Source code for dipy.tracking.life

"""
This is an implementation of the Linear Fascicle Evaluation (LiFE) algorithm
described in :footcite:p:`Pestilli2014`.

References
----------
.. footbibliography::
"""

import numpy as np
import scipy.linalg as la
import scipy.sparse as sps

import dipy.core.optimize as opt
import dipy.data as dpd
from dipy.reconst.base import ReconstFit, ReconstModel
from dipy.testing.decorators import warning_for_keywords
from dipy.tracking.streamline import transform_streamlines
from dipy.tracking.utils import unique_rows
from dipy.tracking.vox2track import _voxel2streamline


[docs] def gradient(f): """ Return the gradient of an N-dimensional array. The gradient is computed using central differences in the interior and first differences at the boundaries. The returned gradient hence has the same shape as the input array. Parameters ---------- f : array_like An N-dimensional array containing samples of a scalar function. Returns ------- gradient : ndarray N arrays of the same shape as `f` giving the derivative of `f` with respect to each dimension. Examples -------- >>> x = np.array([1, 2, 4, 7, 11, 16], dtype=float) >>> gradient(x) array([ 1. , 1.5, 2.5, 3.5, 4.5, 5. ]) >>> gradient(np.array([[1, 2, 6], [3, 4, 5]], dtype=float)) [array([[ 2., 2., -1.], [ 2., 2., -1.]]), array([[ 1. , 2.5, 4. ], [ 1. , 1. , 1. ]])] Notes ----- This is a simplified implementation of gradient that is part of numpy 1.8. In order to mitigate the effects of changes added to this implementation in version 1.9 of numpy, we include this implementation here. """ f = np.asanyarray(f) N = len(f.shape) # number of dimensions dx = [1.0] * N # use central differences on interior and first differences on endpoints outvals = [] # create slice objects --- initially all are [:, :, ..., :] slice1 = [slice(None)] * N slice2 = [slice(None)] * N slice3 = [slice(None)] * N for axis in range(N): # select out appropriate parts for this dimension out = np.empty_like(f) slice1[axis] = slice(1, -1) slice2[axis] = slice(2, None) slice3[axis] = slice(None, -2) # 1D equivalent -- out[1:-1] = (f[2:] - f[:-2])/2.0 out[tuple(slice1)] = (f[tuple(slice2)] - f[tuple(slice3)]) / 2.0 slice1[axis] = 0 slice2[axis] = 1 slice3[axis] = 0 # 1D equivalent -- out[0] = (f[1] - f[0]) out[tuple(slice1)] = f[tuple(slice2)] - f[tuple(slice3)] slice1[axis] = -1 slice2[axis] = -1 slice3[axis] = -2 # 1D equivalent -- out[-1] = (f[-1] - f[-2]) out[tuple(slice1)] = f[tuple(slice2)] - f[tuple(slice3)] # divide by step size outvals.append(out / dx[axis]) # reset the slice object in this dimension to ":" slice1[axis] = slice(None) slice2[axis] = slice(None) slice3[axis] = slice(None) if N == 1: return outvals[0] else: return outvals
[docs] def streamline_gradients(streamline): """ Calculate the gradients of the streamline along the spatial dimension Parameters ---------- streamline : array-like of shape (n, 3) The 3d coordinates of a single streamline Returns ------- Array of shape (3, n): Spatial gradients along the length of the streamline. """ return np.array(gradient(np.asarray(streamline))[0])
[docs] def grad_tensor(grad, evals): """ Calculate the 3 by 3 tensor for a given spatial gradient, given a canonical tensor shape (also as a 3 by 3), pointing at [1,0,0] Parameters ---------- grad : 1d array of shape (3,) The spatial gradient (e.g between two nodes of a streamline). evals: 1d array of shape (3,) The eigenvalues of a canonical tensor to be used as a response function. """ # This is the rotation matrix from [1, 0, 0] to this gradient of the sl: R = la.svd([grad], overwrite_a=True)[2] # This is the 3 by 3 tensor after rotation: T = np.linalg.multi_dot([R, np.diag(evals), R.T]) return T
[docs] @warning_for_keywords() def streamline_tensors(streamline, *, evals=(0.001, 0, 0)): """ The tensors generated by this fiber. Parameters ---------- streamline : array-like of shape (n, 3) The 3d coordinates of a single streamline evals : iterable with three entries, optional The estimated eigenvalues of a single fiber tensor. Returns ------- An n_nodes by 3 by 3 array with the tensor for each node in the fiber. Notes ----- Estimates of the radial/axial diffusivities may rely on empirical measurements (for example, the AD in the Corpus Callosum), or may be based on a biophysical model of some kind. """ grad = streamline_gradients(streamline) # Preallocate: tensors = np.empty((grad.shape[0], 3, 3)) for grad_idx, this_grad in enumerate(grad): tensors[grad_idx] = grad_tensor(this_grad, evals) return tensors
[docs] @warning_for_keywords() def streamline_signal(streamline, gtab, *, evals=(0.001, 0, 0)): """ The signal from a single streamline estimate along each of its nodes. Parameters ---------- streamline : a single streamline Streamline data. gtab : GradientTable class instance Gradient table. evals : array-like of length 3, optional The eigenvalues of the canonical tensor used as an estimate of the signal generated by each node of the streamline. """ # Gotta have those tensors: tensors = streamline_tensors(streamline, evals=evals) sig = np.empty((len(streamline), np.sum(~gtab.b0s_mask))) # Extract them once: bvecs = gtab.bvecs[~gtab.b0s_mask] bvals = gtab.bvals[~gtab.b0s_mask] for ii, tensor in enumerate(tensors): ADC = np.diag(np.linalg.multi_dot([bvecs, tensor, bvecs.T])) # Use the Stejskal-Tanner equation with the ADC as input, and S0 = 1: sig[ii] = np.exp(-bvals * ADC) return sig - np.mean(sig)
[docs] class LifeSignalMaker: """ A class for generating signals from streamlines in an efficient and speedy manner. """ @warning_for_keywords() def __init__(self, gtab, *, evals=(0.001, 0, 0), sphere=None): """ Initialize a signal maker Parameters ---------- gtab : GradientTable class instance The gradient table on which the signal is calculated. evals : array-like of 3 items, optional The eigenvalues of the canonical tensor to use in calculating the signal. sphere : `dipy.core.Sphere` class instance, optional The discrete sphere to use as an approximation for the continuous sphere on which the signal is represented. If integer - we will use an instance of one of the symmetric spheres cached in `dps.get_sphere`. If a 'dipy.core.Sphere' class instance is provided, we will use this object. Default: the :mod:`dipy.data` symmetric sphere with 724 vertices """ self.sphere = sphere or dpd.default_sphere self.gtab = gtab self.evals = evals # Initialize an empty dict to fill with signals for each of the sphere # vertices: self.signal = np.empty((self.sphere.vertices.shape[0], np.sum(~gtab.b0s_mask))) # We'll need to keep track of what we've already calculated: self._calculated = []
[docs] def calc_signal(self, xyz): idx = self.sphere.find_closest(xyz) if idx not in self._calculated: bvecs = self.gtab.bvecs[~self.gtab.b0s_mask] bvals = self.gtab.bvals[~self.gtab.b0s_mask] tensor = grad_tensor(self.sphere.vertices[idx], self.evals) ADC = np.diag(np.linalg.multi_dot([bvecs, tensor, bvecs.T])) sig = np.exp(-bvals * ADC) sig = sig - np.mean(sig) self.signal[idx] = sig self._calculated.append(idx) return self.signal[idx]
[docs] def streamline_signal(self, streamline): """ Approximate the signal for a given streamline """ grad = streamline_gradients(streamline) sig_out = np.zeros((grad.shape[0], self.signal.shape[-1])) for ii, g in enumerate(grad): sig_out[ii] = self.calc_signal(g) return sig_out
[docs] @warning_for_keywords() def voxel2streamline(streamline, affine, *, unique_idx=None): """ Maps voxels to streamlines and streamlines to voxels, for setting up the LiFE equations matrix Parameters ---------- streamline : list A collection of streamlines, each n by 3, with n being the number of nodes in the fiber. affine : array_like (4, 4) The mapping from voxel coordinates to streamline points. The voxel_to_rasmm matrix, typically from a NIFTI file. unique_idx : array, optional. The unique indices in the streamlines Returns ------- v2f, v2fn : tuple of dicts The first dict in the tuple answers the question: Given a voxel (from the unique indices in this model), which fibers pass through it? The second answers the question: Given a streamline, for each voxel that this streamline passes through, which nodes of that streamline are in that voxel? """ transformed_streamline = transform_streamlines(streamline, affine) if unique_idx is None: all_coords = np.concatenate(transformed_streamline) unique_idx = unique_rows(np.round(all_coords)) return _voxel2streamline(transformed_streamline, unique_idx.astype(np.intp))
[docs] class FiberModel(ReconstModel): """ A class for representing and solving predictive models based on tractography solutions. Notes ----- This is an implementation of the LiFE model described in :footcite:p:`Pestilli2014`. References ---------- .. footbibliography:: """ def __init__(self, gtab): """ Parameters ---------- gtab : a GradientTable class instance """ # Initialize the super-class: ReconstModel.__init__(self, gtab)
[docs] @warning_for_keywords() def setup(self, streamline, affine, *, evals=(0.001, 0, 0), sphere=None): """ Set up the necessary components for the LiFE model: the matrix of fiber-contributions to the DWI signal, and the coordinates of voxels for which the equations will be solved Parameters ---------- streamline : list Streamlines, each is an array of shape (n, 3) affine : array_like (4, 4) The mapping from voxel coordinates to streamline points. The voxel_to_rasmm matrix, typically from a NIFTI file. evals : array-like (3 items, optional) The eigenvalues of the canonical tensor used as a response function. Default:[0.001, 0, 0]. sphere: `dipy.core.Sphere` instance, optional Whether to approximate (and cache) the signal on a discrete sphere. This may confer a significant speed-up in setting up the problem, but is not as accurate. If `False`, we use the exact gradients along the streamlines to calculate the matrix, instead of an approximation. Defaults to use the 724-vertex symmetric sphere from :mod:`dipy.data` """ if sphere is not False: SignalMaker = LifeSignalMaker(self.gtab, evals=evals, sphere=sphere) streamline = transform_streamlines(streamline, affine) # Assign some local variables, for shorthand: all_coords = np.concatenate(streamline) vox_coords = unique_rows(np.round(all_coords).astype(np.intp)) del all_coords # We only consider the diffusion-weighted signals: n_bvecs = self.gtab.bvals[~self.gtab.b0s_mask].shape[0] v2f, v2fn = voxel2streamline(streamline, np.eye(4), unique_idx=vox_coords) # How many fibers in each voxel (this will determine how many # components are in the matrix): n_unique_f = len(np.hstack(list(v2f.values()))) # Preallocate these, which will be used to generate the sparse # matrix: f_matrix_sig = np.zeros(n_unique_f * n_bvecs, dtype=float) f_matrix_row = np.zeros(n_unique_f * n_bvecs, dtype=np.intp) f_matrix_col = np.zeros(n_unique_f * n_bvecs, dtype=np.intp) fiber_signal = [] for _, s in enumerate(streamline): if sphere is not False: fiber_signal.append(SignalMaker.streamline_signal(s)) else: fiber_signal.append(streamline_signal(s, self.gtab, evals=evals)) del streamline if sphere is not False: del SignalMaker keep_ct = 0 range_bvecs = np.arange(n_bvecs).astype(int) # In each voxel: for v_idx in range(vox_coords.shape[0]): mat_row_idx = (range_bvecs + v_idx * n_bvecs).astype(np.intp) # For each fiber in that voxel: for f_idx in v2f[v_idx]: # For each fiber-voxel combination, store the row/column # indices in the pre-allocated linear arrays f_matrix_row[keep_ct : keep_ct + n_bvecs] = mat_row_idx f_matrix_col[keep_ct : keep_ct + n_bvecs] = f_idx vox_fiber_sig = np.zeros(n_bvecs) for node_idx in v2fn[f_idx][v_idx]: # Sum the signal from each node of the fiber in that voxel: vox_fiber_sig += fiber_signal[f_idx][node_idx] # And add the summed thing into the corresponding rows: f_matrix_sig[keep_ct : keep_ct + n_bvecs] += vox_fiber_sig keep_ct = keep_ct + n_bvecs del v2f, v2fn # Allocate the sparse matrix, using the more memory-efficient 'csr' # format: life_matrix = sps.csr_array((f_matrix_sig, [f_matrix_row, f_matrix_col])) return life_matrix, vox_coords
def _signals(self, data, vox_coords): """ Helper function to extract and separate all the signals we need to fit and evaluate a fit of this model Parameters ---------- data : 4D array vox_coords: n by 3 array The coordinates into the data array of the fiber nodes. """ # Fitting is done on the S0-normalized-and-demeaned diffusion-weighted # signal: idx_tuple = (vox_coords[:, 0], vox_coords[:, 1], vox_coords[:, 2]) # We'll look at a 2D array, extracting the data from the voxels: vox_data = data[idx_tuple] weighted_signal = vox_data[:, ~self.gtab.b0s_mask] b0_signal = np.mean(vox_data[:, self.gtab.b0s_mask], -1) relative_signal = weighted_signal / b0_signal[:, None] # The mean of the relative signal across directions in each voxel: mean_sig = np.mean(relative_signal, -1) to_fit = (relative_signal - mean_sig[:, None]).ravel() to_fit[np.isnan(to_fit)] = 0 return (to_fit, weighted_signal, b0_signal, relative_signal, mean_sig, vox_data)
[docs] @warning_for_keywords() def fit(self, data, streamline, affine, *, evals=(0.001, 0, 0), sphere=None): """ Fit the LiFE FiberModel for data and a set of streamlines associated with this data Parameters ---------- data : 4D array Diffusion-weighted data streamline : list A bunch of streamlines affine : array_like (4, 4) The mapping from voxel coordinates to streamline points. The voxel_to_rasmm matrix, typically from a NIFTI file. evals : array-like, optional The eigenvalues of the tensor response function used in constructing the model signal. Default: [0.001, 0, 0] sphere: `dipy.core.Sphere` instance or False, optional Whether to approximate (and cache) the signal on a discrete sphere. This may confer a significant speed-up in setting up the problem, but is not as accurate. If `False`, we use the exact gradients along the streamlines to calculate the matrix, instead of an approximation. Returns ------- FiberFit class instance """ if affine is None: affine = np.eye(4) sl_len = np.array([len(s) for s in streamline]) if np.any(sl_len < 2): raise ValueError( "Input contains streamlines with only one node." " The LiFE model cannot be fit with these streamlines included." ) life_matrix, vox_coords = self.setup( streamline, affine, evals=evals, sphere=sphere ) (to_fit, weighted_signal, b0_signal, relative_signal, mean_sig, vox_data) = ( self._signals(data, vox_coords) ) beta = opt.sparse_nnls(to_fit, life_matrix) return FiberFit( self, life_matrix, vox_coords, to_fit, beta, weighted_signal, b0_signal, relative_signal, mean_sig, vox_data, streamline, affine, evals, )
[docs] class FiberFit(ReconstFit): """ A fit of the LiFE model to diffusion data """ def __init__( self, fiber_model, life_matrix, vox_coords, to_fit, beta, weighted_signal, b0_signal, relative_signal, mean_sig, vox_data, streamline, affine, evals, ): """ Parameters ---------- fiber_model : A FiberModel class instance params : the parameters derived from a fit of the model to the data. """ ReconstFit.__init__(self, fiber_model, vox_data) self.life_matrix = life_matrix self.vox_coords = vox_coords self.fit_data = to_fit self.beta = beta self.weighted_signal = weighted_signal self.b0_signal = b0_signal self.relative_signal = relative_signal self.mean_signal = mean_sig self.streamline = streamline self.affine = affine self.evals = evals
[docs] @warning_for_keywords() def predict(self, *, gtab=None, S0=None): """ Predict the signal Parameters ---------- gtab : GradientTable Default: use self.gtab S0 : float or array The non-diffusion-weighted signal in the voxels for which a prediction is made. Default: use self.b0_signal Returns ------- prediction : ndarray of shape (voxels, bvecs) An array with a prediction of the signal in each voxel/direction """ # We generate the prediction and in each voxel, we add the # offset, according to the isotropic part of the signal, which was # removed prior to fitting: if gtab is None: _matrix = self.life_matrix gtab = self.model.gtab else: _model = FiberModel(gtab) _matrix, _ = _model.setup(self.streamline, self.affine, evals=self.evals) pred_weighted = np.reshape( opt.spdot(_matrix, self.beta), (self.vox_coords.shape[0], np.sum(~gtab.b0s_mask)), ) pred = np.empty((self.vox_coords.shape[0], gtab.bvals.shape[0])) if S0 is None: S0 = self.b0_signal pred[..., gtab.b0s_mask] = S0[:, None] pred[..., ~gtab.b0s_mask] = (pred_weighted + self.mean_signal[:, None]) * S0[ :, None ] return pred