/
discrete.py
executable file
·322 lines (267 loc) · 13.4 KB
/
discrete.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
"""
Methods for decoding subsets of voxels (e.g., ROIs) or experiments (e.g., from
meta-analytic clustering on a database) into text.
"""
import numpy as np
import pandas as pd
import nibabel as nib
from scipy.stats import binom
from scipy import special
from statsmodels.sandbox.stats.multicomp import multipletests
from .utils import weight_priors
from ..stats import p_to_z, one_way, two_way
from ..due import due
from .. import references
@due.dcite(references.GCLDA_DECODING, description='Citation for GCLDA decoding.')
def gclda_decode_roi(model, roi, topic_priors=None, prior_weight=1.):
r"""
Perform image-to-text decoding for discrete image inputs (e.g., regions
of interest, significant clusters) according to the method described in
[1]_.
Parameters
----------
model : :obj:`nimare.annotate.topic.GCLDAModel`
Model object needed for decoding.
roi : :obj:`nibabel.nifti1.Nifti1Image` or :obj:`str`
Binary image to decode into text. If string, path to a file with
the binary image.
topic_priors : :obj:`numpy.ndarray` of :obj:`float`, optional
A 1d array of size (n_topics) with values for topic weighting.
If None, no weighting is done. Default is None.
prior_weight : :obj:`float`, optional
The weight by which the prior will affect the decoding.
Default is 1.
Returns
-------
decoded_df : :obj:`pandas.DataFrame`
A DataFrame with the word-tokens and their associated weights.
topic_weights : :obj:`numpy.ndarray` of :obj:`float`
The weights of the topics used in decoding.
Notes
-----
====================== ==============================================================
Notation Meaning
====================== ==============================================================
:math:`v` Voxel
:math:`t` Topic
:math:`w` Word type
:math:`r` Region of interest (ROI)
:math:`p(v|t)` Probability of topic given voxel (``p_topic_g_voxel``)
:math:`\\tau_{t}` Topic weight vector (``topic_weights``)
:math:`p(w|t)` Probability of word type given topic (``p_word_g_topic``)
====================== ==============================================================
1. Compute
:math:`p(v|t)`.
- From :obj:`gclda.model.Model.get_spatial_probs()`
2. Compute topic weight vector (:math:`\\tau_{t}`) by adding across voxels
within ROI.
- :math:`\\tau_{t} = \sum_{i} {p(t|v_{i})}`
3. Multiply :math:`\\tau_{t}` by
:math:`p(w|t)`.
- :math:`p(w|r) \propto \\tau_{t} \cdot p(w|t)`
4. The resulting vector (``word_weights``) reflects arbitrarily scaled
term weights for the ROI.
References
----------
.. [1] Rubin, Timothy N., et al. "Decoding brain activity using a
large-scale probabilistic functional-anatomical atlas of human
cognition." PLoS computational biology 13.10 (2017): e1005649.
https://doi.org/10.1371/journal.pcbi.1005649
"""
if isinstance(roi, str):
roi = nib.load(roi)
elif not isinstance(roi, nib.Nifti1Image):
raise IOError('Input roi must be either a nifti image '
'(nibabel.Nifti1Image) or a path to one.')
dset_aff = model.mask.affine
if not np.array_equal(roi.affine, dset_aff):
raise ValueError('Input roi must have same affine as mask img:'
'\n{0}\n{1}'.format(np.array2string(roi.affine),
np.array2string(dset_aff)))
# Load ROI file and get ROI voxels overlapping with brain mask
mask_vec = model.mask.get_data().ravel().astype(bool)
roi_vec = roi.get_data().astype(bool).ravel()
roi_vec = roi_vec[mask_vec]
roi_idx = np.where(roi_vec)[0]
p_topic_g_roi = model.p_topic_g_voxel[roi_idx, :] # p(T|V) for voxels in ROI only
topic_weights = np.sum(p_topic_g_roi, axis=0) # Sum across words
if topic_priors is not None:
weighted_priors = weight_priors(topic_priors, prior_weight)
topic_weights *= weighted_priors
# Multiply topic_weights by topic-by-word matrix (p_word_g_topic).
# n_word_tokens_per_topic = np.sum(model.n_word_tokens_word_by_topic, axis=0)
# p_word_g_topic = model.n_word_tokens_word_by_topic / n_word_tokens_per_topic[None, :]
# p_word_g_topic = np.nan_to_num(p_word_g_topic, 0)
word_weights = np.dot(model.p_word_g_topic, topic_weights)
decoded_df = pd.DataFrame(index=model.vocabulary,
columns=['Weight'], data=word_weights)
decoded_df.index.name = 'Term'
return decoded_df, topic_weights
@due.dcite(references.BRAINMAP_DECODING,
description='Citation for BrainMap-style decoding.')
def brainmap_decode(coordinates, annotations, ids, ids2=None, features=None,
frequency_threshold=0.001, u=0.05, correction='fdr_bh'):
"""
Perform image-to-text decoding for discrete image inputs (e.g., regions
of interest, significant clusters) according to the BrainMap method [1]_.
References
----------
.. [1] Amft, Maren, et al. "Definition and characterization of an extended
social-affective default network." Brain Structure and Function 220.2
(2015): 1031-1049. https://doi.org/10.1007/s00429-013-0698-0
"""
id_cols = ['id', 'study_id', 'contrast_id']
dataset_ids = sorted(list(set(coordinates['id'].values)))
if ids2 is None:
unselected = sorted(list(set(dataset_ids) - set(ids)))
else:
unselected = ids2[:]
if features is None:
features = annotations.columns.values
features = [f for f in features if f not in id_cols]
# Binarize with frequency threshold
features_df = annotations.set_index('id', drop=True)
features_df = features_df[features].ge(frequency_threshold)
sel_array = features_df.loc[ids].values
unsel_array = features_df.loc[unselected].values
n_selected = len(ids)
n_unselected = len(unselected)
# the number of times any term is used (e.g., if one experiment uses
# two terms, that counts twice). Why though?
n_exps_across_terms = np.sum(np.sum(features_df))
n_selected_term = np.sum(sel_array, axis=0)
n_unselected_term = np.sum(unsel_array, axis=0)
n_selected_noterm = n_selected - n_selected_term
n_unselected_noterm = n_unselected - n_unselected_term
n_term = n_selected_term + n_unselected_term
p_term = n_term / n_exps_across_terms
n_foci_in_database = coordinates.shape[0]
p_selected = n_selected / n_foci_in_database
# I hope there's a way to do this without the for loop
n_term_foci = np.zeros(len(features))
n_noterm_foci = np.zeros(len(features))
for i, term in enumerate(features):
term_ids = features_df.loc[features_df[term] == 1].index.values
noterm_ids = features_df.loc[features_df[term] == 0].index.values
n_term_foci[i] = coordinates['id'].isin(term_ids).sum()
n_noterm_foci[i] = coordinates['id'].isin(noterm_ids).sum()
p_selected_g_term = n_selected_term / n_term_foci # probForward
l_selected_g_term = p_selected_g_term / p_selected # likelihoodForward
p_selected_g_noterm = n_selected_noterm / n_noterm_foci
p_term_g_selected = p_selected_g_term * p_term / p_selected # probReverse
p_term_g_selected = p_term_g_selected / np.sum(p_term_g_selected) # Normalize
# Significance testing
# Forward inference significance is determined with a binomial distribution
p_fi = 1 - binom.cdf(k=n_selected_term, n=n_term_foci, p=p_selected)
sign_fi = np.sign(n_selected_term - np.mean(n_selected_term)).ravel() # pylint: disable=no-member
# Two-way chi-square test for specificity of activation
cells = np.array([[n_selected_term, n_selected_noterm], # pylint: disable=no-member
[n_unselected_term, n_unselected_noterm]]).T
chi2_ri = two_way(cells)
p_ri = special.chdtrc(1, chi2_ri)
sign_ri = np.sign(p_selected_g_term - p_selected_g_noterm).ravel() # pylint: disable=no-member
# Ignore rare features
p_fi[n_selected_term < 5] = 1.
p_ri[n_selected_term < 5] = 1.
# Multiple comparisons correction across features. Separately done for FI and RI.
if correction is not None:
_, p_corr_fi, _, _ = multipletests(p_fi, alpha=u, method=correction,
returnsorted=False)
_, p_corr_ri, _, _ = multipletests(p_ri, alpha=u, method=correction,
returnsorted=False)
else:
p_corr_fi = p_fi
p_corr_ri = p_ri
# Compute z-values
z_corr_fi = p_to_z(p_corr_fi, 'two') * sign_fi
z_corr_ri = p_to_z(p_corr_ri, 'two') * sign_ri
# Effect size
arr = np.array([p_corr_fi, z_corr_fi, l_selected_g_term, # pylint: disable=no-member
p_corr_ri, z_corr_ri, p_term_g_selected]).T
out_df = pd.DataFrame(data=arr, index=features,
columns=['pForward', 'zForward', 'likelihoodForward',
'pReverse', 'zReverse', 'probReverse'])
out_df.index.name = 'Term'
return out_df
@due.dcite(references.NEUROSYNTH, description='Introduces Neurosynth.')
def neurosynth_decode(coordinates, annotations, ids, ids2=None, features=None,
frequency_threshold=0.001, prior=0.5, u=0.05,
correction='fdr_bh'):
"""
Perform discrete functional decoding according to Neurosynth's
meta-analytic method [1]_. This does not employ correlations between
unthresholded maps, which are the method of choice for decoding within
Neurosynth and Neurovault.
Metadata (i.e., feature labels) for studies within the selected sample
(`ids`) are compared to the unselected studies remaining in the database
(`dataset`).
References
----------
.. [1] Yarkoni, Tal, et al. "Large-scale automated synthesis of human
functional neuroimaging data." Nature methods 8.8 (2011): 665.
https://doi.org/10.1038/nmeth.1635
"""
id_cols = ['id', 'study_id', 'contrast_id']
dataset_ids = sorted(list(set(coordinates['id'].values)))
if ids2 is None:
unselected = sorted(list(set(dataset_ids) - set(ids)))
else:
unselected = ids2[:]
if features is None:
features = annotations.columns.values
features = [f for f in features if f not in id_cols]
# Binarize with frequency threshold
features_df = annotations.set_index('id', drop=True)
features_df = features_df[features].ge(frequency_threshold)
sel_array = features_df.loc[ids].values
unsel_array = features_df.loc[unselected].values
n_selected = len(ids)
n_unselected = len(unselected)
n_selected_term = np.sum(sel_array, axis=0)
n_unselected_term = np.sum(unsel_array, axis=0)
n_selected_noterm = n_selected - n_selected_term
n_unselected_noterm = n_unselected - n_unselected_term
n_term = n_selected_term + n_unselected_term
n_noterm = n_selected_noterm + n_unselected_noterm
p_term = n_term / (n_term + n_noterm)
p_selected_g_term = n_selected_term / n_term
p_selected_g_noterm = n_selected_noterm / n_noterm
# Recompute conditions with empirically derived prior (or inputted one)
if prior is None:
# if this is used, p_term_g_selected_prior = p_selected (regardless of term)
prior = p_term
# Significance testing
# One-way chi-square test for consistency of term frequency across terms
chi2_fi = one_way(n_selected_term, n_term)
p_fi = special.chdtrc(1, chi2_fi)
sign_fi = np.sign(n_selected_term - np.mean(n_selected_term)).ravel() # pylint: disable=no-member
# Two-way chi-square test for specificity of activation
cells = np.array([[n_selected_term, n_selected_noterm], # pylint: disable=no-member
[n_unselected_term, n_unselected_noterm]]).T
chi2_ri = two_way(cells)
p_ri = special.chdtrc(1, chi2_ri)
sign_ri = np.sign(p_selected_g_term - p_selected_g_noterm).ravel() # pylint: disable=no-member
# Multiple comparisons correction across terms. Separately done for FI and RI.
if correction is not None:
_, p_corr_fi, _, _ = multipletests(p_fi, alpha=u, method=correction,
returnsorted=False)
_, p_corr_ri, _, _ = multipletests(p_ri, alpha=u, method=correction,
returnsorted=False)
else:
p_corr_fi = p_fi
p_corr_ri = p_ri
# Compute z-values
z_corr_fi = p_to_z(p_corr_fi, 'two') * sign_fi
z_corr_ri = p_to_z(p_corr_ri, 'two') * sign_ri
# Effect size
# est. prob. of brain state described by term finding activation in ROI
p_selected_g_term_g_prior = prior * p_selected_g_term + (1 - prior) * p_selected_g_noterm
# est. prob. of activation in ROI reflecting brain state described by term
p_term_g_selected_g_prior = p_selected_g_term * prior / p_selected_g_term_g_prior
arr = np.array([p_corr_fi, z_corr_fi, p_selected_g_term_g_prior, # pylint: disable=no-member
p_corr_ri, z_corr_ri, p_term_g_selected_g_prior]).T
out_df = pd.DataFrame(data=arr, index=features,
columns=['pForward', 'zForward', 'probForward',
'pReverse', 'zReverse', 'probReverse'])
out_df.index.name = 'Term'
return out_df