align

Bunch(**kwds)

floating

alias of numpy.float32

Module: align.imaffine

Affine image registration module consisting of the following classes:

AffineMap: encapsulates the necessary information to perform affine

transforms between two domains, defined by a static and a moving image. The domain of the transform is the set of points in the static image’s grid, and the codomain is the set of points in the moving image. When we call the transform method, AffineMap maps each point x of the domain (static grid) to the codomain (moving grid) and interpolates the moving image at that point to obtain the intensity value to be placed at x in the resulting grid. The transform_inverse method performs the opposite operation mapping points in the codomain to points in the domain.

ParzenJointHistogram: computes the marginal and joint distributions of

intensities of a pair of images, using Parzen windows [Parzen62] with a cubic spline kernel, as proposed by Mattes et al. [Mattes03]. It also computes the gradient of the joint histogram w.r.t. the parameters of a given transform.

MutualInformationMetric: computes the value and gradient of the mutual

information metric the way Optimizer needs them. That is, given a set of transform parameters, it will use ParzenJointHistogram to compute the value and gradient of the joint intensity histogram evaluated at the given parameters, and evaluate the the value and gradient of the histogram’s mutual information.

AffineRegistration: it runs the multi-resolution registration, putting

all the pieces together. It needs to create the scale space of the images and run the multi-resolution registration by using the Metric and the Optimizer at each level of the Gaussian pyramid. At each level, it will setup the metric to compute value and gradient of the metric with the input images with different levels of smoothing.

References

[Parzen62] E. Parzen. On the estimation of a probability density

function and the mode. Annals of Mathematical Statistics, 33(3), 1065-1076, 1962.

[Mattes03] Mattes, D., Haynor, D. R., Vesselle, H., Lewellen, T. K.,

& Eubank, W. PET-CT image registration in the chest using free-form deformations. IEEE Transactions on Medical Imaging, 22(1), 120-8, 2003.

AffineInvalidValuesError

AffineInversionError

AffineMap(affine[, domain_grid_shape, …])

Methods

AffineRegistration([metric, level_iters, …])

Methods

IsotropicScaleSpace(image, factors, sigmas)

Methods

MutualInformationMetric([nbins, …])

Methods

Optimizer(fun, x0[, args, method, jac, …])

Attributes

ParzenJointHistogram

Methods

ScaleSpace(image, num_levels[, …])

Methods

compute_parzen_mi()

Computes the mutual information and its gradient (if requested)

get_direction_and_spacings(affine, dim)

Extracts the rotational and spacing components from a matrix

interpolate_scalar_2d(image, locations)

Bilinear interpolation of a 2D scalar image

interpolate_scalar_3d(image, locations)

Trilinear interpolation of a 3D scalar image

sample_domain_regular()

Take floor(total_voxels/k) samples from a (2D or 3D) grid

transform_centers_of_mass(static, …)

Transformation to align the center of mass of the input images.

transform_geometric_centers(static, …)

Transformation to align the geometric center of the input images.

transform_origins(static, static_grid2world, …)

Transformation to align the origins of the input images.

warn(message[, category, stacklevel, source])

Issue a warning, or maybe ignore it or raise an exception.

Module: align.imwarp

Classes and functions for Symmetric Diffeomorphic Registration

Bunch(**kwds)

DiffeomorphicMap(dim, disp_shape[, …])

Methods

DiffeomorphicRegistration([metric])

Methods

ScaleSpace(image, num_levels[, …])

Methods

SymmetricDiffeomorphicRegistration(metric[, …])

Methods

floating

alias of numpy.float32

get_direction_and_spacings(affine, dim)

Extracts the rotational and spacing components from a matrix

mult_aff(A, B)

Returns the matrix product A.dot(B) considering None as the identity

Module: align.metrics

Metrics for Symmetric Diffeomorphic Registration

CCMetric(dim[, sigma_diff, radius])

Methods

EMMetric(dim[, smooth, inner_iter, …])

Methods

SSDMetric(dim[, smooth, inner_iter, step_type])

Methods

SimilarityMetric(dim)

Methods

floating

alias of numpy.float32

gradient(f, \*varargs, \*\*kwargs)

Return the gradient of an N-dimensional array.

v_cycle_2d(n, k, delta_field, …[, depth])

Multi-resolution Gauss-Seidel solver using V-type cycles

v_cycle_3d(n, k, delta_field, …[, depth])

Multi-resolution Gauss-Seidel solver using V-type cycles

Module: align.reslice

Pool([processes, initializer, initargs, …])

Returns a process pool object

affine_transform(input, matrix[, offset, …])

Apply an affine transformation.

cpu_count()

Returns the number of CPUs in the system

reslice(data, affine, zooms, new_zooms[, …])

Reslice data with new voxel resolution defined by new_zooms

Module: align.scalespace

IsotropicScaleSpace(image, factors, sigmas)

Methods

ScaleSpace(image, num_levels[, …])

Methods

floating

alias of numpy.float32

Module: align.streamlinear

BundleMinDistanceAsymmetricMetric([num_threads])

Asymmetric Bundle-based Minimum distance

BundleMinDistanceMatrixMetric([num_threads])

Bundle-based Minimum Distance aka BMD

BundleMinDistanceMetric([num_threads])

Bundle-based Minimum Distance aka BMD

BundleSumDistanceMatrixMetric([num_threads])

Bundle-based Sum Distance aka BMD

Optimizer(fun, x0[, args, method, jac, …])

Attributes

StreamlineDistanceMetric([num_threads])

Methods

StreamlineLinearRegistration([metric, x0, …])

Methods

StreamlineRegistrationMap(matopt, xopt, …)

Methods

Streamlines

alias of nibabel.streamlines.array_sequence.ArraySequence

bundle_min_distance(t, static, moving)

MDF-based pairwise distance optimization function (MIN)

bundle_min_distance_asymmetric_fast(t, …)

MDF-based pairwise distance optimization function (MIN)

bundle_min_distance_fast(t, static, moving, …)

MDF-based pairwise distance optimization function (MIN)

bundle_sum_distance(t, static, moving[, …])

MDF distance optimization function (SUM)

center_streamlines(streamlines)

Move streamlines to the origin

compose_matrix([scale, shear, angles, …])

Return 4x4 transformation matrix from sequence of transformations.

compose_matrix44(t[, dtype])

Compose a 4x4 transformation matrix

compose_transformations(\*mats)

Compose multiple 4x4 affine transformations in one 4x4 matrix

decompose_matrix(matrix)

Return sequence of transformations from transformation matrix.

decompose_matrix44(mat[, size])

Given a 4x4 homogeneous matrix return the parameter vector

distance_matrix_mdf()

Minimum direct flipped distance matrix between two streamline sets

length()

Euclidean length of streamlines

progressive_slr(static, moving, metric, x0, …)

Progressive SLR

qbx_and_merge(streamlines, thresholds[, …])

Run QuickBundlesX and then run again on the centroids of the last layer

remove_clusters_by_size(clusters[, min_size])

select_random_set_of_streamlines(…[, rng])

Select a random set of streamlines

set_number_of_points()

Change the number of points of streamlines

slr_with_qbx(static, moving[, x0, …])

Utility function for registering large tractograms.

time()

Return the current time in seconds since the Epoch.

transform_streamlines(streamlines, mat[, …])

Apply affine transformation to streamlines

unlist_streamlines(streamlines)

Return the streamlines not as a list but as an array and an offset

whole_brain_slr(static, moving[, x0, …])

Utility function for registering large tractograms.

Bunch

class dipy.align.Bunch(**kwds)

Bases: object

__init__(self, **kwds)

A ‘bunch’ of values (a replacement of Enum)

This is a temporary replacement of Enum, which is not available on all versions of Python 2

floating

dipy.align.floating

alias of numpy.float32

AffineInvalidValuesError

class dipy.align.imaffine.AffineInvalidValuesError

Bases: Exception

Attributes
args

Methods

with_traceback()

Exception.with_traceback(tb) – set self.__traceback__ to tb and return self.

__init__(self, /, *args, **kwargs)

Initialize self. See help(type(self)) for accurate signature.

AffineInversionError

class dipy.align.imaffine.AffineInversionError

Bases: Exception

Attributes
args

Methods

with_traceback()

Exception.with_traceback(tb) – set self.__traceback__ to tb and return self.

__init__(self, /, *args, **kwargs)

Initialize self. See help(type(self)) for accurate signature.

AffineMap

class dipy.align.imaffine.AffineMap(affine, domain_grid_shape=None, domain_grid2world=None, codomain_grid_shape=None, codomain_grid2world=None)

Bases: object

Methods

get_affine(self)

Return the value of the transformation, not a reference.

set_affine(self, affine)

Set the affine transform (operating in physical space).

transform(self, image[, interp, …])

Transform the input image from co-domain to domain space.

transform_inverse(self, image[, interp, …])

Transform the input image from domain to co-domain space.

__init__(self, affine, domain_grid_shape=None, domain_grid2world=None, codomain_grid_shape=None, codomain_grid2world=None)

AffineMap

Implements an affine transformation whose domain is given by domain_grid and domain_grid2world, and whose co-domain is given by codomain_grid and codomain_grid2world.

The actual transform is represented by the affine matrix, which operate in world coordinates. Therefore, to transform a moving image towards a static image, we first map each voxel (i,j,k) of the static image to world coordinates (x,y,z) by applying domain_grid2world. Then we apply the affine transform to (x,y,z) obtaining (x’, y’, z’) in moving image’s world coordinates. Finally, (x’, y’, z’) is mapped to voxel coordinates (i’, j’, k’) in the moving image by multiplying (x’, y’, z’) by the inverse of codomain_grid2world. The codomain_grid_shape is used analogously to transform the static image towards the moving image when calling transform_inverse.

If the domain/co-domain information is not provided (None) then the sampling information needs to be specified each time the transform or transform_inverse is called to transform images. Note that such sampling information is not necessary to transform points defined in physical space, such as stream lines.

Parameters
affinearray, shape (dim + 1, dim + 1)

the matrix defining the affine transform, where dim is the dimension of the space this map operates in (2 for 2D images, 3 for 3D images). If None, then self represents the identity transformation.

domain_grid_shapesequence, shape (dim,), optional

the shape of the default domain sampling grid. When transform is called to transform an image, the resulting image will have this shape, unless a different sampling information is provided. If None, then the sampling grid shape must be specified each time the transform method is called.

domain_grid2worldarray, shape (dim + 1, dim + 1), optional

the grid-to-world transform associated with the domain grid. If None (the default), then the grid-to-world transform is assumed to be the identity.

codomain_grid_shapesequence of integers, shape (dim,)

the shape of the default co-domain sampling grid. When transform_inverse is called to transform an image, the resulting image will have this shape, unless a different sampling information is provided. If None (the default), then the sampling grid shape must be specified each time the transform_inverse method is called.

codomain_grid2worldarray, shape (dim + 1, dim + 1)

the grid-to-world transform associated with the co-domain grid. If None (the default), then the grid-to-world transform is assumed to be the identity.

get_affine(self)

Return the value of the transformation, not a reference.

Returns
affinendarray

Copy of the transform, not a reference.

set_affine(self, affine)

Set the affine transform (operating in physical space).

Also sets self.affine_inv - the inverse of affine, or None if there is no inverse.

Parameters
affinearray, shape (dim + 1, dim + 1)

the matrix representing the affine transform operating in physical space. The domain and co-domain information remains unchanged. If None, then self represents the identity transformation.

transform(self, image, interp='linear', image_grid2world=None, sampling_grid_shape=None, sampling_grid2world=None, resample_only=False)

Transform the input image from co-domain to domain space.

By default, the transformed image is sampled at a grid defined by self.domain_shape and self.domain_grid2world. If such information was not provided then sampling_grid_shape is mandatory.

Parameters
image2D or 3D array

the image to be transformed

interpstring, either ‘linear’ or ‘nearest’

the type of interpolation to be used, either ‘linear’ (for k-linear interpolation) or ‘nearest’ for nearest neighbor

image_grid2worldarray, shape (dim + 1, dim + 1), optional

the grid-to-world transform associated with image. If None (the default), then the grid-to-world transform is assumed to be the identity.

sampling_grid_shapesequence, shape (dim,), optional

the shape of the grid where the transformed image must be sampled. If None (the default), then self.codomain_shape is used instead (which must have been set at initialization, otherwise an exception will be raised).

sampling_grid2worldarray, shape (dim + 1, dim + 1), optional

the grid-to-world transform associated with the sampling grid (specified by sampling_grid_shape, or by default self.codomain_shape). If None (the default), then the grid-to-world transform is assumed to be the identity.

resample_onlyBoolean, optional

If False (the default) the affine transform is applied normally. If True, then the affine transform is not applied, and the input image is just re-sampled on the domain grid of this transform.

Returns
transformedarray, shape sampling_grid_shape or

self.codomain_shape

the transformed image, sampled at the requested grid

transform_inverse(self, image, interp='linear', image_grid2world=None, sampling_grid_shape=None, sampling_grid2world=None, resample_only=False)

Transform the input image from domain to co-domain space.

By default, the transformed image is sampled at a grid defined by self.codomain_shape and self.codomain_grid2world. If such information was not provided then sampling_grid_shape is mandatory.

Parameters
image2D or 3D array

the image to be transformed

interpstring, either ‘linear’ or ‘nearest’

the type of interpolation to be used, either ‘linear’ (for k-linear interpolation) or ‘nearest’ for nearest neighbor

image_grid2worldarray, shape (dim + 1, dim + 1), optional

the grid-to-world transform associated with image. If None (the default), then the grid-to-world transform is assumed to be the identity.

sampling_grid_shapesequence, shape (dim,), optional

the shape of the grid where the transformed image must be sampled. If None (the default), then self.codomain_shape is used instead (which must have been set at initialization, otherwise an exception will be raised).

sampling_grid2worldarray, shape (dim + 1, dim + 1), optional

the grid-to-world transform associated with the sampling grid (specified by sampling_grid_shape, or by default self.codomain_shape). If None (the default), then the grid-to-world transform is assumed to be the identity.

resample_onlyBoolean, optional

If False (the default) the affine transform is applied normally. If True, then the affine transform is not applied, and the input image is just re-sampled on the domain grid of this transform.

Returns
transformedarray, shape sampling_grid_shape or

self.codomain_shape

the transformed image, sampled at the requested grid

AffineRegistration

class dipy.align.imaffine.AffineRegistration(metric=None, level_iters=None, sigmas=None, factors=None, method='L-BFGS-B', ss_sigma_factor=None, options=None, verbosity=1)

Bases: object

Methods

optimize(self, static, moving, transform, …)

Start the optimization process.

__init__(self, metric=None, level_iters=None, sigmas=None, factors=None, method='L-BFGS-B', ss_sigma_factor=None, options=None, verbosity=1)

Initialize an instance of the AffineRegistration class.

Parameters
metricNone or object, optional

an instance of a metric. The default is None, implying the Mutual Information metric with default settings.

level_iterssequence, optional

the number of iterations at each scale of the scale space. level_iters[0] corresponds to the coarsest scale, level_iters[-1] the finest, where n is the length of the sequence. By default, a 3-level scale space with iterations sequence equal to [10000, 1000, 100] will be used.

sigmassequence of floats, optional

custom smoothing parameter to build the scale space (one parameter for each scale). By default, the sequence of sigmas will be [3, 1, 0].

factorssequence of floats, optional

custom scale factors to build the scale space (one factor for each scale). By default, the sequence of factors will be [4, 2, 1].

methodstring, optional

optimization method to be used. If Scipy version < 0.12, then only L-BFGS-B is available. Otherwise, method can be any gradient-based method available in dipy.core.Optimize: CG, BFGS, Newton-CG, dogleg or trust-ncg. The default is ‘L-BFGS-B’.

ss_sigma_factorfloat, optional

If None, this parameter is not used and an isotropic scale space with the given factors and sigmas will be built. If not None, an anisotropic scale space will be used by automatically selecting the smoothing sigmas along each axis according to the voxel dimensions of the given image. The ss_sigma_factor is used to scale the automatically computed sigmas. For example, in the isotropic case, the sigma of the kernel will be \(factor * (2 ^ i)\) where \(i = 1, 2, ..., n_scales - 1\) is the scale (the finest resolution image \(i=0\) is never smoothed). The default is None.

optionsdict, optional

extra optimization options. The default is None, implying no extra options are passed to the optimizer.

verbosity: int (one of {0, 1, 2, 3}), optional

Set the verbosity level of the algorithm: 0 : do not print anything 1 : print information about the current status of the algorithm 2 : print high level information of the components involved in

the registration that can be used to detect a failing component.

3print as much information as possible to isolate the cause

of a bug.

Default: 1

docstring_addendum = 'verbosity: int (one of {0, 1, 2, 3}), optional\n Set the verbosity level of the algorithm:\n 0 : do not print anything\n 1 : print information about the current status of the algorithm\n 2 : print high level information of the components involved in\n the registration that can be used to detect a failing\n component.\n 3 : print as much information as possible to isolate the cause\n of a bug.\n Default: 1\n '
optimize(self, static, moving, transform, params0, static_grid2world=None, moving_grid2world=None, starting_affine=None, ret_metric=False)

Start the optimization process.

Parameters
static2D or 3D array

the image to be used as reference during optimization.

moving2D or 3D array

the image to be used as “moving” during optimization. It is necessary to pre-align the moving image to ensure its domain lies inside the domain of the deformation fields. This is assumed to be accomplished by “pre-aligning” the moving image towards the static using an affine transformation given by the ‘starting_affine’ matrix

transforminstance of Transform

the transformation with respect to whose parameters the gradient must be computed

params0array, shape (n,)

parameters from which to start the optimization. If None, the optimization will start at the identity transform. n is the number of parameters of the specified transformation.

static_grid2worldarray, shape (dim+1, dim+1), optional

the voxel-to-space transformation associated with the static image. The default is None, implying the transform is the identity.

moving_grid2worldarray, shape (dim+1, dim+1), optional

the voxel-to-space transformation associated with the moving image. The default is None, implying the transform is the identity.

starting_affinestring, or matrix, or None, optional
If string:

‘mass’: align centers of gravity ‘voxel-origin’: align physical coordinates of voxel (0,0,0) ‘centers’: align physical coordinates of central voxels

If matrix:

array, shape (dim+1, dim+1).

If None:

Start from identity.

The default is None.

ret_metricboolean, optional

if True, it returns the parameters for measuring the similarity between the images (default ‘False’). The metric containing optimal parameters and the distance between the images.

Returns
affine_mapinstance of AffineMap

the affine resulting affine transformation

xoptoptimal parameters

the optimal parameters (translation, rotation shear etc.)

foptSimilarity metric

the value of the function at the optimal parameters.

IsotropicScaleSpace

class dipy.align.imaffine.IsotropicScaleSpace(image, factors, sigmas, image_grid2world=None, input_spacing=None, mask0=False)

Bases: dipy.align.scalespace.ScaleSpace

Methods

get_affine(self, level)

Voxel-to-space transformation at a given level

get_affine_inv(self, level)

Space-to-voxel transformation at a given level

get_domain_shape(self, level)

Shape the sub-sampled image must have at a particular level

get_expand_factors(self, from_level, to_level)

Ratio of voxel size from pyramid level from_level to to_level

get_image(self, level)

Smoothed image at a given level

get_scaling(self, level)

Adjustment factor for input-spacing to reflect voxel sizes at level

get_sigmas(self, level)

Smoothing parameters used at a given level

get_spacing(self, level)

Spacings the sub-sampled image must have at a particular level

print_level(self, level)

Prints properties of a pyramid level

__init__(self, image, factors, sigmas, image_grid2world=None, input_spacing=None, mask0=False)

IsotropicScaleSpace

Computes the Scale Space representation of an image using isotropic smoothing kernels for all scales. The scale space is simply a list of images produced by smoothing the input image with a Gaussian kernel with different smoothing parameters.

This specialization of ScaleSpace allows the user to provide custom scale and smoothing factors for all scales.

Parameters
imagearray, shape (r,c) or (s, r, c) where s is the number of

slices, r is the number of rows and c is the number of columns of the input image.

factorslist of floats

custom scale factors to build the scale space (one factor for each scale).

sigmaslist of floats

custom smoothing parameter to build the scale space (one parameter for each scale).

image_grid2worldarray, shape (dim + 1, dim + 1), optional

the grid-to-space transform of the image grid. The default is the identity matrix.

input_spacingarray, shape (dim,), optional

the spacing (voxel size) between voxels in physical space. The default if 1.0 along all axes.

mask0Boolean, optional

if True, all smoothed images will be zero at all voxels that are zero in the input image. The default is False.

MutualInformationMetric

class dipy.align.imaffine.MutualInformationMetric(nbins=32, sampling_proportion=None)

Bases: object

Methods

distance(self, params)

Numeric value of the negative Mutual Information.

distance_and_gradient(self, params)

Numeric value of the metric and its gradient at given parameters.

gradient(self, params)

Numeric value of the metric’s gradient at the given parameters.

setup(self, transform, static, moving[, …])

Prepare the metric to compute intensity densities and gradients.

__init__(self, nbins=32, sampling_proportion=None)

Initialize an instance of the Mutual Information metric.

This class implements the methods required by Optimizer to drive the registration process.

Parameters
nbinsint, optional

the number of bins to be used for computing the intensity histograms. The default is 32.

sampling_proportionNone or float in interval (0, 1], optional

There are two types of sampling: dense and sparse. Dense sampling uses all voxels for estimating the (joint and marginal) intensity histograms, while sparse sampling uses a subset of them. If sampling_proportion is None, then dense sampling is used. If sampling_proportion is a floating point value in (0,1] then sparse sampling is used, where sampling_proportion specifies the proportion of voxels to be used. The default is None.

Notes

Since we use linear interpolation, images are not, in general, differentiable at exact voxel coordinates, but they are differentiable between voxel coordinates. When using sparse sampling, selected voxels are slightly moved by adding a small random displacement within one voxel to prevent sampling points from being located exactly at voxel coordinates. When using dense sampling, this random displacement is not applied.

distance(self, params)

Numeric value of the negative Mutual Information.

We need to change the sign so we can use standard minimization algorithms.

Parameters
paramsarray, shape (n,)

the parameter vector of the transform currently used by the metric (the transform name is provided when self.setup is called), n is the number of parameters of the transform

Returns
neg_mifloat

the negative mutual information of the input images after transforming the moving image by the currently set transform with params parameters

distance_and_gradient(self, params)

Numeric value of the metric and its gradient at given parameters.

Parameters
paramsarray, shape (n,)

the parameter vector of the transform currently used by the metric (the transform name is provided when self.setup is called), n is the number of parameters of the transform

Returns
neg_mifloat

the negative mutual information of the input images after transforming the moving image by the currently set transform with params parameters

neg_mi_gradarray, shape (n,)

the gradient of the negative Mutual Information

gradient(self, params)

Numeric value of the metric’s gradient at the given parameters.

Parameters
paramsarray, shape (n,)

the parameter vector of the transform currently used by the metric (the transform name is provided when self.setup is called), n is the number of parameters of the transform

Returns
gradarray, shape (n,)

the gradient of the negative Mutual Information

setup(self, transform, static, moving, static_grid2world=None, moving_grid2world=None, starting_affine=None)

Prepare the metric to compute intensity densities and gradients.

The histograms will be setup to compute probability densities of intensities within the minimum and maximum values of static and moving

Parameters
transform: instance of Transform

the transformation with respect to whose parameters the gradient must be computed

staticarray, shape (S, R, C) or (R, C)

static image

movingarray, shape (S’, R’, C’) or (R’, C’)

moving image. The dimensions of the static (S, R, C) and moving (S’, R’, C’) images do not need to be the same.

static_grid2worldarray (dim+1, dim+1), optional

the grid-to-space transform of the static image. The default is None, implying the transform is the identity.

moving_grid2worldarray (dim+1, dim+1)

the grid-to-space transform of the moving image. The default is None, implying the spacing along all axes is 1.

starting_affinearray, shape (dim+1, dim+1), optional

the pre-aligning matrix (an affine transform) that roughly aligns the moving image towards the static image. If None, no pre-alignment is performed. If a pre-alignment matrix is available, it is recommended to provide this matrix as starting_affine instead of manually transforming the moving image to reduce interpolation artifacts. The default is None, implying no pre-alignment is performed.

Optimizer

class dipy.align.imaffine.Optimizer(fun, x0, args=(), method='L-BFGS-B', jac=None, hess=None, hessp=None, bounds=None, constraints=(), tol=None, callback=None, options=None, evolution=False)

Bases: object

Attributes
evolution
fopt
message
nfev
nit
xopt

Methods

print_summary

__init__(self, fun, x0, args=(), method='L-BFGS-B', jac=None, hess=None, hessp=None, bounds=None, constraints=(), tol=None, callback=None, options=None, evolution=False)

A class for handling minimization of scalar function of one or more variables.

Parameters
funcallable

Objective function.

x0ndarray

Initial guess.

argstuple, optional

Extra arguments passed to the objective function and its derivatives (Jacobian, Hessian).

methodstr, optional

Type of solver. Should be one of

  • ‘Nelder-Mead’

  • ‘Powell’

  • ‘CG’

  • ‘BFGS’

  • ‘Newton-CG’

  • ‘Anneal’

  • ‘L-BFGS-B’

  • ‘TNC’

  • ‘COBYLA’

  • ‘SLSQP’

  • ‘dogleg’

  • ‘trust-ncg’

jacbool or callable, optional

Jacobian of objective function. Only for CG, BFGS, Newton-CG, dogleg, trust-ncg. If jac is a Boolean and is True, fun is assumed to return the value of Jacobian along with the objective function. If False, the Jacobian will be estimated numerically. jac can also be a callable returning the Jacobian of the objective. In this case, it must accept the same arguments as fun.

hess, hesspcallable, optional

Hessian of objective function or Hessian of objective function times an arbitrary vector p. Only for Newton-CG, dogleg, trust-ncg. Only one of hessp or hess needs to be given. If hess is provided, then hessp will be ignored. If neither hess nor hessp is provided, then the hessian product will be approximated using finite differences on jac. hessp must compute the Hessian times an arbitrary vector.

boundssequence, optional

Bounds for variables (only for L-BFGS-B, TNC and SLSQP). (min, max) pairs for each element in x, defining the bounds on that parameter. Use None for one of min or max when there is no bound in that direction.

constraintsdict or sequence of dict, optional

Constraints definition (only for COBYLA and SLSQP). Each constraint is defined in a dictionary with fields:

typestr

Constraint type: ‘eq’ for equality, ‘ineq’ for inequality.

funcallable

The function defining the constraint.

jaccallable, optional

The Jacobian of fun (only for SLSQP).

argssequence, optional

Extra arguments to be passed to the function and Jacobian.

Equality constraint means that the constraint function result is to be zero whereas inequality means that it is to be non-negative. Note that COBYLA only supports inequality constraints.

tolfloat, optional

Tolerance for termination. For detailed control, use solver-specific options.

callbackcallable, optional

Called after each iteration, as callback(xk), where xk is the current parameter vector. Only available using Scipy >= 0.12.

optionsdict, optional

A dictionary of solver options. All methods accept the following generic options:

maxiterint

Maximum number of iterations to perform.

dispbool

Set to True to print convergence messages.

For method-specific options, see show_options(‘minimize’, method).

evolutionbool, optional

save history of x for each iteration. Only available using Scipy >= 0.12.

See also

scipy.optimize.minimize
property evolution
property fopt
property message
property nfev
property nit
print_summary(self)
property xopt

ParzenJointHistogram

class dipy.align.imaffine.ParzenJointHistogram

Bases: object

Methods

bin_index(self, xnorm)

Bin index associated with the given normalized intensity

bin_normalize_moving(self, x)

Maps intensity x to the range covered by the moving histogram

bin_normalize_static(self, x)

Maps intensity x to the range covered by the static histogram

setup(self, static, moving[, smask, mmask])

Compute histogram settings to store the PDF of input images

update_gradient_dense(self, theta, …[, …])

Computes the Gradient of the joint PDF w.r.t.

update_gradient_sparse(self, theta, …)

Computes the Gradient of the joint PDF w.r.t.

update_pdfs_dense(self, static, moving[, …])

Computes the Probability Density Functions of two images

update_pdfs_sparse(self, sval, mval)

Computes the Probability Density Functions from a set of samples

__init__(self, nbins)

Computes joint histogram and derivatives with Parzen windows

Base class to compute joint and marginal probability density functions and their derivatives with respect to a transform’s parameters. The smooth histograms are computed by using Parzen windows [Parzen62] with a cubic spline kernel, as proposed by Mattes et al. [Mattes03]. This implementation is not tied to any optimization (registration) method, the idea is that information-theoretic matching functionals (such as Mutual Information) can inherit from this class to perform the low-level computations of the joint intensity distributions and its gradient w.r.t. the transform parameters. The derived class can then compute the similarity/dissimilarity measure and gradient, and finally communicate the results to the appropriate optimizer.

Parameters
nbinsint

the number of bins of the joint and marginal probability density functions (the actual number of bins of the joint PDF is nbins**2)

Notes

We need this class in cython to allow _joint_pdf_gradient_dense_2d and _joint_pdf_gradient_dense_3d to use a nogil Jacobian function (obtained from an instance of the Transform class), which allows us to evaluate Jacobians at all the sampling points (maybe the full grid) inside a nogil loop.

The reason we need a class is to encapsulate all the parameters related to the joint and marginal distributions.

References

[Parzen62] E. Parzen. On the estimation of a probability density

function and the mode. Annals of Mathematical Statistics, 33(3), 1065-1076, 1962.

[Mattes03] Mattes, D., Haynor, D. R., Vesselle, H., Lewellen, T. K.,

& Eubank, W. PET-CT image registration in the chest using free-form deformations. IEEE Transactions on Medical Imaging, 22(1), 120-8, 2003.

bin_index(self, xnorm)

Bin index associated with the given normalized intensity

The return value is an integer in [padding, nbins - 1 - padding]

Parameters
xnormfloat

intensity value normalized to the range covered by the histogram

Returns
binint

the bin index associated with the given normalized intensity

bin_normalize_moving(self, x)

Maps intensity x to the range covered by the moving histogram

If the input intensity is in [self.mmin, self.mmax] then the normalized intensity will be in [self.padding, self.nbins - self.padding]

Parameters
xfloat

the intensity to be normalized

Returns
xnormfloat

normalized intensity to the range covered by the moving histogram

bin_normalize_static(self, x)

Maps intensity x to the range covered by the static histogram

If the input intensity is in [self.smin, self.smax] then the normalized intensity will be in [self.padding, self.nbins - self.padding]

Parameters
xfloat

the intensity to be normalized

Returns
xnormfloat

normalized intensity to the range covered by the static histogram

setup(self, static, moving, smask=None, mmask=None)

Compute histogram settings to store the PDF of input images

Parameters
staticarray

static image

movingarray

moving image

smaskarray

mask of static object being registered (a binary array with 1’s inside the object of interest and 0’s along the background). If None, the behaviour is equivalent to smask=ones_like(static)

mmaskarray

mask of moving object being registered (a binary array with 1’s inside the object of interest and 0’s along the background). If None, the behaviour is equivalent to mmask=ones_like(static)

update_gradient_dense(self, theta, transform, static, moving, grid2world, mgradient, smask=None, mmask=None)

Computes the Gradient of the joint PDF w.r.t. transform parameters

Computes the vector of partial derivatives of the joint histogram w.r.t. each transformation parameter.

The gradient is stored in self.joint_grad.

Parameters
thetaarray, shape (n,)

parameters of the transformation to compute the gradient from

transforminstance of Transform

the transformation with respect to whose parameters the gradient must be computed

staticarray, shape (S, R, C)

static image

movingarray, shape (S, R, C)

moving image

grid2worldarray, shape (4, 4)

we assume that both images have already been sampled at a common grid. This transform must map voxel coordinates of this common grid to physical coordinates of its corresponding voxel in the moving image. For example, if the moving image was sampled on the static image’s grid (this is the typical setting) using an aligning matrix A, then

  1. grid2world = A.dot(static_affine)

where static_affine is the transformation mapping static image’s grid coordinates to physical space.

mgradientarray, shape (S, R, C, 3)

the gradient of the moving image

smaskarray, shape (S, R, C), optional

mask of static object being registered (a binary array with 1’s inside the object of interest and 0’s along the background). The default is None, indicating all voxels are considered.

mmaskarray, shape (S, R, C), optional

mask of moving object being registered (a binary array with 1’s inside the object of interest and 0’s along the background). The default is None, indicating all voxels are considered.

update_gradient_sparse(self, theta, transform, sval, mval, sample_points, mgradient)

Computes the Gradient of the joint PDF w.r.t. transform parameters

Computes the vector of partial derivatives of the joint histogram w.r.t. each transformation parameter.

The list of intensities sval and mval are assumed to be sampled from the static and moving images, respectively, at the same physical points. Of course, the images may not be perfectly aligned at the moment the sampling was performed. The resulting gradient corresponds to the paired intensities according to the alignment at the moment the images were sampled.

The gradient is stored in self.joint_grad.

Parameters
thetaarray, shape (n,)

parameters to compute the gradient at

transforminstance of Transform

the transformation with respect to whose parameters the gradient must be computed

svalarray, shape (m,)

sampled intensities from the static image at sampled_points

mvalarray, shape (m,)

sampled intensities from the moving image at sampled_points

sample_pointsarray, shape (m, 3)

coordinates (in physical space) of the points the images were sampled at

mgradientarray, shape (m, 3)

the gradient of the moving image at the sample points

update_pdfs_dense(self, static, moving, smask=None, mmask=None)

Computes the Probability Density Functions of two images

The joint PDF is stored in self.joint. The marginal distributions corresponding to the static and moving images are computed and stored in self.smarginal and self.mmarginal, respectively.

Parameters
staticarray, shape (S, R, C)

static image

movingarray, shape (S, R, C)

moving image

smaskarray, shape (S, R, C)

mask of static object being registered (a binary array with 1’s inside the object of interest and 0’s along the background). If None, ones_like(static) is used as mask.

mmaskarray, shape (S, R, C)

mask of moving object being registered (a binary array with 1’s inside the object of interest and 0’s along the background). If None, ones_like(moving) is used as mask.

update_pdfs_sparse(self, sval, mval)

Computes the Probability Density Functions from a set of samples

The list of intensities sval and mval are assumed to be sampled from the static and moving images, respectively, at the same physical points. Of course, the images may not be perfectly aligned at the moment the sampling was performed. The resulting distributions corresponds to the paired intensities according to the alignment at the moment the images were sampled.

The joint PDF is stored in self.joint. The marginal distributions corresponding to the static and moving images are computed and stored in self.smarginal and self.mmarginal, respectively.

Parameters
svalarray, shape (n,)

sampled intensities from the static image at sampled_points

mvalarray, shape (n,)

sampled intensities from the moving image at sampled_points

ScaleSpace

class dipy.align.imaffine.ScaleSpace(image, num_levels, image_grid2world=None, input_spacing=None, sigma_factor=0.2, mask0=False)

Bases: object

Methods

get_affine(self, level)

Voxel-to-space transformation at a given level

get_affine_inv(self, level)

Space-to-voxel transformation at a given level

get_domain_shape(self, level)

Shape the sub-sampled image must have at a particular level

get_expand_factors(self, from_level, to_level)

Ratio of voxel size from pyramid level from_level to to_level

get_image(self, level)

Smoothed image at a given level

get_scaling(self, level)

Adjustment factor for input-spacing to reflect voxel sizes at level

get_sigmas(self, level)

Smoothing parameters used at a given level

get_spacing(self, level)

Spacings the sub-sampled image must have at a particular level

print_level(self, level)

Prints properties of a pyramid level

__init__(self, image, num_levels, image_grid2world=None, input_spacing=None, sigma_factor=0.2, mask0=False)

ScaleSpace

Computes the Scale Space representation of an image. The scale space is simply a list of images produced by smoothing the input image with a Gaussian kernel with increasing smoothing parameter. If the image’s voxels are isotropic, the smoothing will be the same along all directions: at level L = 0, 1, …, the sigma is given by \(s * ( 2^L - 1 )\). If the voxel dimensions are not isotropic, then the smoothing is weaker along low resolution directions.

Parameters
imagearray, shape (r,c) or (s, r, c) where s is the number of

slices, r is the number of rows and c is the number of columns of the input image.

num_levelsint

the desired number of levels (resolutions) of the scale space

image_grid2worldarray, shape (dim + 1, dim + 1), optional

the grid-to-space transform of the image grid. The default is the identity matrix

input_spacingarray, shape (dim,), optional

the spacing (voxel size) between voxels in physical space. The default is 1.0 along all axes

sigma_factorfloat, optional

the smoothing factor to be used in the construction of the scale space. The default is 0.2

mask0Boolean, optional

if True, all smoothed images will be zero at all voxels that are zero in the input image. The default is False.

get_affine(self, level)

Voxel-to-space transformation at a given level

Returns the voxel-to-space transformation associated with the sub-sampled image at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get affine transform from

Returns
the affine (voxel-to-space) transform at the requested resolution

or None if an invalid level was requested

get_affine_inv(self, level)

Space-to-voxel transformation at a given level

Returns the space-to-voxel transformation associated with the sub-sampled image at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the inverse transform from

Returns
the inverse (space-to-voxel) transform at the requested resolution or
None if an invalid level was requested
get_domain_shape(self, level)

Shape the sub-sampled image must have at a particular level

Returns the shape the sub-sampled image must have at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the sub-sampled shape from

Returns
the sub-sampled shape at the requested resolution or None if an

invalid level was requested

get_expand_factors(self, from_level, to_level)

Ratio of voxel size from pyramid level from_level to to_level

Given two scale space resolutions a = from_level, b = to_level, returns the ratio of voxels size at level b to voxel size at level a (the factor that must be used to multiply voxels at level a to ‘expand’ them to level b).

Parameters
from_levelint, 0 <= from_level < L, (L = number of resolutions)

the resolution to expand voxels from

to_levelint, 0 <= to_level < from_level

the resolution to expand voxels to

Returns
factorsarray, shape (k,), k = 2, 3

the expand factors (a scalar for each voxel dimension)

get_image(self, level)

Smoothed image at a given level

Returns the smoothed image at the requested level in the Scale Space.

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the smooth image from

Returns
the smooth image at the requested resolution or None if an invalid

level was requested

get_scaling(self, level)

Adjustment factor for input-spacing to reflect voxel sizes at level

Returns the scaling factor that needs to be applied to the input spacing (the voxel sizes of the image at level 0 of the scale space) to transform them to voxel sizes at the requested level.

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the scalings from

Returns
the scaling factors from the original spacing to the spacings at the
requested level
get_sigmas(self, level)

Smoothing parameters used at a given level

Returns the smoothing parameters (a scalar for each axis) used at the requested level of the scale space

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the smoothing parameters from

Returns
the smoothing parameters at the requested level
get_spacing(self, level)

Spacings the sub-sampled image must have at a particular level

Returns the spacings (voxel sizes) the sub-sampled image must have at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the sub-sampled shape from

Returns
the spacings (voxel sizes) at the requested resolution or None if an
invalid level was requested
print_level(self, level)

Prints properties of a pyramid level

Prints the properties of a level of this scale space to standard output

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to be printed

compute_parzen_mi

dipy.align.imaffine.compute_parzen_mi()

Computes the mutual information and its gradient (if requested)

Parameters
jointarray, shape (nbins, nbins)

the joint intensity distribution

joint_gradientarray, shape (nbins, nbins, n)

the gradient of the joint distribution w.r.t. the transformation parameters

smarginalarray, shape (nbins,)

the marginal intensity distribution of the static image

mmarginalarray, shape (nbins,)

the marginal intensity distribution of the moving image

mi_gradientarray, shape (n,)

the buffer in which to write the gradient of the mutual information. If None, the gradient is not computed

get_direction_and_spacings

dipy.align.imaffine.get_direction_and_spacings(affine, dim)

Extracts the rotational and spacing components from a matrix

Extracts the rotational and spacing (voxel dimensions) components from a matrix. An image gradient represents the local variation of the image’s gray values per voxel. Since we are iterating on the physical space, we need to compute the gradients as variation per millimeter, so we need to divide each gradient’s component by the voxel size along the corresponding axis, that’s what the spacings are used for. Since the image’s gradients are oriented along the grid axes, we also need to re-orient the gradients to be given in physical space coordinates.

Parameters
affinearray, shape (k, k), k = 3, 4

the matrix transforming grid coordinates to physical space.

Returns
directionarray, shape (k-1, k-1)

the rotational component of the input matrix

spacingsarray, shape (k-1,)

the scaling component (voxel size) of the matrix

interpolate_scalar_2d

dipy.align.imaffine.interpolate_scalar_2d(image, locations)

Bilinear interpolation of a 2D scalar image

Interpolates the 2D image at the given locations. This function is a wrapper for _interpolate_scalar_2d for testing purposes, it is equivalent to scipy.ndimage.interpolation.map_coordinates with bilinear interpolation

Parameters
fieldarray, shape (S, R)

the 2D image to be interpolated

locationsarray, shape (n, 2)

(locations[i,0], locations[i,1]), 0<=i<n must contain the row and column coordinates to interpolate the image at

Returns
outarray, shape (n,)

out[i], 0<=i<n will be the interpolated scalar at coordinates locations[i,:], or 0 if locations[i,:] is outside the image

insidearray, (n,)

if locations[i:] is inside the image then inside[i]=1, else inside[i]=0

interpolate_scalar_3d

dipy.align.imaffine.interpolate_scalar_3d(image, locations)

Trilinear interpolation of a 3D scalar image

Interpolates the 3D image at the given locations. This function is a wrapper for _interpolate_scalar_3d for testing purposes, it is equivalent to scipy.ndimage.interpolation.map_coordinates with trilinear interpolation

Parameters
fieldarray, shape (S, R, C)

the 3D image to be interpolated

locationsarray, shape (n, 3)

(locations[i,0], locations[i,1], locations[i,2), 0<=i<n must contain the coordinates to interpolate the image at

Returns
outarray, shape (n,)

out[i], 0<=i<n will be the interpolated scalar at coordinates locations[i,:], or 0 if locations[i,:] is outside the image

insidearray, (n,)

if locations[i,:] is inside the image then inside[i]=1, else inside[i]=0

sample_domain_regular

dipy.align.imaffine.sample_domain_regular()

Take floor(total_voxels/k) samples from a (2D or 3D) grid

The sampling is made by taking all pixels whose index (in lexicographical order) is a multiple of k. Each selected point is slightly perturbed by adding a realization of a normally distributed random variable and then mapped to physical space by the given grid-to-space transform.

The lexicographical order of a pixels in a grid of shape (a, b, c) is defined by assigning to each voxel position (i, j, k) the integer index

F((i, j, k)) = i * (b * c) + j * (c) + k

and sorting increasingly by this index.

Parameters
kint

the sampling rate, as described before

shapearray, shape (dim,)

the shape of the grid to be sampled

grid2worldarray, shape (dim+1, dim+1)

the grid-to-space transform

sigmafloat

the standard deviation of the Normal random distortion to be applied to the sampled points

Returns
samplesarray, shape (total_pixels//k, dim)

the matrix whose rows are the sampled points

Examples

>>> from dipy.align.parzenhist import sample_domain_regular
>>> import dipy.align.vector_fields as vf
>>> shape = np.array((10, 10), dtype=np.int32)
>>> sigma = 0
>>> dim = len(shape)
>>> grid2world = np.eye(dim+1)
>>> n = shape[0]*shape[1]
>>> k = 2
>>> samples = sample_domain_regular(k, shape, grid2world, sigma)
>>> (samples.shape[0], samples.shape[1]) == (n//k, dim)
True
>>> isamples = np.array(samples, dtype=np.int32)
>>> indices = (isamples[:, 0] * shape[1] + isamples[:, 1])
>>> len(set(indices)) == len(indices)
True
>>> (indices%k).sum()
0

transform_centers_of_mass

dipy.align.imaffine.transform_centers_of_mass(static, static_grid2world, moving, moving_grid2world)

Transformation to align the center of mass of the input images.

Parameters
staticarray, shape (S, R, C)

static image

static_grid2worldarray, shape (dim+1, dim+1)

the voxel-to-space transformation of the static image

movingarray, shape (S, R, C)

moving image

moving_grid2worldarray, shape (dim+1, dim+1)

the voxel-to-space transformation of the moving image

Returns
affine_mapinstance of AffineMap

the affine transformation (translation only, in this case) aligning the center of mass of the moving image towards the one of the static image

transform_geometric_centers

dipy.align.imaffine.transform_geometric_centers(static, static_grid2world, moving, moving_grid2world)

Transformation to align the geometric center of the input images.

With “geometric center” of a volume we mean the physical coordinates of its central voxel

Parameters
staticarray, shape (S, R, C)

static image

static_grid2worldarray, shape (dim+1, dim+1)

the voxel-to-space transformation of the static image

movingarray, shape (S, R, C)

moving image

moving_grid2worldarray, shape (dim+1, dim+1)

the voxel-to-space transformation of the moving image

Returns
affine_mapinstance of AffineMap

the affine transformation (translation only, in this case) aligning the geometric center of the moving image towards the one of the static image

transform_origins

dipy.align.imaffine.transform_origins(static, static_grid2world, moving, moving_grid2world)

Transformation to align the origins of the input images.

With “origin” of a volume we mean the physical coordinates of voxel (0,0,0)

Parameters
staticarray, shape (S, R, C)

static image

static_grid2worldarray, shape (dim+1, dim+1)

the voxel-to-space transformation of the static image

movingarray, shape (S, R, C)

moving image

moving_grid2worldarray, shape (dim+1, dim+1)

the voxel-to-space transformation of the moving image

Returns
affine_mapinstance of AffineMap

the affine transformation (translation only, in this case) aligning the origin of the moving image towards the one of the static image

warn

dipy.align.imaffine.warn(message, category=None, stacklevel=1, source=None)

Issue a warning, or maybe ignore it or raise an exception.

Bunch

class dipy.align.imwarp.Bunch(**kwds)

Bases: object

__init__(self, **kwds)

A ‘bunch’ of values (a replacement of Enum)

This is a temporary replacement of Enum, which is not available on all versions of Python 2

DiffeomorphicMap

class dipy.align.imwarp.DiffeomorphicMap(dim, disp_shape, disp_grid2world=None, domain_shape=None, domain_grid2world=None, codomain_shape=None, codomain_grid2world=None, prealign=None)

Bases: object

Methods

allocate(self)

Creates a zero displacement field

compute_inversion_error(self)

Inversion error of the displacement fields

expand_fields(self, expand_factors, new_shape)

Expands the displacement fields from current shape to new_shape

get_backward_field(self)

Deformation field to transform an image in the backward direction

get_forward_field(self)

Deformation field to transform an image in the forward direction

get_simplified_transform(self)

Constructs a simplified version of this Diffeomorhic Map

interpret_matrix(self, obj)

Try to interpret obj as a matrix

inverse(self)

Inverse of this DiffeomorphicMap instance

shallow_copy(self)

Shallow copy of this DiffeomorphicMap instance

transform(self, image[, interpolation, …])

Warps an image in the forward direction

transform_inverse(self, image[, …])

Warps an image in the backward direction

warp_endomorphism(self, phi)

Composition of this DiffeomorphicMap with a given endomorphism

__init__(self, dim, disp_shape, disp_grid2world=None, domain_shape=None, domain_grid2world=None, codomain_shape=None, codomain_grid2world=None, prealign=None)

DiffeomorphicMap

Implements a diffeomorphic transformation on the physical space. The deformation fields encoding the direct and inverse transformations share the same domain discretization (both the discretization grid shape and voxel-to-space matrix). The input coordinates (physical coordinates) are first aligned using prealign, and then displaced using the corresponding vector field interpolated at the aligned coordinates.

Parameters
dimint, 2 or 3

the transformation’s dimension

disp_shapearray, shape (dim,)

the number of slices (if 3D), rows and columns of the deformation field’s discretization

disp_grid2worldthe voxel-to-space transform between the def. fields

grid and space

domain_shapearray, shape (dim,)

the number of slices (if 3D), rows and columns of the default discretizatio of this map’s domain

domain_grid2worldarray, shape (dim+1, dim+1)

the default voxel-to-space transformation between this map’s discretization and physical space

codomain_shapearray, shape (dim,)

the number of slices (if 3D), rows and columns of the images that are ‘normally’ warped using this transformation in the forward direction (this will provide default transformation parameters to warp images under this transformation). By default, we assume that the inverse transformation is ‘normally’ used to warp images with the same discretization and voxel-to-space transformation as the deformation field grid.

codomain_grid2worldarray, shape (dim+1, dim+1)

the voxel-to-space transformation of images that are ‘normally’ warped using this transformation (in the forward direction).

prealignarray, shape (dim+1, dim+1)

the linear transformation to be applied to align input images to the reference space before warping under the deformation field.

allocate(self)

Creates a zero displacement field

Creates a zero displacement field (the identity transformation).

compute_inversion_error(self)

Inversion error of the displacement fields

Estimates the inversion error of the displacement fields by computing statistics of the residual vectors obtained after composing the forward and backward displacement fields.

Returns
residualarray, shape (R, C) or (S, R, C)

the displacement field resulting from composing the forward and backward displacement fields of this transformation (the residual should be zero for a perfect diffeomorphism)

statsarray, shape (3,)

statistics from the norms of the vectors of the residual displacement field: maximum, mean and standard deviation

Notes

Since the forward and backward displacement fields have the same discretization, the final composition is given by

comp[i] = forward[ i + Dinv * backward[i]]

where Dinv is the space-to-grid transformation of the displacement fields

expand_fields(self, expand_factors, new_shape)

Expands the displacement fields from current shape to new_shape

Up-samples the discretization of the displacement fields to be of new_shape shape.

Parameters
expand_factorsarray, shape (dim,)

the factors scaling current spacings (voxel sizes) to spacings in the expanded discretization.

new_shapearray, shape (dim,)

the shape of the arrays holding the up-sampled discretization

get_backward_field(self)

Deformation field to transform an image in the backward direction

Returns the deformation field that must be used to warp an image under this transformation in the backward direction (note the ‘is_inverse’ flag).

get_forward_field(self)

Deformation field to transform an image in the forward direction

Returns the deformation field that must be used to warp an image under this transformation in the forward direction (note the ‘is_inverse’ flag).

get_simplified_transform(self)

Constructs a simplified version of this Diffeomorhic Map

The simplified version incorporates the pre-align transform, as well as the domain and codomain affine transforms into the displacement field. The resulting transformation may be regarded as operating on the image spaces given by the domain and codomain discretization. As a result, self.prealign, self.disp_grid2world, self.domain_grid2world and self.codomain affine will be None (denoting Identity) in the resulting diffeomorphic map.

interpret_matrix(self, obj)

Try to interpret obj as a matrix

Some operations are performed faster if we know in advance if a matrix is the identity (so we can skip the actual matrix-vector multiplication). This function returns None if the given object is None or the ‘identity’ string. It returns the same object if it is a numpy array. It raises an exception otherwise.

Parameters
objobject

any object

Returns
objobject

the same object given as argument if obj is None or a numpy array. None if obj is the ‘identity’ string.

inverse(self)

Inverse of this DiffeomorphicMap instance

Returns a diffeomorphic map object representing the inverse of this transformation. The internal arrays are not copied but just referenced.

Returns
invDiffeomorphicMap object

the inverse of this diffeomorphic map.

shallow_copy(self)

Shallow copy of this DiffeomorphicMap instance

Creates a shallow copy of this diffeomorphic map (the arrays are not copied but just referenced)

Returns
new_mapDiffeomorphicMap object

the shallow copy of this diffeomorphic map

transform(self, image, interpolation='linear', image_world2grid=None, out_shape=None, out_grid2world=None)

Warps an image in the forward direction

Transforms the input image under this transformation in the forward direction. It uses the “is_inverse” flag to switch between “forward” and “backward” (if is_inverse is False, then transform(…) warps the image forwards, else it warps the image backwards).

Parameters
imagearray, shape (s, r, c) if dim = 3 or (r, c) if dim = 2

the image to be warped under this transformation in the forward direction

interpolationstring, either ‘linear’ or ‘nearest’

the type of interpolation to be used for warping, either ‘linear’ (for k-linear interpolation) or ‘nearest’ for nearest neighbor

image_world2gridarray, shape (dim+1, dim+1)

the transformation bringing world (space) coordinates to voxel coordinates of the image given as input

out_shapearray, shape (dim,)

the number of slices, rows and columns of the desired warped image

out_grid2worldthe transformation bringing voxel coordinates of the

warped image to physical space

Returns
warpedarray, shape = out_shape or self.codomain_shape if None

the warped image under this transformation in the forward direction

Notes

See _warp_forward and _warp_backward documentation for further information.

transform_inverse(self, image, interpolation='linear', image_world2grid=None, out_shape=None, out_grid2world=None)

Warps an image in the backward direction

Transforms the input image under this transformation in the backward direction. It uses the “is_inverse” flag to switch between “forward” and “backward” (if is_inverse is False, then transform_inverse(…) warps the image backwards, else it warps the image forwards)

Parameters
imagearray, shape (s, r, c) if dim = 3 or (r, c) if dim = 2

the image to be warped under this transformation in the forward direction

interpolationstring, either ‘linear’ or ‘nearest’

the type of interpolation to be used for warping, either ‘linear’ (for k-linear interpolation) or ‘nearest’ for nearest neighbor

image_world2gridarray, shape (dim+1, dim+1)

the transformation bringing world (space) coordinates to voxel coordinates of the image given as input

out_shapearray, shape (dim,)

the number of slices, rows and columns of the desired warped image

out_grid2worldthe transformation bringing voxel coordinates of the

warped image to physical space

Returns
warpedarray, shape = out_shape or self.codomain_shape if None

warped image under this transformation in the backward direction

Notes

See _warp_forward and _warp_backward documentation for further information.

warp_endomorphism(self, phi)

Composition of this DiffeomorphicMap with a given endomorphism

Creates a new DiffeomorphicMap C with the same properties as self and composes its displacement fields with phi’s corresponding fields. The resulting diffeomorphism is of the form C(x) = phi(self(x)) with inverse C^{-1}(y) = self^{-1}(phi^{-1}(y)). We assume that phi is an endomorphism with the same discretization and domain affine as self to ensure that the composition inherits self’s properties (we also assume that the pre-aligning matrix of phi is None or identity).

Parameters
phiDiffeomorphicMap object

the endomorphism to be warped by this diffeomorphic map

Returns
compositionthe composition of this diffeomorphic map with the

endomorphism given as input

Notes

The problem with our current representation of a DiffeomorphicMap is that the set of Diffeomorphism that can be represented this way (a pre-aligning matrix followed by a non-linear endomorphism given as a displacement field) is not closed under the composition operation.

Supporting a general DiffeomorphicMap class, closed under composition, may be extremely costly computationally, and the kind of transformations we actually need for Avants’ mid-point algorithm (SyN) are much simpler.

DiffeomorphicRegistration

class dipy.align.imwarp.DiffeomorphicRegistration(metric=None)

Bases: object

Methods

get_map(self)

Returns the resulting diffeomorphic map after optimization

optimize(self)

Starts the metric optimization

set_level_iters(self, level_iters)

Sets the number of iterations at each pyramid level

__init__(self, metric=None)

Diffeomorphic Registration

This abstract class defines the interface to be implemented by any optimization algorithm for diffeomorphic registration.

Parameters
metricSimilarityMetric object

the object measuring the similarity of the two images. The registration algorithm will minimize (or maximize) the provided similarity.

abstract get_map(self)

Returns the resulting diffeomorphic map after optimization

abstract optimize(self)

Starts the metric optimization

This is the main function each specialized class derived from this must implement. Upon completion, the deformation field must be available from the forward transformation model.

set_level_iters(self, level_iters)

Sets the number of iterations at each pyramid level

Establishes the maximum number of iterations to be performed at each level of the Gaussian pyramid, similar to ANTS.

Parameters
level_iterslist

the number of iterations at each level of the Gaussian pyramid. level_iters[0] corresponds to the finest level, level_iters[n-1] the coarsest, where n is the length of the list

ScaleSpace

class dipy.align.imwarp.ScaleSpace(image, num_levels, image_grid2world=None, input_spacing=None, sigma_factor=0.2, mask0=False)

Bases: object

Methods

get_affine(self, level)

Voxel-to-space transformation at a given level

get_affine_inv(self, level)

Space-to-voxel transformation at a given level

get_domain_shape(self, level)

Shape the sub-sampled image must have at a particular level

get_expand_factors(self, from_level, to_level)

Ratio of voxel size from pyramid level from_level to to_level

get_image(self, level)

Smoothed image at a given level

get_scaling(self, level)

Adjustment factor for input-spacing to reflect voxel sizes at level

get_sigmas(self, level)

Smoothing parameters used at a given level

get_spacing(self, level)

Spacings the sub-sampled image must have at a particular level

print_level(self, level)

Prints properties of a pyramid level

__init__(self, image, num_levels, image_grid2world=None, input_spacing=None, sigma_factor=0.2, mask0=False)

ScaleSpace

Computes the Scale Space representation of an image. The scale space is simply a list of images produced by smoothing the input image with a Gaussian kernel with increasing smoothing parameter. If the image’s voxels are isotropic, the smoothing will be the same along all directions: at level L = 0, 1, …, the sigma is given by \(s * ( 2^L - 1 )\). If the voxel dimensions are not isotropic, then the smoothing is weaker along low resolution directions.

Parameters
imagearray, shape (r,c) or (s, r, c) where s is the number of

slices, r is the number of rows and c is the number of columns of the input image.

num_levelsint

the desired number of levels (resolutions) of the scale space

image_grid2worldarray, shape (dim + 1, dim + 1), optional

the grid-to-space transform of the image grid. The default is the identity matrix

input_spacingarray, shape (dim,), optional

the spacing (voxel size) between voxels in physical space. The default is 1.0 along all axes

sigma_factorfloat, optional

the smoothing factor to be used in the construction of the scale space. The default is 0.2

mask0Boolean, optional

if True, all smoothed images will be zero at all voxels that are zero in the input image. The default is False.

get_affine(self, level)

Voxel-to-space transformation at a given level

Returns the voxel-to-space transformation associated with the sub-sampled image at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get affine transform from

Returns
the affine (voxel-to-space) transform at the requested resolution

or None if an invalid level was requested

get_affine_inv(self, level)

Space-to-voxel transformation at a given level

Returns the space-to-voxel transformation associated with the sub-sampled image at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the inverse transform from

Returns
the inverse (space-to-voxel) transform at the requested resolution or
None if an invalid level was requested
get_domain_shape(self, level)

Shape the sub-sampled image must have at a particular level

Returns the shape the sub-sampled image must have at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the sub-sampled shape from

Returns
the sub-sampled shape at the requested resolution or None if an

invalid level was requested

get_expand_factors(self, from_level, to_level)

Ratio of voxel size from pyramid level from_level to to_level

Given two scale space resolutions a = from_level, b = to_level, returns the ratio of voxels size at level b to voxel size at level a (the factor that must be used to multiply voxels at level a to ‘expand’ them to level b).

Parameters
from_levelint, 0 <= from_level < L, (L = number of resolutions)

the resolution to expand voxels from

to_levelint, 0 <= to_level < from_level

the resolution to expand voxels to

Returns
factorsarray, shape (k,), k = 2, 3

the expand factors (a scalar for each voxel dimension)

get_image(self, level)

Smoothed image at a given level

Returns the smoothed image at the requested level in the Scale Space.

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the smooth image from

Returns
the smooth image at the requested resolution or None if an invalid

level was requested

get_scaling(self, level)

Adjustment factor for input-spacing to reflect voxel sizes at level

Returns the scaling factor that needs to be applied to the input spacing (the voxel sizes of the image at level 0 of the scale space) to transform them to voxel sizes at the requested level.

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the scalings from

Returns
the scaling factors from the original spacing to the spacings at the
requested level
get_sigmas(self, level)

Smoothing parameters used at a given level

Returns the smoothing parameters (a scalar for each axis) used at the requested level of the scale space

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the smoothing parameters from

Returns
the smoothing parameters at the requested level
get_spacing(self, level)

Spacings the sub-sampled image must have at a particular level

Returns the spacings (voxel sizes) the sub-sampled image must have at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the sub-sampled shape from

Returns
the spacings (voxel sizes) at the requested resolution or None if an
invalid level was requested
print_level(self, level)

Prints properties of a pyramid level

Prints the properties of a level of this scale space to standard output

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to be printed

SymmetricDiffeomorphicRegistration

class dipy.align.imwarp.SymmetricDiffeomorphicRegistration(metric, level_iters=None, step_length=0.25, ss_sigma_factor=0.2, opt_tol=1e-05, inv_iter=20, inv_tol=0.001, callback=None)

Bases: dipy.align.imwarp.DiffeomorphicRegistration

Methods

get_map(self)

Returns the resulting diffeomorphic map Returns the DiffeomorphicMap registering the moving image towards the static image.

optimize(self, static, moving[, …])

Starts the optimization

set_level_iters(self, level_iters)

Sets the number of iterations at each pyramid level

update(self, current_displacement, …)

Composition of the current displacement field with the given field

__init__(self, metric, level_iters=None, step_length=0.25, ss_sigma_factor=0.2, opt_tol=1e-05, inv_iter=20, inv_tol=0.001, callback=None)

Symmetric Diffeomorphic Registration (SyN) Algorithm

Performs the multi-resolution optimization algorithm for non-linear registration using a given similarity metric.

Parameters
metricSimilarityMetric object

the metric to be optimized

level_iterslist of int

the number of iterations at each level of the Gaussian Pyramid (the length of the list defines the number of pyramid levels to be used)

opt_tolfloat

the optimization will stop when the estimated derivative of the energy profile w.r.t. time falls below this threshold

inv_iterint

the number of iterations to be performed by the displacement field inversion algorithm

step_lengthfloat

the length of the maximum displacement vector of the update displacement field at each iteration

ss_sigma_factorfloat

parameter of the scale-space smoothing kernel. For example, the std. dev. of the kernel will be factor*(2^i) in the isotropic case where i = 0, 1, …, n_scales is the scale

inv_tolfloat

the displacement field inversion algorithm will stop iterating when the inversion error falls below this threshold

callbackfunction(SymmetricDiffeomorphicRegistration)

a function receiving a SymmetricDiffeomorphicRegistration object to be called after each iteration (this optimizer will call this function passing self as parameter)

get_map(self)

Returns the resulting diffeomorphic map Returns the DiffeomorphicMap registering the moving image towards the static image.

optimize(self, static, moving, static_grid2world=None, moving_grid2world=None, prealign=None)

Starts the optimization

Parameters
staticarray, shape (S, R, C) or (R, C)

the image to be used as reference during optimization. The displacement fields will have the same discretization as the static image.

movingarray, shape (S, R, C) or (R, C)

the image to be used as “moving” during optimization. Since the deformation fields’ discretization is the same as the static image, it is necessary to pre-align the moving image to ensure its domain lies inside the domain of the deformation fields. This is assumed to be accomplished by “pre-aligning” the moving image towards the static using an affine transformation given by the ‘prealign’ matrix

static_grid2worldarray, shape (dim+1, dim+1)

the voxel-to-space transformation associated to the static image

moving_grid2worldarray, shape (dim+1, dim+1)

the voxel-to-space transformation associated to the moving image

prealignarray, shape (dim+1, dim+1)

the affine transformation (operating on the physical space) pre-aligning the moving image towards the static

Returns
static_to_refDiffeomorphicMap object

the diffeomorphic map that brings the moving image towards the static one in the forward direction (i.e. by calling static_to_ref.transform) and the static image towards the moving one in the backward direction (i.e. by calling static_to_ref.transform_inverse).

update(self, current_displacement, new_displacement, disp_world2grid, time_scaling)

Composition of the current displacement field with the given field

Interpolates new displacement at the locations defined by current_displacement. Equivalently, computes the composition C of the given displacement fields as C(x) = B(A(x)), where A is current_displacement and B is new_displacement. This function is intended to be used with deformation fields of the same sampling (e.g. to be called by a registration algorithm).

Parameters
current_displacementarray, shape (R’, C’, 2) or (S’, R’, C’, 3)

the displacement field defining where to interpolate new_displacement

new_displacementarray, shape (R, C, 2) or (S, R, C, 3)

the displacement field to be warped by current_displacement

disp_world2gridarray, shape (dim+1, dim+1)

the space-to-grid transform associated with the displacements’ grid (we assume that both displacements are discretized over the same grid)

time_scalingfloat

scaling factor applied to d2. The effect may be interpreted as moving d1 displacements along a factor (time_scaling) of d2.

Returns
updatedarray, shape (the same as new_displacement)

the warped displacement field

mean_normthe mean norm of all vectors in current_displacement

floating

dipy.align.imwarp.floating

alias of numpy.float32

get_direction_and_spacings

dipy.align.imwarp.get_direction_and_spacings(affine, dim)

Extracts the rotational and spacing components from a matrix

Extracts the rotational and spacing (voxel dimensions) components from a matrix. An image gradient represents the local variation of the image’s gray values per voxel. Since we are iterating on the physical space, we need to compute the gradients as variation per millimeter, so we need to divide each gradient’s component by the voxel size along the corresponding axis, that’s what the spacings are used for. Since the image’s gradients are oriented along the grid axes, we also need to re-orient the gradients to be given in physical space coordinates.

Parameters
affinearray, shape (k, k), k = 3, 4

the matrix transforming grid coordinates to physical space.

Returns
directionarray, shape (k-1, k-1)

the rotational component of the input matrix

spacingsarray, shape (k-1,)

the scaling component (voxel size) of the matrix

mult_aff

dipy.align.imwarp.mult_aff(A, B)

Returns the matrix product A.dot(B) considering None as the identity

Parameters
Aarray, shape (n,k)
Barray, shape (k,m)
Returns
The matrix product A.dot(B). If any of the input matrices is None, it is
treated as the identity matrix. If both matrices are None, None is returned

CCMetric

class dipy.align.metrics.CCMetric(dim, sigma_diff=2.0, radius=4)

Bases: dipy.align.metrics.SimilarityMetric

Methods

compute_backward(self)

Computes one step bringing the static image towards the moving.

compute_forward(self)

Computes one step bringing the moving image towards the static.

free_iteration(self)

Frees the resources allocated during initialization

get_energy(self)

Numerical value assigned by this metric to the current image pair

initialize_iteration(self)

Prepares the metric to compute one displacement field iteration.

set_levels_above(self, levels)

Informs the metric how many pyramid levels are above the current one

set_levels_below(self, levels)

Informs the metric how many pyramid levels are below the current one

set_moving_image(self, moving_image, …)

Sets the moving image being compared against the static one.

set_static_image(self, static_image, …)

Sets the static image being compared against the moving one.

use_moving_image_dynamics(self, …)

This is called by the optimizer just after setting the moving image

use_static_image_dynamics(self, …)

This is called by the optimizer just after setting the static image.

__init__(self, dim, sigma_diff=2.0, radius=4)

Normalized Cross-Correlation Similarity metric.

Parameters
dimint (either 2 or 3)

the dimension of the image domain

sigma_diffthe standard deviation of the Gaussian smoothing kernel to

be applied to the update field at each iteration

radiusint

the radius of the squared (cubic) neighborhood at each voxel to be considered to compute the cross correlation

compute_backward(self)

Computes one step bringing the static image towards the moving.

Computes the update displacement field to be used for registration of the static image towards the moving image

compute_forward(self)

Computes one step bringing the moving image towards the static.

Computes the update displacement field to be used for registration of the moving image towards the static image

free_iteration(self)

Frees the resources allocated during initialization

get_energy(self)

Numerical value assigned by this metric to the current image pair

Returns the Cross Correlation (data term) energy computed at the largest iteration

initialize_iteration(self)

Prepares the metric to compute one displacement field iteration.

Pre-computes the cross-correlation factors for efficient computation of the gradient of the Cross Correlation w.r.t. the displacement field. It also pre-computes the image gradients in the physical space by re-orienting the gradients in the voxel space using the corresponding affine transformations.

EMMetric

class dipy.align.metrics.EMMetric(dim, smooth=1.0, inner_iter=5, q_levels=256, double_gradient=True, step_type='gauss_newton')

Bases: dipy.align.metrics.SimilarityMetric

Methods

compute_backward(self)

Computes one step bringing the static image towards the moving.

compute_demons_step(self[, forward_step])

Demons step for EM metric

compute_forward(self)

Computes one step bringing the reference image towards the static.

compute_gauss_newton_step(self[, forward_step])

Computes the Gauss-Newton energy minimization step

free_iteration(self)

Frees the resources allocated during initialization

get_energy(self)

The numerical value assigned by this metric to the current image pair

initialize_iteration(self)

Prepares the metric to compute one displacement field iteration.

set_levels_above(self, levels)

Informs the metric how many pyramid levels are above the current one

set_levels_below(self, levels)

Informs the metric how many pyramid levels are below the current one

set_moving_image(self, moving_image, …)

Sets the moving image being compared against the static one.

set_static_image(self, static_image, …)

Sets the static image being compared against the moving one.

use_moving_image_dynamics(self, …)

This is called by the optimizer just after setting the moving image.

use_static_image_dynamics(self, …)

This is called by the optimizer just after setting the static image.

__init__(self, dim, smooth=1.0, inner_iter=5, q_levels=256, double_gradient=True, step_type='gauss_newton')

Expectation-Maximization Metric

Similarity metric based on the Expectation-Maximization algorithm to handle multi-modal images. The transfer function is modeled as a set of hidden random variables that are estimated at each iteration of the algorithm.

Parameters
dimint (either 2 or 3)

the dimension of the image domain

smoothfloat

smoothness parameter, the larger the value the smoother the deformation field

inner_iterint

number of iterations to be performed at each level of the multi- resolution Gauss-Seidel optimization algorithm (this is not the number of steps per Gaussian Pyramid level, that parameter must be set for the optimizer, not the metric)

q_levelsnumber of quantization levels (equal to the number of hidden

variables in the EM algorithm)

double_gradientboolean

if True, the gradient of the expected static image under the moving modality will be added to the gradient of the moving image, similarly, the gradient of the expected moving image under the static modality will be added to the gradient of the static image.

step_typestring (‘gauss_newton’, ‘demons’)

the optimization schedule to be used in the multi-resolution Gauss-Seidel optimization algorithm (not used if Demons Step is selected)

compute_backward(self)

Computes one step bringing the static image towards the moving.

Computes the update displacement field to be used for registration of the static image towards the moving image

compute_demons_step(self, forward_step=True)

Demons step for EM metric

Parameters
forward_stepboolean

if True, computes the Demons step in the forward direction (warping the moving towards the static image). If False, computes the backward step (warping the static image to the moving image)

Returns
displacementarray, shape (R, C, 2) or (S, R, C, 3)

the Demons step

compute_forward(self)

Computes one step bringing the reference image towards the static.

Computes the forward update field to register the moving image towards the static image in a gradient-based optimization algorithm

compute_gauss_newton_step(self, forward_step=True)

Computes the Gauss-Newton energy minimization step

Computes the Newton step to minimize this energy, i.e., minimizes the linearized energy function with respect to the regularized displacement field (this step does not require post-smoothing, as opposed to the demons step, which does not include regularization). To accelerate convergence we use the multi-grid Gauss-Seidel algorithm proposed by Bruhn and Weickert et al [Bruhn05]

Parameters
forward_stepboolean

if True, computes the Newton step in the forward direction (warping the moving towards the static image). If False, computes the backward step (warping the static image to the moving image)

Returns
displacementarray, shape (R, C, 2) or (S, R, C, 3)

the Newton step

References

[Bruhn05] Andres Bruhn and Joachim Weickert, “Towards ultimate motion

estimation: combining highest accuracy with real-time performance”, 10th IEEE International Conference on Computer Vision, 2005. ICCV 2005.

free_iteration(self)

Frees the resources allocated during initialization

get_energy(self)

The numerical value assigned by this metric to the current image pair

Returns the EM (data term) energy computed at the largest iteration

initialize_iteration(self)

Prepares the metric to compute one displacement field iteration.

Pre-computes the transfer functions (hidden random variables) and variances of the estimators. Also pre-computes the gradient of both input images. Note that once the images are transformed to the opposite modality, the gradient of the transformed images can be used with the gradient of the corresponding modality in the same fashion as diff-demons does for mono-modality images. If the flag self.use_double_gradient is True these gradients are averaged.

use_moving_image_dynamics(self, original_moving_image, transformation)

This is called by the optimizer just after setting the moving image.

EMMetric takes advantage of the image dynamics by computing the current moving image mask from the original_moving_image mask (warped by nearest neighbor interpolation)

Parameters
original_moving_imagearray, shape (R, C) or (S, R, C)

the original moving image from which the current moving image was generated, the current moving image is the one that was provided via ‘set_moving_image(…)’, which may not be the same as the original moving image but a warped version of it.

transformationDiffeomorphicMap object

the transformation that was applied to the original_moving_image to generate the current moving image

use_static_image_dynamics(self, original_static_image, transformation)

This is called by the optimizer just after setting the static image.

EMMetric takes advantage of the image dynamics by computing the current static image mask from the originalstaticImage mask (warped by nearest neighbor interpolation)

Parameters
original_static_imagearray, shape (R, C) or (S, R, C)

the original static image from which the current static image was generated, the current static image is the one that was provided via ‘set_static_image(…)’, which may not be the same as the original static image but a warped version of it (even the static image changes during Symmetric Normalization, not only the moving one).

transformationDiffeomorphicMap object

the transformation that was applied to the original_static_image to generate the current static image

SSDMetric

class dipy.align.metrics.SSDMetric(dim, smooth=4, inner_iter=10, step_type='demons')

Bases: dipy.align.metrics.SimilarityMetric

Methods

compute_backward(self)

Computes one step bringing the static image towards the moving.

compute_demons_step(self[, forward_step])

Demons step for SSD metric

compute_forward(self)

Computes one step bringing the reference image towards the static.

compute_gauss_newton_step(self[, forward_step])

Computes the Gauss-Newton energy minimization step

free_iteration(self)

Nothing to free for the SSD metric

get_energy(self)

The numerical value assigned by this metric to the current image pair

initialize_iteration(self)

Prepares the metric to compute one displacement field iteration.

set_levels_above(self, levels)

Informs the metric how many pyramid levels are above the current one

set_levels_below(self, levels)

Informs the metric how many pyramid levels are below the current one

set_moving_image(self, moving_image, …)

Sets the moving image being compared against the static one.

set_static_image(self, static_image, …)

Sets the static image being compared against the moving one.

use_moving_image_dynamics(self, …)

This is called by the optimizer just after setting the moving image

use_static_image_dynamics(self, …)

This is called by the optimizer just after setting the static image.

__init__(self, dim, smooth=4, inner_iter=10, step_type='demons')

Sum of Squared Differences (SSD) Metric

Similarity metric for (mono-modal) nonlinear image registration defined by the sum of squared differences (SSD)

Parameters
dimint (either 2 or 3)

the dimension of the image domain

smoothfloat

smoothness parameter, the larger the value the smoother the deformation field

inner_iterint

number of iterations to be performed at each level of the multi- resolution Gauss-Seidel optimization algorithm (this is not the number of steps per Gaussian Pyramid level, that parameter must be set for the optimizer, not the metric)

step_typestring

the displacement field step to be computed when ‘compute_forward’ and ‘compute_backward’ are called. Either ‘demons’ or ‘gauss_newton’

compute_backward(self)

Computes one step bringing the static image towards the moving.

Computes the update displacement field to be used for registration of the static image towards the moving image

compute_demons_step(self, forward_step=True)

Demons step for SSD metric

Computes the demons step proposed by Vercauteren et al.[Vercauteren09] for the SSD metric.

Parameters
forward_stepboolean

if True, computes the Demons step in the forward direction (warping the moving towards the static image). If False, computes the backward step (warping the static image to the moving image)

Returns
displacementarray, shape (R, C, 2) or (S, R, C, 3)

the Demons step

References

[Vercauteren09] Tom Vercauteren, Xavier Pennec, Aymeric Perchant,

Nicholas Ayache, “Diffeomorphic Demons: Efficient Non-parametric Image Registration”, Neuroimage 2009

compute_forward(self)

Computes one step bringing the reference image towards the static.

Computes the update displacement field to be used for registration of the moving image towards the static image

compute_gauss_newton_step(self, forward_step=True)

Computes the Gauss-Newton energy minimization step

Minimizes the linearized energy function (Newton step) defined by the sum of squared differences of corresponding pixels of the input images with respect to the displacement field.

Parameters
forward_stepboolean

if True, computes the Newton step in the forward direction (warping the moving towards the static image). If False, computes the backward step (warping the static image to the moving image)

Returns
displacementarray, shape = static_image.shape + (3,)

if forward_step==True, the forward SSD Gauss-Newton step, else, the backward step

free_iteration(self)

Nothing to free for the SSD metric

get_energy(self)

The numerical value assigned by this metric to the current image pair

Returns the Sum of Squared Differences (data term) energy computed at the largest iteration

initialize_iteration(self)

Prepares the metric to compute one displacement field iteration.

Pre-computes the gradient of the input images to be used in the computation of the forward and backward steps.

SimilarityMetric

class dipy.align.metrics.SimilarityMetric(dim)

Bases: object

Methods

compute_backward(self)

Computes one step bringing the static image towards the moving.

compute_forward(self)

Computes one step bringing the reference image towards the static.

free_iteration(self)

Releases the resources no longer needed by the metric

get_energy(self)

Numerical value assigned by this metric to the current image pair

initialize_iteration(self)

Prepares the metric to compute one displacement field iteration.

set_levels_above(self, levels)

Informs the metric how many pyramid levels are above the current one

set_levels_below(self, levels)

Informs the metric how many pyramid levels are below the current one

set_moving_image(self, moving_image, …)

Sets the moving image being compared against the static one.

set_static_image(self, static_image, …)

Sets the static image being compared against the moving one.

use_moving_image_dynamics(self, …)

This is called by the optimizer just after setting the moving image

use_static_image_dynamics(self, …)

This is called by the optimizer just after setting the static image.

__init__(self, dim)

Similarity Metric abstract class

A similarity metric is in charge of keeping track of the numerical value of the similarity (or distance) between the two given images. It also computes the update field for the forward and inverse displacement fields to be used in a gradient-based optimization algorithm. Note that this metric does not depend on any transformation (affine or non-linear) so it assumes the static and moving images are already warped

Parameters
dimint (either 2 or 3)

the dimension of the image domain

abstract compute_backward(self)

Computes one step bringing the static image towards the moving.

Computes the backward update field to register the static image towards the moving image in a gradient-based optimization algorithm

abstract compute_forward(self)

Computes one step bringing the reference image towards the static.

Computes the forward update field to register the moving image towards the static image in a gradient-based optimization algorithm

abstract free_iteration(self)

Releases the resources no longer needed by the metric

This method is called by the RegistrationOptimizer after the required iterations have been computed (forward and / or backward) so that the SimilarityMetric can safely delete any data it computed as part of the initialization

abstract get_energy(self)

Numerical value assigned by this metric to the current image pair

Must return the numeric value of the similarity between the given static and moving images

abstract initialize_iteration(self)

Prepares the metric to compute one displacement field iteration.

This method will be called before any compute_forward or compute_backward call, this allows the Metric to pre-compute any useful information for speeding up the update computations. This initialization was needed in ANTS because the updates are called once per voxel. In Python this is unpractical, though.

set_levels_above(self, levels)

Informs the metric how many pyramid levels are above the current one

Informs this metric the number of pyramid levels above the current one. The metric may change its behavior (e.g. number of inner iterations) accordingly

Parameters
levelsint

the number of levels above the current Gaussian Pyramid level

set_levels_below(self, levels)

Informs the metric how many pyramid levels are below the current one

Informs this metric the number of pyramid levels below the current one. The metric may change its behavior (e.g. number of inner iterations) accordingly

Parameters
levelsint

the number of levels below the current Gaussian Pyramid level

set_moving_image(self, moving_image, moving_affine, moving_spacing, moving_direction)

Sets the moving image being compared against the static one.

Sets the moving image. The default behavior (of this abstract class) is simply to assign the reference to an attribute, but generalizations of the metric may need to perform other operations

Parameters
moving_imagearray, shape (R, C) or (S, R, C)

the moving image

set_static_image(self, static_image, static_affine, static_spacing, static_direction)

Sets the static image being compared against the moving one.

Sets the static image. The default behavior (of this abstract class) is simply to assign the reference to an attribute, but generalizations of the metric may need to perform other operations

Parameters
static_imagearray, shape (R, C) or (S, R, C)

the static image

use_moving_image_dynamics(self, original_moving_image, transformation)

This is called by the optimizer just after setting the moving image

This method allows the metric to compute any useful information from knowing how the current static image was generated (as the transformation of an original static image). This method is called by the optimizer just after it sets the static image. Transformation will be an instance of DiffeomorficMap or None if the original_moving_image equals self.moving_image.

Parameters
original_moving_imagearray, shape (R, C) or (S, R, C)

original image from which the current moving image was generated

transformationDiffeomorphicMap object

the transformation that was applied to original image to generate the current moving image

use_static_image_dynamics(self, original_static_image, transformation)

This is called by the optimizer just after setting the static image.

This method allows the metric to compute any useful information from knowing how the current static image was generated (as the transformation of an original static image). This method is called by the optimizer just after it sets the static image. Transformation will be an instance of DiffeomorficMap or None if the original_static_image equals self.moving_image.

Parameters
original_static_imagearray, shape (R, C) or (S, R, C)

original image from which the current static image was generated

transformationDiffeomorphicMap object

the transformation that was applied to original image to generate the current static image

floating

dipy.align.metrics.floating

alias of numpy.float32

gradient

dipy.align.metrics.gradient(f, *varargs, **kwargs)

Return the gradient of an N-dimensional array.

The gradient is computed using second order accurate central differences in the interior points and either first or second order accurate one-sides (forward or backwards) differences at the boundaries. The returned gradient hence has the same shape as the input array.

Parameters
farray_like

An N-dimensional array containing samples of a scalar function.

varargslist of scalar or array, optional

Spacing between f values. Default unitary spacing for all dimensions. Spacing can be specified using:

  1. single scalar to specify a sample distance for all dimensions.

  2. N scalars to specify a constant sample distance for each dimension. i.e. dx, dy, dz, …

  3. N arrays to specify the coordinates of the values along each dimension of F. The length of the array must match the size of the corresponding dimension

  4. Any combination of N scalars/arrays with the meaning of 2. and 3.

If axis is given, the number of varargs must equal the number of axes. Default: 1.

edge_order{1, 2}, optional

Gradient is calculated using N-th order accurate differences at the boundaries. Default: 1.

New in version 1.9.1.

axisNone or int or tuple of ints, optional

Gradient is calculated only along the given axis or axes The default (axis = None) is to calculate the gradient for all the axes of the input array. axis may be negative, in which case it counts from the last to the first axis.

New in version 1.11.0.

Returns
gradientndarray or list of ndarray

A set of ndarrays (or a single ndarray if there is only one dimension) corresponding to the derivatives of f with respect to each dimension. Each derivative has the same shape as f.

Notes

Assuming that \(f\in C^{3}\) (i.e., \(f\) has at least 3 continuous derivatives) and let \(h_{*}\) be a non-homogeneous stepsize, we minimize the “consistency error” \(\eta_{i}\) between the true gradient and its estimate from a linear combination of the neighboring grid-points:

\[\eta_{i} = f_{i}^{\left(1\right)} - \left[ \alpha f\left(x_{i}\right) + \beta f\left(x_{i} + h_{d}\right) + \gamma f\left(x_{i}-h_{s}\right) \right]\]

By substituting \(f(x_{i} + h_{d})\) and \(f(x_{i} - h_{s})\) with their Taylor series expansion, this translates into solving the following the linear system:

\[\begin{split}\left\{ \begin{array}{r} \alpha+\beta+\gamma=0 \\ \beta h_{d}-\gamma h_{s}=1 \\ \beta h_{d}^{2}+\gamma h_{s}^{2}=0 \end{array} \right.\end{split}\]

The resulting approximation of \(f_{i}^{(1)}\) is the following:

\[\hat f_{i}^{(1)} = \frac{ h_{s}^{2}f\left(x_{i} + h_{d}\right) + \left(h_{d}^{2} - h_{s}^{2}\right)f\left(x_{i}\right) - h_{d}^{2}f\left(x_{i}-h_{s}\right)} { h_{s}h_{d}\left(h_{d} + h_{s}\right)} + \mathcal{O}\left(\frac{h_{d}h_{s}^{2} + h_{s}h_{d}^{2}}{h_{d} + h_{s}}\right)\]

It is worth noting that if \(h_{s}=h_{d}\) (i.e., data are evenly spaced) we find the standard second order approximation:

\[\hat f_{i}^{(1)}= \frac{f\left(x_{i+1}\right) - f\left(x_{i-1}\right)}{2h} + \mathcal{O}\left(h^{2}\right)\]

With a similar procedure the forward/backward approximations used for boundaries can be derived.

References

1

Quarteroni A., Sacco R., Saleri F. (2007) Numerical Mathematics (Texts in Applied Mathematics). New York: Springer.

2

Durran D. R. (1999) Numerical Methods for Wave Equations in Geophysical Fluid Dynamics. New York: Springer.

3

Fornberg B. (1988) Generation of Finite Difference Formulas on Arbitrarily Spaced Grids, Mathematics of Computation 51, no. 184 : 699-706. PDF.

Examples

>>> f = np.array([1, 2, 4, 7, 11, 16], dtype=float)
>>> np.gradient(f)
array([1. , 1.5, 2.5, 3.5, 4.5, 5. ])
>>> np.gradient(f, 2)
array([0.5 ,  0.75,  1.25,  1.75,  2.25,  2.5 ])

Spacing can be also specified with an array that represents the coordinates of the values F along the dimensions. For instance a uniform spacing:

>>> x = np.arange(f.size)
>>> np.gradient(f, x)
array([1. ,  1.5,  2.5,  3.5,  4.5,  5. ])

Or a non uniform one:

>>> x = np.array([0., 1., 1.5, 3.5, 4., 6.], dtype=float)
>>> np.gradient(f, x)
array([1. ,  3. ,  3.5,  6.7,  6.9,  2.5])

For two dimensional arrays, the return will be two arrays ordered by axis. In this example the first array stands for the gradient in rows and the second one in columns direction:

>>> np.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. ]])]

In this example the spacing is also specified: uniform for axis=0 and non uniform for axis=1

>>> dx = 2.
>>> y = [1., 1.5, 3.5]
>>> np.gradient(np.array([[1, 2, 6], [3, 4, 5]], dtype=float), dx, y)
[array([[ 1. ,  1. , -0.5],
       [ 1. ,  1. , -0.5]]), array([[2. , 2. , 2. ],
       [2. , 1.7, 0.5]])]

It is possible to specify how boundaries are treated using edge_order

>>> x = np.array([0, 1, 2, 3, 4])
>>> f = x**2
>>> np.gradient(f, edge_order=1)
array([1.,  2.,  4.,  6.,  7.])
>>> np.gradient(f, edge_order=2)
array([0., 2., 4., 6., 8.])

The axis keyword can be used to specify a subset of axes of which the gradient is calculated

>>> np.gradient(np.array([[1, 2, 6], [3, 4, 5]], dtype=float), axis=0)
array([[ 2.,  2., -1.],
       [ 2.,  2., -1.]])

v_cycle_2d

dipy.align.metrics.v_cycle_2d(n, k, delta_field, sigma_sq_field, gradient_field, target, lambda_param, displacement, depth=0)

Multi-resolution Gauss-Seidel solver using V-type cycles

Multi-resolution Gauss-Seidel solver: solves the Gauss-Newton linear system by first filtering (GS-iterate) the current level, then solves for the residual at a coarser resolution and finally refines the solution at the current resolution. This scheme corresponds to the V-cycle proposed by Bruhn and Weickert[Bruhn05].

Parameters
nint

number of levels of the multi-resolution algorithm (it will be called recursively until level n == 0)

kint

the number of iterations at each multi-resolution level

delta_fieldarray, shape (R, C)

the difference between the static and moving image (the ‘derivative w.r.t. time’ in the optical flow model)

sigma_sq_fieldarray, shape (R, C)

the variance of the gray level value at each voxel, according to the EM model (for SSD, it is 1 for all voxels). Inf and 0 values are processed specially to support infinite and zero variance.

gradient_fieldarray, shape (R, C, 2)

the gradient of the moving image

targetarray, shape (R, C, 2)

right-hand side of the linear system to be solved in the Weickert’s multi-resolution algorithm

lambda_paramfloat

smoothness parameter, the larger its value the smoother the displacement field

displacementarray, shape (R, C, 2)

the displacement field to start the optimization from

Returns
energythe energy of the EM (or SSD if sigmafield[…]==1) metric at this

iteration

References

[Bruhn05] Andres Bruhn and Joachim Weickert, “Towards ultimate motion

estimation: combining highest accuracy with real-time performance”, 10th IEEE International Conference on Computer Vision, 2005. ICCV 2005.

v_cycle_3d

dipy.align.metrics.v_cycle_3d(n, k, delta_field, sigma_sq_field, gradient_field, target, lambda_param, displacement, depth=0)

Multi-resolution Gauss-Seidel solver using V-type cycles

Multi-resolution Gauss-Seidel solver: solves the linear system by first filtering (GS-iterate) the current level, then solves for the residual at a coarser resolution and finally refines the solution at the current resolution. This scheme corresponds to the V-cycle proposed by Bruhn and Weickert[1]. [1] Andres Bruhn and Joachim Weickert, “Towards ultimate motion estimation:

combining highest accuracy with real-time performance”, 10th IEEE International Conference on Computer Vision, 2005. ICCV 2005.

Parameters
nint

number of levels of the multi-resolution algorithm (it will be called recursively until level n == 0)

kint

the number of iterations at each multi-resolution level

delta_fieldarray, shape (S, R, C)

the difference between the static and moving image (the ‘derivative w.r.t. time’ in the optical flow model)

sigma_sq_fieldarray, shape (S, R, C)

the variance of the gray level value at each voxel, according to the EM model (for SSD, it is 1 for all voxels). Inf and 0 values are processed specially to support infinite and zero variance.

gradient_fieldarray, shape (S, R, C, 3)

the gradient of the moving image

targetarray, shape (S, R, C, 3)

right-hand side of the linear system to be solved in the Weickert’s multi-resolution algorithm

lambda_paramfloat

smoothness parameter, the larger its value the smoother the displacement field

displacementarray, shape (S, R, C, 3)

the displacement field to start the optimization from

Returns
energythe energy of the EM (or SSD if sigmafield[…]==1) metric at this

iteration

Pool

dipy.align.reslice.Pool(processes=None, initializer=None, initargs=(), maxtasksperchild=None)

Returns a process pool object

affine_transform

dipy.align.reslice.affine_transform(input, matrix, offset=0.0, output_shape=None, output=None, order=3, mode='constant', cval=0.0, prefilter=True)

Apply an affine transformation.

Given an output image pixel index vector o, the pixel value is determined from the input image at position np.dot(matrix, o) + offset.

This does ‘pull’ (or ‘backward’) resampling, transforming the output space to the input to locate data. Affine transformations are often described in the ‘push’ (or ‘forward’) direction, transforming input to output. If you have a matrix for the ‘push’ transformation, use its inverse (numpy.linalg.inv()) in this function.

Parameters
inputarray_like

The input array.

matrixndarray

The inverse coordinate transformation matrix, mapping output coordinates to input coordinates. If ndim is the number of dimensions of input, the given matrix must have one of the following shapes:

  • (ndim, ndim): the linear transformation matrix for each output coordinate.

  • (ndim,): assume that the 2D transformation matrix is diagonal, with the diagonal specified by the given value. A more efficient algorithm is then used that exploits the separability of the problem.

  • (ndim + 1, ndim + 1): assume that the transformation is specified using homogeneous coordinates [1]. In this case, any value passed to offset is ignored.

  • (ndim, ndim + 1): as above, but the bottom row of a homogeneous transformation matrix is always [0, 0, ..., 1], and may be omitted.

offsetfloat or sequence, optional

The offset into the array where the transform is applied. If a float, offset is the same for each axis. If a sequence, offset should contain one value for each axis.

output_shapetuple of ints, optional

Shape tuple.

outputarray or dtype, optional

The array in which to place the output, or the dtype of the returned array. By default an array of the same dtype as input will be created.

orderint, optional

The order of the spline interpolation, default is 3. The order has to be in the range 0-5.

mode{‘reflect’, ‘constant’, ‘nearest’, ‘mirror’, ‘wrap’}, optional

The mode parameter determines how the input array is extended beyond its boundaries. Default is ‘constant’. Behavior for each valid value is as follows:

‘reflect’ (d c b a | a b c d | d c b a)

The input is extended by reflecting about the edge of the last pixel.

‘constant’ (k k k k | a b c d | k k k k)

The input is extended by filling all values beyond the edge with the same constant value, defined by the cval parameter.

‘nearest’ (a a a a | a b c d | d d d d)

The input is extended by replicating the last pixel.

‘mirror’ (d c b | a b c d | c b a)

The input is extended by reflecting about the center of the last pixel.

‘wrap’ (a b c d | a b c d | a b c d)

The input is extended by wrapping around to the opposite edge.

cvalscalar, optional

Value to fill past edges of input if mode is ‘constant’. Default is 0.0.

prefilterbool, optional

Determines if the input array is prefiltered with spline_filter before interpolation. The default is True, which will create a temporary float64 array of filtered values if order > 1. If setting this to False, the output will be slightly blurred if order > 1, unless the input is prefiltered, i.e. it is the result of calling spline_filter on the original input.

Returns
affine_transformndarray

The transformed input.

Notes

The given matrix and offset are used to find for each point in the output the corresponding coordinates in the input by an affine transformation. The value of the input at those coordinates is determined by spline interpolation of the requested order. Points outside the boundaries of the input are filled according to the given mode.

Changed in version 0.18.0: Previously, the exact interpretation of the affine transformation depended on whether the matrix was supplied as a one-dimensional or two-dimensional array. If a one-dimensional array was supplied to the matrix parameter, the output pixel value at index o was determined from the input image at position matrix * (o + offset).

References

1

https://en.wikipedia.org/wiki/Homogeneous_coordinates

cpu_count

dipy.align.reslice.cpu_count()

Returns the number of CPUs in the system

reslice

dipy.align.reslice.reslice(data, affine, zooms, new_zooms, order=1, mode='constant', cval=0, num_processes=1)

Reslice data with new voxel resolution defined by new_zooms

Parameters
dataarray, shape (I,J,K) or (I,J,K,N)

3d volume or 4d volume with datasets

affinearray, shape (4,4)

mapping from voxel coordinates to world coordinates

zoomstuple, shape (3,)

voxel size for (i,j,k) dimensions

new_zoomstuple, shape (3,)

new voxel size for (i,j,k) after resampling

orderint, from 0 to 5

order of interpolation for resampling/reslicing, 0 nearest interpolation, 1 trilinear etc.. if you don’t want any smoothing 0 is the option you need.

modestring (‘constant’, ‘nearest’, ‘reflect’ or ‘wrap’)

Points outside the boundaries of the input are filled according to the given mode.

cvalfloat

Value used for points outside the boundaries of the input if mode=’constant’.

num_processesint

Split the calculation to a pool of children processes. This only applies to 4D data arrays. If a positive integer then it defines the size of the multiprocessing pool that will be used. If 0, then the size of the pool will equal the number of cores available.

Returns
data2array, shape (I,J,K) or (I,J,K,N)

datasets resampled into isotropic voxel size

affine2array, shape (4,4)

new affine for the resampled image

Examples

>>> from dipy.io.image import load_nifti
>>> from dipy.align.reslice import reslice
>>> from dipy.data import get_fnames
>>> f_name = get_fnames('aniso_vox')
>>> data, affine, zooms = load_nifti(f_name, return_voxsize=True)
>>> data.shape == (58, 58, 24)
True
>>> zooms
(4.0, 4.0, 5.0)
>>> new_zooms = (3.,3.,3.)
>>> new_zooms
(3.0, 3.0, 3.0)
>>> data2, affine2 = reslice(data, affine, zooms, new_zooms)
>>> data2.shape == (77, 77, 40)
True

IsotropicScaleSpace

class dipy.align.scalespace.IsotropicScaleSpace(image, factors, sigmas, image_grid2world=None, input_spacing=None, mask0=False)

Bases: dipy.align.scalespace.ScaleSpace

Methods

get_affine(self, level)

Voxel-to-space transformation at a given level

get_affine_inv(self, level)

Space-to-voxel transformation at a given level

get_domain_shape(self, level)

Shape the sub-sampled image must have at a particular level

get_expand_factors(self, from_level, to_level)

Ratio of voxel size from pyramid level from_level to to_level

get_image(self, level)

Smoothed image at a given level

get_scaling(self, level)

Adjustment factor for input-spacing to reflect voxel sizes at level

get_sigmas(self, level)

Smoothing parameters used at a given level

get_spacing(self, level)

Spacings the sub-sampled image must have at a particular level

print_level(self, level)

Prints properties of a pyramid level

__init__(self, image, factors, sigmas, image_grid2world=None, input_spacing=None, mask0=False)

IsotropicScaleSpace

Computes the Scale Space representation of an image using isotropic smoothing kernels for all scales. The scale space is simply a list of images produced by smoothing the input image with a Gaussian kernel with different smoothing parameters.

This specialization of ScaleSpace allows the user to provide custom scale and smoothing factors for all scales.

Parameters
imagearray, shape (r,c) or (s, r, c) where s is the number of

slices, r is the number of rows and c is the number of columns of the input image.

factorslist of floats

custom scale factors to build the scale space (one factor for each scale).

sigmaslist of floats

custom smoothing parameter to build the scale space (one parameter for each scale).

image_grid2worldarray, shape (dim + 1, dim + 1), optional

the grid-to-space transform of the image grid. The default is the identity matrix.

input_spacingarray, shape (dim,), optional

the spacing (voxel size) between voxels in physical space. The default if 1.0 along all axes.

mask0Boolean, optional

if True, all smoothed images will be zero at all voxels that are zero in the input image. The default is False.

ScaleSpace

class dipy.align.scalespace.ScaleSpace(image, num_levels, image_grid2world=None, input_spacing=None, sigma_factor=0.2, mask0=False)

Bases: object

Methods

get_affine(self, level)

Voxel-to-space transformation at a given level

get_affine_inv(self, level)

Space-to-voxel transformation at a given level

get_domain_shape(self, level)

Shape the sub-sampled image must have at a particular level

get_expand_factors(self, from_level, to_level)

Ratio of voxel size from pyramid level from_level to to_level

get_image(self, level)

Smoothed image at a given level

get_scaling(self, level)

Adjustment factor for input-spacing to reflect voxel sizes at level

get_sigmas(self, level)

Smoothing parameters used at a given level

get_spacing(self, level)

Spacings the sub-sampled image must have at a particular level

print_level(self, level)

Prints properties of a pyramid level

__init__(self, image, num_levels, image_grid2world=None, input_spacing=None, sigma_factor=0.2, mask0=False)

ScaleSpace

Computes the Scale Space representation of an image. The scale space is simply a list of images produced by smoothing the input image with a Gaussian kernel with increasing smoothing parameter. If the image’s voxels are isotropic, the smoothing will be the same along all directions: at level L = 0, 1, …, the sigma is given by \(s * ( 2^L - 1 )\). If the voxel dimensions are not isotropic, then the smoothing is weaker along low resolution directions.

Parameters
imagearray, shape (r,c) or (s, r, c) where s is the number of

slices, r is the number of rows and c is the number of columns of the input image.

num_levelsint

the desired number of levels (resolutions) of the scale space

image_grid2worldarray, shape (dim + 1, dim + 1), optional

the grid-to-space transform of the image grid. The default is the identity matrix

input_spacingarray, shape (dim,), optional

the spacing (voxel size) between voxels in physical space. The default is 1.0 along all axes

sigma_factorfloat, optional

the smoothing factor to be used in the construction of the scale space. The default is 0.2

mask0Boolean, optional

if True, all smoothed images will be zero at all voxels that are zero in the input image. The default is False.

get_affine(self, level)

Voxel-to-space transformation at a given level

Returns the voxel-to-space transformation associated with the sub-sampled image at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get affine transform from

Returns
the affine (voxel-to-space) transform at the requested resolution

or None if an invalid level was requested

get_affine_inv(self, level)

Space-to-voxel transformation at a given level

Returns the space-to-voxel transformation associated with the sub-sampled image at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the inverse transform from

Returns
the inverse (space-to-voxel) transform at the requested resolution or
None if an invalid level was requested
get_domain_shape(self, level)

Shape the sub-sampled image must have at a particular level

Returns the shape the sub-sampled image must have at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the sub-sampled shape from

Returns
the sub-sampled shape at the requested resolution or None if an

invalid level was requested

get_expand_factors(self, from_level, to_level)

Ratio of voxel size from pyramid level from_level to to_level

Given two scale space resolutions a = from_level, b = to_level, returns the ratio of voxels size at level b to voxel size at level a (the factor that must be used to multiply voxels at level a to ‘expand’ them to level b).

Parameters
from_levelint, 0 <= from_level < L, (L = number of resolutions)

the resolution to expand voxels from

to_levelint, 0 <= to_level < from_level

the resolution to expand voxels to

Returns
factorsarray, shape (k,), k = 2, 3

the expand factors (a scalar for each voxel dimension)

get_image(self, level)

Smoothed image at a given level

Returns the smoothed image at the requested level in the Scale Space.

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the smooth image from

Returns
the smooth image at the requested resolution or None if an invalid

level was requested

get_scaling(self, level)

Adjustment factor for input-spacing to reflect voxel sizes at level

Returns the scaling factor that needs to be applied to the input spacing (the voxel sizes of the image at level 0 of the scale space) to transform them to voxel sizes at the requested level.

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the scalings from

Returns
the scaling factors from the original spacing to the spacings at the
requested level
get_sigmas(self, level)

Smoothing parameters used at a given level

Returns the smoothing parameters (a scalar for each axis) used at the requested level of the scale space

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the smoothing parameters from

Returns
the smoothing parameters at the requested level
get_spacing(self, level)

Spacings the sub-sampled image must have at a particular level

Returns the spacings (voxel sizes) the sub-sampled image must have at a particular resolution of the scale space (note that this object does not explicitly subsample the smoothed images, but only provides the properties the sub-sampled images must have).

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to get the sub-sampled shape from

Returns
the spacings (voxel sizes) at the requested resolution or None if an
invalid level was requested
print_level(self, level)

Prints properties of a pyramid level

Prints the properties of a level of this scale space to standard output

Parameters
levelint, 0 <= from_level < L, (L = number of resolutions)

the scale space level to be printed

floating

dipy.align.scalespace.floating

alias of numpy.float32

BundleMinDistanceAsymmetricMetric

class dipy.align.streamlinear.BundleMinDistanceAsymmetricMetric(num_threads=None)

Bases: dipy.align.streamlinear.BundleMinDistanceMetric

Asymmetric Bundle-based Minimum distance

This is a cost function that can be used by the StreamlineLinearRegistration class.

Methods

distance(self, xopt)

Distance calculated from this Metric

setup(self, static, moving)

Setup static and moving sets of streamlines

__init__(self, num_threads=None)

An abstract class for the metric used for streamline registration

If the two sets of streamlines match exactly then method distance of this object should be minimum.

Parameters
num_threadsint

Number of threads. If None (default) then all available threads will be used. Only metrics using OpenMP will use this variable.

distance(self, xopt)

Distance calculated from this Metric

Parameters
xoptsequence

List of affine parameters as an 1D vector

BundleMinDistanceMatrixMetric

class dipy.align.streamlinear.BundleMinDistanceMatrixMetric(num_threads=None)

Bases: dipy.align.streamlinear.StreamlineDistanceMetric

Bundle-based Minimum Distance aka BMD

This is the cost function used by the StreamlineLinearRegistration

Notes

The difference with BundleMinDistanceMetric is that this creates the entire distance matrix and therefore requires more memory.

Methods

setup(static, moving)

distance(xopt)

__init__(self, num_threads=None)

An abstract class for the metric used for streamline registration

If the two sets of streamlines match exactly then method distance of this object should be minimum.

Parameters
num_threadsint

Number of threads. If None (default) then all available threads will be used. Only metrics using OpenMP will use this variable.

distance(self, xopt)

Distance calculated from this Metric

Parameters
xoptsequence

List of affine parameters as an 1D vector

setup(self, static, moving)

Setup static and moving sets of streamlines

Parameters
staticstreamlines

Fixed or reference set of streamlines.

movingstreamlines

Moving streamlines.

Notes

Call this after the object is initiated and before distance.

Num_threads is not used in this class. Use BundleMinDistanceMetric for a faster, threaded and less memory hungry metric

BundleMinDistanceMetric

class dipy.align.streamlinear.BundleMinDistanceMetric(num_threads=None)

Bases: dipy.align.streamlinear.StreamlineDistanceMetric

Bundle-based Minimum Distance aka BMD

This is the cost function used by the StreamlineLinearRegistration

References

Ra052abf06f20-Garyfallidis14

Garyfallidis et al., “Direct native-space fiber bundle alignment for group comparisons”, ISMRM, 2014.

Methods

setup(static, moving)

distance(xopt)

__init__(self, num_threads=None)

An abstract class for the metric used for streamline registration

If the two sets of streamlines match exactly then method distance of this object should be minimum.

Parameters
num_threadsint

Number of threads. If None (default) then all available threads will be used. Only metrics using OpenMP will use this variable.

distance(self, xopt)

Distance calculated from this Metric

Parameters
xoptsequence

List of affine parameters as an 1D vector,

setup(self, static, moving)

Setup static and moving sets of streamlines

Parameters
staticstreamlines

Fixed or reference set of streamlines.

movingstreamlines

Moving streamlines.

num_threadsint

Number of threads. If None (default) then all available threads will be used.

Notes

Call this after the object is initiated and before distance.

BundleSumDistanceMatrixMetric

class dipy.align.streamlinear.BundleSumDistanceMatrixMetric(num_threads=None)

Bases: dipy.align.streamlinear.BundleMinDistanceMatrixMetric

Bundle-based Sum Distance aka BMD

This is a cost function that can be used by the StreamlineLinearRegistration class.

Notes

The difference with BundleMinDistanceMatrixMetric is that it uses uses the sum of the distance matrix and not the sum of mins.

Methods

setup(static, moving)

distance(xopt)

__init__(self, num_threads=None)

An abstract class for the metric used for streamline registration

If the two sets of streamlines match exactly then method distance of this object should be minimum.

Parameters
num_threadsint

Number of threads. If None (default) then all available threads will be used. Only metrics using OpenMP will use this variable.

distance(self, xopt)

Distance calculated from this Metric

Parameters
xoptsequence

List of affine parameters as an 1D vector

Optimizer

class dipy.align.streamlinear.Optimizer(fun, x0, args=(), method='L-BFGS-B', jac=None, hess=None, hessp=None, bounds=None, constraints=(), tol=None, callback=None, options=None, evolution=False)

Bases: object

Attributes
evolution
fopt
message
nfev
nit
xopt

Methods

print_summary

__init__(self, fun, x0, args=(), method='L-BFGS-B', jac=None, hess=None, hessp=None, bounds=None, constraints=(), tol=None, callback=None, options=None, evolution=False)

A class for handling minimization of scalar function of one or more variables.

Parameters
funcallable

Objective function.

x0ndarray

Initial guess.

argstuple, optional

Extra arguments passed to the objective function and its derivatives (Jacobian, Hessian).

methodstr, optional

Type of solver. Should be one of

  • ‘Nelder-Mead’

  • ‘Powell’

  • ‘CG’

  • ‘BFGS’

  • ‘Newton-CG’

  • ‘Anneal’

  • ‘L-BFGS-B’

  • ‘TNC’

  • ‘COBYLA’

  • ‘SLSQP’

  • ‘dogleg’

  • ‘trust-ncg’

jacbool or callable, optional

Jacobian of objective function. Only for CG, BFGS, Newton-CG, dogleg, trust-ncg. If jac is a Boolean and is True, fun is assumed to return the value of Jacobian along with the objective function. If False, the Jacobian will be estimated numerically. jac can also be a callable returning the Jacobian of the objective. In this case, it must accept the same arguments as fun.

hess, hesspcallable, optional

Hessian of objective function or Hessian of objective function times an arbitrary vector p. Only for Newton-CG, dogleg, trust-ncg. Only one of hessp or hess needs to be given. If hess is provided, then hessp will be ignored. If neither hess nor hessp is provided, then the hessian product will be approximated using finite differences on jac. hessp must compute the Hessian times an arbitrary vector.

boundssequence, optional

Bounds for variables (only for L-BFGS-B, TNC and SLSQP). (min, max) pairs for each element in x, defining the bounds on that parameter. Use None for one of min or max when there is no bound in that direction.

constraintsdict or sequence of dict, optional

Constraints definition (only for COBYLA and SLSQP). Each constraint is defined in a dictionary with fields:

typestr

Constraint type: ‘eq’ for equality, ‘ineq’ for inequality.

funcallable

The function defining the constraint.

jaccallable, optional

The Jacobian of fun (only for SLSQP).

argssequence, optional

Extra arguments to be passed to the function and Jacobian.

Equality constraint means that the constraint function result is to be zero whereas inequality means that it is to be non-negative. Note that COBYLA only supports inequality constraints.

tolfloat, optional

Tolerance for termination. For detailed control, use solver-specific options.

callbackcallable, optional

Called after each iteration, as callback(xk), where xk is the current parameter vector. Only available using Scipy >= 0.12.

optionsdict, optional

A dictionary of solver options. All methods accept the following generic options:

maxiterint

Maximum number of iterations to perform.

dispbool

Set to True to print convergence messages.

For method-specific options, see show_options(‘minimize’, method).

evolutionbool, optional

save history of x for each iteration. Only available using Scipy >= 0.12.

See also

scipy.optimize.minimize
property evolution
property fopt
property message
property nfev
property nit
print_summary(self)
property xopt

StreamlineDistanceMetric

class dipy.align.streamlinear.StreamlineDistanceMetric(num_threads=None)

Bases: object

Methods

distance(self, xopt)

calculate distance for current set of parameters

setup

__init__(self, num_threads=None)

An abstract class for the metric used for streamline registration

If the two sets of streamlines match exactly then method distance of this object should be minimum.

Parameters
num_threadsint

Number of threads. If None (default) then all available threads will be used. Only metrics using OpenMP will use this variable.

abstract distance(self, xopt)

calculate distance for current set of parameters

abstract setup(self, static, moving)

StreamlineLinearRegistration

class dipy.align.streamlinear.StreamlineLinearRegistration(metric=None, x0='rigid', method='L-BFGS-B', bounds=None, verbose=False, options=None, evolution=False, num_threads=None)

Bases: object

Methods

optimize(self, static, moving[, mat])

Find the minimum of the provided metric.

__init__(self, metric=None, x0='rigid', method='L-BFGS-B', bounds=None, verbose=False, options=None, evolution=False, num_threads=None)

Linear registration of 2 sets of streamlines [Garyfallidis15].

Parameters
metricStreamlineDistanceMetric,

If None and fast is False then the BMD distance is used. If fast is True then a faster implementation of BMD is used. Otherwise, use the given distance metric.

x0array or int or str

Initial parametrization for the optimization.

If 1D array with:

a) 6 elements then only rigid registration is performed with the 3 first elements for translation and 3 for rotation. b) 7 elements also isotropic scaling is performed (similarity). c) 12 elements then translation, rotation (in degrees), scaling and shearing is performed (affine).

Here is an example of x0 with 12 elements: x0=np.array([0, 10, 0, 40, 0, 0, 2., 1.5, 1, 0.1, -0.5, 0])

This has translation (0, 10, 0), rotation (40, 0, 0) in degrees, scaling (2., 1.5, 1) and shearing (0.1, -0.5, 0).

If int:
  1. 6

    x0 = np.array([0, 0, 0, 0, 0, 0])

  2. 7

    x0 = np.array([0, 0, 0, 0, 0, 0, 1.])

  3. 12

    x0 = np.array([0, 0, 0, 0, 0, 0, 1., 1., 1, 0, 0, 0])

If str:
  1. “rigid”

    x0 = np.array([0, 0, 0, 0, 0, 0])

  2. “similarity”

    x0 = np.array([0, 0, 0, 0, 0, 0, 1.])

  3. “affine”

    x0 = np.array([0, 0, 0, 0, 0, 0, 1., 1., 1, 0, 0, 0])

methodstr,

‘L_BFGS_B’ or ‘Powell’ optimizers can be used. Default is ‘L_BFGS_B’.

boundslist of tuples or None,

If method == ‘L_BFGS_B’ then we can use bounded optimization. For example for the six parameters of rigid rotation we can set the bounds = [(-30, 30), (-30, 30), (-30, 30),

(-45, 45), (-45, 45), (-45, 45)]

That means that we have set the bounds for the three translations and three rotation axes (in degrees).

verbosebool, optional.

If True, if True then information about the optimization is shown. Default: False.

optionsNone or dict,

Extra options to be used with the selected method.

evolutionboolean

If True save the transformation for each iteration of the optimizer. Default is False. Supported only with Scipy >= 0.11.

num_threadsint

Number of threads. If None (default) then all available threads will be used. Only metrics using OpenMP will use this variable.

References

Garyfallidis15

Garyfallidis et al. “Robust and efficient linear registration of white-matter fascicles in the space of streamlines”, NeuroImage, 117, 124–140, 2015

Garyfallidis14

Garyfallidis et al., “Direct native-space fiber bundle alignment for group comparisons”, ISMRM, 2014.

Garyfallidis17

Garyfallidis et al. Recognition of white matter bundles using local and global streamline-based registration and clustering, Neuroimage, 2017.

optimize(self, static, moving, mat=None)

Find the minimum of the provided metric.

Parameters
staticstreamlines

Reference or fixed set of streamlines.

movingstreamlines

Moving set of streamlines.

matarray

Transformation (4, 4) matrix to start the registration. mat is applied to moving. Default value None which means that initial transformation will be generated by shifting the centers of moving and static sets of streamlines to the origin.

Returns
mapStreamlineRegistrationMap

StreamlineRegistrationMap

class dipy.align.streamlinear.StreamlineRegistrationMap(matopt, xopt, fopt, matopt_history, funcs, iterations)

Bases: object

Methods

transform(self, moving)

Transform moving streamlines to the static.

__init__(self, matopt, xopt, fopt, matopt_history, funcs, iterations)

A map holding the optimum affine matrix and some other parameters of the optimization

Parameters
matrixarray,

4x4 affine matrix which transforms the moving to the static streamlines

xoptarray,

1d array with the parameters of the transformation after centering

foptfloat,

final value of the metric

matrix_historyarray

All transformation matrices created during the optimization

funcsint,

Number of function evaluations of the optimizer

iterationsint

Number of iterations of the optimizer

transform(self, moving)

Transform moving streamlines to the static.

Parameters
movingstreamlines
Returns
movedstreamlines

Notes

All this does is apply self.matrix to the input streamlines.

Streamlines

dipy.align.streamlinear.Streamlines

alias of nibabel.streamlines.array_sequence.ArraySequence

bundle_min_distance

dipy.align.streamlinear.bundle_min_distance(t, static, moving)

MDF-based pairwise distance optimization function (MIN)

We minimize the distance between moving streamlines as they align with the static streamlines.

Parameters
tndarray

t is a vector of of affine transformation parameters with size at least 6. If size is 6, t is interpreted as translation + rotation. If size is 7, t is interpreted as translation + rotation + isotropic scaling. If size is 12, t is interpreted as translation + rotation + scaling + shearing.

staticlist

Static streamlines

movinglist

Moving streamlines.

num_threadsint

Number of threads. If None (default) then all available threads will be used.

Returns
cost: float

bundle_min_distance_asymmetric_fast

dipy.align.streamlinear.bundle_min_distance_asymmetric_fast(t, static, moving, block_size)

MDF-based pairwise distance optimization function (MIN)

We minimize the distance between moving streamlines as they align with the static streamlines.

Parameters
tarray

1D array. t is a vector of of affine transformation parameters with size at least 6. If size is 6, t is interpreted as translation + rotation. If size is 7, t is interpreted as translation + rotation + isotropic scaling. If size is 12, t is interpreted as translation + rotation + scaling + shearing.

staticarray

N*M x 3 array. All the points of the static streamlines. With order of streamlines intact. Where N is the number of streamlines and M is the number of points per streamline.

movingarray

K*M x 3 array. All the points of the moving streamlines. With order of streamlines intact. Where K is the number of streamlines and M is the number of points per streamline.

block_sizeint

Number of points per streamline. All streamlines in static and moving should have the same number of points M.

Returns
cost: float

bundle_min_distance_fast

dipy.align.streamlinear.bundle_min_distance_fast(t, static, moving, block_size, num_threads)

MDF-based pairwise distance optimization function (MIN)

We minimize the distance between moving streamlines as they align with the static streamlines.

Parameters
tarray

1D array. t is a vector of of affine transformation parameters with size at least 6. If size is 6, t is interpreted as translation + rotation. If size is 7, t is interpreted as translation + rotation + isotropic scaling. If size is 12, t is interpreted as translation + rotation + scaling + shearing.

staticarray

N*M x 3 array. All the points of the static streamlines. With order of streamlines intact. Where N is the number of streamlines and M is the number of points per streamline.

movingarray

K*M x 3 array. All the points of the moving streamlines. With order of streamlines intact. Where K is the number of streamlines and M is the number of points per streamline.

block_sizeint

Number of points per streamline. All streamlines in static and moving should have the same number of points M.

num_threadsint

Number of threads. If None (default) then all available threads will be used.

Returns
cost: float

Notes

This is a faster implementation of bundle_min_distance, which requires that all the points of each streamline are allocated into an ndarray (of shape N*M by 3, with N the number of points per streamline and M the number of streamlines). This can be done by calling dipy.tracking.streamlines.unlist_streamlines.

bundle_sum_distance

dipy.align.streamlinear.bundle_sum_distance(t, static, moving, num_threads=None)

MDF distance optimization function (SUM)

We minimize the distance between moving streamlines as they align with the static streamlines.

Parameters
tndarray

t is a vector of of affine transformation parameters with size at least 6. If size is 6, t is interpreted as translation + rotation. If size is 7, t is interpreted as translation + rotation + isotropic scaling. If size is 12, t is interpreted as translation + rotation + scaling + shearing.

staticlist

Static streamlines

movinglist

Moving streamlines. These will be transform to align with the static streamlines

Returns
cost: float

center_streamlines

dipy.align.streamlinear.center_streamlines(streamlines)

Move streamlines to the origin

Parameters
streamlineslist

List of 2D ndarrays of shape[-1]==3

Returns
new_streamlineslist

List of 2D ndarrays of shape[-1]==3

inv_shiftndarray

Translation in x,y,z to go back in the initial position

compose_matrix

dipy.align.streamlinear.compose_matrix(scale=None, shear=None, angles=None, translate=None, perspective=None)

Return 4x4 transformation matrix from sequence of transformations.

Code modified from the work of Christoph Gohlke link provided here http://www.lfd.uci.edu/~gohlke/code/transformations.py.html

This is the inverse of the decompose_matrix function.

Parameters
scale(3,) array_like

Scaling factors.

sheararray_like

Shear factors for x-y, x-z, y-z axes.

anglesarray_like

Euler angles about static x, y, z axes.

translatearray_like

Translation vector along x, y, z axes.

perspectivearray_like

Perspective partition of matrix.

Returns
matrix4x4 array

Examples

>>> import math
>>> import numpy as np
>>> import dipy.core.geometry as gm
>>> scale = np.random.random(3) - 0.5
>>> shear = np.random.random(3) - 0.5
>>> angles = (np.random.random(3) - 0.5) * (2*math.pi)
>>> trans = np.random.random(3) - 0.5
>>> persp = np.random.random(4) - 0.5
>>> M0 = gm.compose_matrix(scale, shear, angles, trans, persp)

compose_matrix44

dipy.align.streamlinear.compose_matrix44(t, dtype=<class 'numpy.float64'>)

Compose a 4x4 transformation matrix

Parameters
tndarray

This is a 1D vector of of affine transformation parameters with size at least 3. If size is 3, t is interpreted as translation. If size is 6, t is interpreted as translation + rotation. If size is 7, t is interpreted as translation + rotation + isotropic scaling. If size is 9, t is interpreted as translation + rotation + anisotropic scaling. If size is 12, t is interpreted as translation + rotation + scaling + shearing.

Returns
Tndarray

Homogeneous transformation matrix of size 4x4.

compose_transformations

dipy.align.streamlinear.compose_transformations(*mats)

Compose multiple 4x4 affine transformations in one 4x4 matrix

Parameters
mat1array, (4, 4)
mat2array, (4, 4)
matNarray, (4, 4)
Returns
matN x … x mat2 x mat1array, (4, 4)

decompose_matrix

dipy.align.streamlinear.decompose_matrix(matrix)

Return sequence of transformations from transformation matrix.

Code modified from the excellent work of Christoph Gohlke link provided here: http://www.lfd.uci.edu/~gohlke/code/transformations.py.html

Parameters
matrixarray_like

Non-degenerative homogeneous transformation matrix

Returns
scale(3,) ndarray

Three scaling factors.

shear(3,) ndarray

Shear factors for x-y, x-z, y-z axes.

angles(3,) ndarray

Euler angles about static x, y, z axes.

translate(3,) ndarray

Translation vector along x, y, z axes.

perspectivendarray

Perspective partition of matrix.

Raises
ValueError

If matrix is of wrong type or degenerative.

Examples

>>> import numpy as np
>>> T0=np.diag([2,1,1,1])
>>> scale, shear, angles, trans, persp = decompose_matrix(T0)

decompose_matrix44

dipy.align.streamlinear.decompose_matrix44(mat, size=12)

Given a 4x4 homogeneous matrix return the parameter vector

Parameters
matarray

Homogeneous 4x4 transformation matrix

sizeint

Size of output vector. 3, for translation, 6 for rigid, 7 for similarity, 9 for scaling and 12 for affine. Default is 12.

Returns
tndarray

One dimensional ndarray of 3, 6, 7, 9 or 12 affine parameters.

distance_matrix_mdf

dipy.align.streamlinear.distance_matrix_mdf()

Minimum direct flipped distance matrix between two streamline sets

All streamlines need to have the same number of points

Parameters
streamlines_asequence

of streamlines as arrays, [(N, 3) .. (N, 3)]

streamlines_bsequence

of streamlines as arrays, [(N, 3) .. (N, 3)]

Returns
DMarray, shape (len(streamlines_a), len(streamlines_b))

distance matrix

length

dipy.align.streamlinear.length()

Euclidean length of streamlines

Length is in mm only if streamlines are expressed in world coordinates.

Parameters
streamlinesndarray or a list or dipy.tracking.Streamlines

If ndarray, must have shape (N,3) where N is the number of points of the streamline. If list, each item must be ndarray shape (Ni,3) where Ni is the number of points of streamline i. If dipy.tracking.Streamlines, its common_shape must be 3.

Returns
lengthsscalar or ndarray shape (N,)

If there is only one streamline, a scalar representing the length of the streamline. If there are several streamlines, ndarray containing the length of every streamline.

Examples

>>> from dipy.tracking.streamline import length
>>> import numpy as np
>>> streamline = np.array([[1, 1, 1], [2, 3, 4], [0, 0, 0]])
>>> expected_length = np.sqrt([1+2**2+3**2, 2**2+3**2+4**2]).sum()
>>> length(streamline) == expected_length
True
>>> streamlines = [streamline, np.vstack([streamline, streamline[::-1]])]
>>> expected_lengths = [expected_length, 2*expected_length]
>>> lengths = [length(streamlines[0]), length(streamlines[1])]
>>> np.allclose(lengths, expected_lengths)
True
>>> length([])
0.0
>>> length(np.array([[1, 2, 3]]))
0.0

progressive_slr

dipy.align.streamlinear.progressive_slr(static, moving, metric, x0, bounds, method='L-BFGS-B', verbose=False, num_threads=None)

Progressive SLR

This is an utility function that allows for example to do affine registration using Streamline-based Linear Registration (SLR) [Garyfallidis15] by starting with translation first, then rigid, then similarity, scaling and finally affine.

Similarly, if for example you want to perform rigid then you start with translation first. This progressive strategy can helps with finding the optimal parameters of the final transformation.

Parameters
staticStreamlines
movingStreamlines
metricStreamlineDistanceMetric
x0string

Could be any of ‘translation’, ‘rigid’, ‘similarity’, ‘scaling’, ‘affine’

boundsarray

Boundaries of registration parameters. See variable DEFAULT_BOUNDS for example.

methodstring

L_BFGS_B’ or ‘Powell’ optimizers can be used. Default is ‘L_BFGS_B’.

verbosebool, optional.

If True, log messages. Default:

num_threadsint

Number of threads. If None (default) then all available threads will be used. Only metrics using OpenMP will use this variable.

References

Garyfallidis15

Garyfallidis et al. “Robust and efficient linear registration of white-matter fascicles in the space of streamlines”, NeuroImage, 117, 124–140, 2015

qbx_and_merge

dipy.align.streamlinear.qbx_and_merge(streamlines, thresholds, nb_pts=20, select_randomly=None, rng=None, verbose=False)

Run QuickBundlesX and then run again on the centroids of the last layer

Running again QuickBundles at a layer has the effect of merging some of the clusters that maybe originally devided because of branching. This function help obtain a result at a QuickBundles quality but with QuickBundlesX speed. The merging phase has low cost because it is applied only on the centroids rather than the entire dataset.

Parameters
streamlinesStreamlines
thresholdssequence

List of distance thresholds for QuickBundlesX.

nb_ptsint

Number of points for discretizing each streamline

select_randomlyint

Randomly select a specific number of streamlines. If None all the streamlines are used.

rngRandomState

If None then RandomState is initialized internally.

verbosebool, optional.

If True, log information. Default False.

Returns
——-
clustersobj

Contains the clusters of the last layer of QuickBundlesX after merging.

References

Garyfallidis12

Garyfallidis E. et al., QuickBundles a method for tractography simplification, Frontiers in Neuroscience, vol 6, no 175, 2012.

Garyfallidis16

Garyfallidis E. et al. QuickBundlesX: Sequential clustering of millions of streamlines in multiple levels of detail at record execution time. Proceedings of the, International Society of Magnetic Resonance in Medicine (ISMRM). Singapore, 4187, 2016.

remove_clusters_by_size

dipy.align.streamlinear.remove_clusters_by_size(clusters, min_size=0)

select_random_set_of_streamlines

dipy.align.streamlinear.select_random_set_of_streamlines(streamlines, select, rng=None)

Select a random set of streamlines

Parameters
streamlinesSteamlines

Object of 2D ndarrays of shape[-1]==3

selectint

Number of streamlines to select. If there are less streamlines than select then select=len(streamlines).

rngRandomState

Default None.

Returns
selected_streamlineslist

Notes

The same streamline will not be selected twice.

set_number_of_points

dipy.align.streamlinear.set_number_of_points()
Change the number of points of streamlines

(either by downsampling or upsampling)

Change the number of points of streamlines in order to obtain nb_points-1 segments of equal length. Points of streamlines will be modified along the curve.

Parameters
streamlinesndarray or a list or dipy.tracking.Streamlines

If ndarray, must have shape (N,3) where N is the number of points of the streamline. If list, each item must be ndarray shape (Ni,3) where Ni is the number of points of streamline i. If dipy.tracking.Streamlines, its common_shape must be 3.

nb_pointsint

integer representing number of points wanted along the curve.

Returns
new_streamlinesndarray or a list or dipy.tracking.Streamlines

Results of the downsampling or upsampling process.

Examples

>>> from dipy.tracking.streamline import set_number_of_points
>>> import numpy as np

One streamline, a semi-circle:

>>> theta = np.pi*np.linspace(0, 1, 100)
>>> x = np.cos(theta)
>>> y = np.sin(theta)
>>> z = 0 * x
>>> streamline = np.vstack((x, y, z)).T
>>> modified_streamline = set_number_of_points(streamline, 3)
>>> len(modified_streamline)
3

Multiple streamlines:

>>> streamlines = [streamline, streamline[::2]]
>>> new_streamlines = set_number_of_points(streamlines, 10)
>>> [len(s) for s in streamlines]
[100, 50]
>>> [len(s) for s in new_streamlines]
[10, 10]

slr_with_qbx

dipy.align.streamlinear.slr_with_qbx(static, moving, x0='affine', rm_small_clusters=50, maxiter=100, select_random=None, verbose=False, greater_than=50, less_than=250, qbx_thr=[40, 30, 20, 15], nb_pts=20, progressive=True, rng=None, num_threads=None)

Utility function for registering large tractograms.

For efficiency we apply the registration on cluster centroids and remove small clusters.

Parameters
staticStreamlines
movingStreamlines
x0str, optional.

rigid, similarity or affine transformation model (default affine)

rm_small_clustersint, optional

Remove clusters that have less than rm_small_clusters (default 50)

select_randomint, optional.

If not None select a random number of streamlines to apply clustering Default None.

verbosebool, optional

If True, logs information about optimization. Default: False

greater_thanint, optional

Keep streamlines that have length greater than this value (default 50)

less_thanint, optional

Keep streamlines have length less than this value (default 250)

qbx_thrvariable int

Thresholds for QuickBundlesX (default [40, 30, 20, 15])

np_ptsint, optional

Number of points for discretizing each streamline (default 20)

progressiveboolean, optional

(default True)

rngRandomState

If None creates RandomState in function.

num_threadsint

Number of threads. If None (default) then all available threads will be used. Only metrics using OpenMP will use this variable.

Notes

The order of operations is the following. First short or long streamlines are removed. Second the tractogram or a random selection of the tractogram is clustered with QuickBundles. Then SLR [Garyfallidis15] is applied.

References

Garyfallidis15

Garyfallidis et al. “Robust and efficient linear

registration of white-matter fascicles in the space of streamlines”, NeuroImage, 117, 124–140, 2015 .. [R778a6c20f622-Garyfallidis14] Garyfallidis et al., “Direct native-space fiber

bundle alignment for group comparisons”, ISMRM, 2014.

Garyfallidis17

Garyfallidis et al. Recognition of white matter

bundles using local and global streamline-based registration and clustering, Neuroimage, 2017.

time

dipy.align.streamlinear.time()

Return the current time in seconds since the Epoch. Fractions of a second may be present if the system clock provides them.

transform_streamlines

dipy.align.streamlinear.transform_streamlines(streamlines, mat, in_place=False)

Apply affine transformation to streamlines

Parameters
streamlinesStreamlines

Streamlines object

matarray, (4, 4)

transformation matrix

in_placebool

If True then change data in place. Be careful changes input streamlines.

Returns
new_streamlinesStreamlines

Sequence transformed 2D ndarrays of shape[-1]==3

unlist_streamlines

dipy.align.streamlinear.unlist_streamlines(streamlines)

Return the streamlines not as a list but as an array and an offset

Parameters
streamlines: sequence
Returns
pointsarray
offsetsarray

whole_brain_slr

dipy.align.streamlinear.whole_brain_slr(static, moving, x0='affine', rm_small_clusters=50, maxiter=100, select_random=None, verbose=False, greater_than=50, less_than=250, qbx_thr=[40, 30, 20, 15], nb_pts=20, progressive=True, rng=None, num_threads=None)

Utility function for registering large tractograms.

For efficiency we apply the registration on cluster centroids and remove small clusters.

Parameters
staticStreamlines
movingStreamlines
x0str, optional.

rigid, similarity or affine transformation model (default affine)

rm_small_clustersint, optional

Remove clusters that have less than rm_small_clusters (default 50)

select_randomint, optional.

If not None select a random number of streamlines to apply clustering Default None.

verbosebool, optional

If True, logs information about optimization. Default: False

greater_thanint, optional

Keep streamlines that have length greater than this value (default 50)

less_thanint, optional

Keep streamlines have length less than this value (default 250)

qbx_thrvariable int

Thresholds for QuickBundlesX (default [40, 30, 20, 15])

np_ptsint, optional

Number of points for discretizing each streamline (default 20)

progressiveboolean, optional

(default True)

rngRandomState

If None creates RandomState in function.

num_threadsint

Number of threads. If None (default) then all available threads will be used. Only metrics using OpenMP will use this variable.

Notes

The order of operations is the following. First short or long streamlines are removed. Second the tractogram or a random selection of the tractogram is clustered with QuickBundles. Then SLR [Garyfallidis15] is applied.

References

Garyfallidis15

Garyfallidis et al. “Robust and efficient linear

registration of white-matter fascicles in the space of streamlines”, NeuroImage, 117, 124–140, 2015 .. [R9eb8c2315518-Garyfallidis14] Garyfallidis et al., “Direct native-space fiber

bundle alignment for group comparisons”, ISMRM, 2014.

Garyfallidis17

Garyfallidis et al. Recognition of white matter

bundles using local and global streamline-based registration and clustering, Neuroimage, 2017.