|
| 1 | +# -*- coding: utf-8 -*- |
| 2 | +# emacs: -*- mode: python; py-indent-offset: 4; indent-tabs-mode: nil -*- |
| 3 | +# vi: set ft=python sts=4 ts=4 sw=4 et: |
| 4 | +''' |
| 5 | +Miscellaneous algorithms |
| 6 | +
|
| 7 | + Change directory to provide relative paths for doctests |
| 8 | + >>> import os |
| 9 | + >>> filepath = os.path.dirname(os.path.realpath(__file__)) |
| 10 | + >>> datadir = os.path.realpath(os.path.join(filepath, '../testing/data')) |
| 11 | + >>> os.chdir(datadir) |
| 12 | +
|
| 13 | +''' |
| 14 | +from ..interfaces.base import (BaseInterfaceInputSpec, TraitedSpec, |
| 15 | + BaseInterface, traits, File) |
| 16 | +from ..pipeline import engine as pe |
| 17 | +from ..interfaces.utility import IdentityInterface |
| 18 | +from .misc import regress_poly |
| 19 | + |
| 20 | +import nibabel as nb |
| 21 | +import numpy as np |
| 22 | +from scipy import linalg, stats |
| 23 | +import os |
| 24 | + |
| 25 | +class CompCorInputSpec(BaseInterfaceInputSpec): |
| 26 | + realigned_file = File(exists=True, mandatory=True, |
| 27 | + desc='already realigned brain image (4D)') |
| 28 | + mask_file = File(exists=True, mandatory=False, |
| 29 | + desc='mask file that determines ROI (3D)') |
| 30 | + components_file = File('components_file.txt', exists=False, |
| 31 | + mandatory=False, usedefault=True, |
| 32 | + desc='filename to store physiological components') |
| 33 | + num_components = traits.Int(6, usedefault=True) # 6 for BOLD, 4 for ASL |
| 34 | + use_regress_poly = traits.Bool(True, usedefault=True, |
| 35 | + desc='use polynomial regression' |
| 36 | + 'pre-component extraction') |
| 37 | + regress_poly_degree = traits.Range(low=1, default=1, usedefault=True, |
| 38 | + desc='the degree polynomial to use') |
| 39 | + |
| 40 | +class CompCorOutputSpec(TraitedSpec): |
| 41 | + components_file = File(exists=True, |
| 42 | + desc='text file containing the noise components') |
| 43 | + |
| 44 | +class CompCor(BaseInterface): |
| 45 | + ''' |
| 46 | + Interface with core CompCor computation, used in aCompCor and tCompCor |
| 47 | +
|
| 48 | + Example |
| 49 | + ------- |
| 50 | +
|
| 51 | + >>> ccinterface = CompCor() |
| 52 | + >>> ccinterface.inputs.realigned_file = 'functional.nii' |
| 53 | + >>> ccinterface.inputs.mask_file = 'mask.nii' |
| 54 | + >>> ccinterface.inputs.num_components = 1 |
| 55 | + >>> ccinterface.inputs.use_regress_poly = True |
| 56 | + >>> ccinterface.inputs.regress_poly_degree = 2 |
| 57 | + ''' |
| 58 | + input_spec = CompCorInputSpec |
| 59 | + output_spec = CompCorOutputSpec |
| 60 | + |
| 61 | + def _run_interface(self, runtime): |
| 62 | + imgseries = nb.load(self.inputs.realigned_file).get_data() |
| 63 | + mask = nb.load(self.inputs.mask_file).get_data() |
| 64 | + voxel_timecourses = imgseries[mask > 0] |
| 65 | + # Zero-out any bad values |
| 66 | + voxel_timecourses[np.isnan(np.sum(voxel_timecourses, axis=1)), :] = 0 |
| 67 | + |
| 68 | + # from paper: |
| 69 | + # "The constant and linear trends of the columns in the matrix M were |
| 70 | + # removed [prior to ...]" |
| 71 | + if self.inputs.use_regress_poly: |
| 72 | + voxel_timecourses = regress_poly(self.inputs.regress_poly_degree, |
| 73 | + voxel_timecourses) |
| 74 | + voxel_timecourses = voxel_timecourses - np.mean(voxel_timecourses, |
| 75 | + axis=1)[:, np.newaxis] |
| 76 | + |
| 77 | + # "Voxel time series from the noise ROI (either anatomical or tSTD) were |
| 78 | + # placed in a matrix M of size Nxm, with time along the row dimension |
| 79 | + # and voxels along the column dimension." |
| 80 | + M = voxel_timecourses.T |
| 81 | + numvols = M.shape[0] |
| 82 | + numvoxels = M.shape[1] |
| 83 | + |
| 84 | + # "[... were removed] prior to column-wise variance normalization." |
| 85 | + M = M / self._compute_tSTD(M, 1.) |
| 86 | + |
| 87 | + # "The covariance matrix C = MMT was constructed and decomposed into its |
| 88 | + # principal components using a singular value decomposition." |
| 89 | + u, _, _ = linalg.svd(M, full_matrices=False) |
| 90 | + components = u[:, :self.inputs.num_components] |
| 91 | + components_file = os.path.join(os.getcwd(), self.inputs.components_file) |
| 92 | + np.savetxt(components_file, components, fmt="%.10f") |
| 93 | + return runtime |
| 94 | + |
| 95 | + def _list_outputs(self): |
| 96 | + outputs = self._outputs().get() |
| 97 | + outputs['components_file'] = os.path.abspath(self.inputs.components_file) |
| 98 | + return outputs |
| 99 | + |
| 100 | + def _compute_tSTD(self, M, x): |
| 101 | + stdM = np.std(M, axis=0) |
| 102 | + # set bad values to x |
| 103 | + stdM[stdM == 0] = x |
| 104 | + stdM[np.isnan(stdM)] = x |
| 105 | + return stdM |
| 106 | + |
| 107 | +class TCompCorInputSpec(CompCorInputSpec): |
| 108 | + # and all the fields in CompCorInputSpec |
| 109 | + percentile_threshold = traits.Range(low=0., high=1., value=.02, |
| 110 | + exclude_low=True, exclude_high=True, |
| 111 | + usedefault=True, desc='the percentile ' |
| 112 | + 'used to select highest-variance ' |
| 113 | + 'voxels, represented by a number ' |
| 114 | + 'between 0 and 1, exclusive. By ' |
| 115 | + 'default, this value is set to .02. ' |
| 116 | + 'That is, the 2% of voxels ' |
| 117 | + 'with the highest variance are used.') |
| 118 | + |
| 119 | +class TCompCor(CompCor): |
| 120 | + ''' |
| 121 | + Interface for tCompCor. Computes a ROI mask based on variance of voxels. |
| 122 | +
|
| 123 | + Example |
| 124 | + ------- |
| 125 | +
|
| 126 | + >>> ccinterface = TCompCor() |
| 127 | + >>> ccinterface.inputs.realigned_file = 'functional.nii' |
| 128 | + >>> ccinterface.inputs.mask_file = 'mask.nii' |
| 129 | + >>> ccinterface.inputs.num_components = 1 |
| 130 | + >>> ccinterface.inputs.use_regress_poly = True |
| 131 | + >>> ccinterface.inputs.regress_poly_degree = 2 |
| 132 | + >>> ccinterface.inputs.percentile_threshold = .03 |
| 133 | + ''' |
| 134 | + |
| 135 | + input_spec = TCompCorInputSpec |
| 136 | + output_spec = CompCorOutputSpec |
| 137 | + |
| 138 | + def _run_interface(self, runtime): |
| 139 | + imgseries = nb.load(self.inputs.realigned_file).get_data() |
| 140 | + |
| 141 | + # From the paper: |
| 142 | + # "For each voxel time series, the temporal standard deviation is |
| 143 | + # defined as the standard deviation of the time series after the removal |
| 144 | + # of low-frequency nuisance terms (e.g., linear and quadratic drift)." |
| 145 | + imgseries = regress_poly(2, imgseries) |
| 146 | + imgseries = imgseries - np.mean(imgseries, axis=1)[:, np.newaxis] |
| 147 | + |
| 148 | + time_voxels = imgseries.T |
| 149 | + num_voxels = np.prod(time_voxels.shape[1:]) |
| 150 | + |
| 151 | + # "To construct the tSTD noise ROI, we sorted the voxels by their |
| 152 | + # temporal standard deviation ..." |
| 153 | + tSTD = self._compute_tSTD(time_voxels, 0) |
| 154 | + sortSTD = np.sort(tSTD, axis=None) # flattened sorted matrix |
| 155 | + |
| 156 | + # use percentile_threshold to pick voxels |
| 157 | + threshold_index = int(num_voxels * (1. - self.inputs.percentile_threshold)) |
| 158 | + threshold_std = sortSTD[threshold_index] |
| 159 | + mask = tSTD >= threshold_std |
| 160 | + mask = mask.astype(int) |
| 161 | + |
| 162 | + # save mask |
| 163 | + mask_file = 'mask.nii' |
| 164 | + nb.nifti1.save(nb.Nifti1Image(mask, np.eye(4)), mask_file) |
| 165 | + self.inputs.mask_file = mask_file |
| 166 | + |
| 167 | + super(TCompCor, self)._run_interface(runtime) |
| 168 | + return runtime |
| 169 | + |
| 170 | +ACompCor = CompCor |
0 commit comments