
"If dataset is a numpy array, a valid global_mask (filename, "
+ "Mask instance, or nibabel image) must be passed.")
if not isinstance(global_mask, Masker):
global_mask = Masker(global_mask)
self.masker = global_mask
if distance_mask is not None:
self.masker.add(distance_mask)
if grid_scale is not None:
self.target_data, _ = nsr.apply_grid(self.data,
masker=self.masker,
scale=grid_scale,
threshold=None)
else:
vox = self.masker.get_current_mask(in_global_mask=True)
self.target_data = self.data[vox, :]
self.masker.reset()
if roi_mask is not None:
self.masker.add(roi_mask)
if grid_scale is not None:
self.data, self.grid = nsr.apply_grid(self.data,
masker=self.masker,
scale=grid_scale,
threshold=None)
else:
vox = self.masker.get_current_mask(in_global_mask=True)
self.data = self.data[vox, :]
if distance_metric is not None:
self.create_distance_matrix(distance_metric=distance_metric)