cornucopia.geometric

ElasticTransform

ElasticTransform(dmax=0.1, unit='fov', shape=5, bound='border', steps=0, order=3, zero_center=False, nearest_if_label=True, *, dtype=None, device=None, shared=True, **kwargs)

Bases: NonFinalTransform

Elastic transform encoded by cubic splines.

The number of control points is fixed but coefficients are randomly sampled from a uniform distribution.

Parameters:

Name	Type	Description	Default
`dmax`	`[list of] float`	Max displacement per dimension	`0.1`
`unit`	`(fov, vox)`	Unit of `dmax`.	`'fov'`
`shape`	`[list of] int`	Number of spline control points	`5`
`bound`	`(zeros, border, reflection)`	Padding mode used for the deformed image.	`'zeros'`
`steps`	`int`	Number of scaling-and-squaring integration steps	`0`
`order`	`1..7`	Order of the splines that encode the smooth deformation.	`3`
`zero_center`	`bool`	Subtract its mean displacement to the flow field so that it has an empirical mean of zero.	`False`
`nearest_if_label`	`bool`	By default, if a tensor has an integer data type, it is deformed using label-specific resampling (each unique label is extracted and resampled using linear interpolation, and an argmax output label map is computed on the fly). If `nearest_if_label=True`, the entire label map will be resampled at once using nearest-neighbour interpolation.	`True`

Other Parameters:

Name	Type	Description
`returns`	`[list or dict of] {'input', 'output', 'flow', 'controls'}`	'input': The input image 'output': The deformed image 'flow': The displacement field 'controls': The control points of the displacement field
`shared`	`(channels, tensors, channels + tensors, '')`	Apply same transform to all images/channels

Final

Final(flow=None, controls=None, steps=0, order=3, bound='border', zero_center=False, nearest_if_label=True, dtype=None, device=None, **kwargs)

Bases: FinalTransform

Final (deterministic) elastic transform

Parameters:

Name	Type	Description	Default
`flow`	`(C, D, *spatial) tensor`	Flow field (if not provided, `control` must be provided)	`None`
`control`	`(C, D, *shape) tensor`	Spline control points (if not provided, `flow` must be provided)	required
`steps`	`int`	Number of scaling and squaring steps	`0`
`order`	`1..7`	Order of the splines that encode the smooth deformation.	`3`
`bound`	`(zeros, border, reflection)`	Padding mode used for the deformed image.	`'zeros'`
`zero_center`	`bool`	Subtract its mean displacement to the flow field so that it has an empirical mean of zero.	`False`
`nearest_if_label`	`bool`		`True`

make_flow

make_flow(control, fullshape)

Upsample the control points to the final full size

Parameters:

Name	Type	Description	Default
`control`	`(C, D, *shape) tensor`	Spline control points	required
`fullshape`	`list[int]`	Target shape	required

Returns:

Name	Type	Description
`flow`	`(C, D, *fullshape) tensor`	Upampled flow field

xform

xform(x)

Deform the input tensor

Parameters:

Name	Type	Description	Default
`x`	`(C, *spatial) tensor`	Input tensor	required

Returns:

Name	Type	Description
`out`	`[dict or list of] tensor`	The tensors returned by this function depend on the value of `self.returns`. See `ElasticTransform`.

make_final

make_final(x, max_depth=float('inf'), flow=True)

Generate a deterministic transform with constant parameters

Parameters:

Name	Type	Description	Default
`x`	`(C, *spatial) tensor`	Tensor to deform	required
`max_depth`	`int`	Maximum number of transforms to unroll	`float('inf')`
`flow`	`bool`	Precompute the upsampled flow field	`True`

Returns:

Name	Type	Description
`xform`	`Final`	Final transform with parameters `flow : (C, D, spatial) tensor`, the upsampled flow field `control : (C, D, shape) tensor`, the spline control points

RandomElasticTransform

RandomElasticTransform(dmax=0.1, shape=5, unit='fov', bound='border', steps=0, order=3, zero_center=False, nearest_if_label=False, *, dtype=None, device=None, shared=True, shared_flow=None, **kwargs)

Bases: NonFinalTransform

Elastic Transform with random parameters.

Parameters:

Name	Type	Description	Default
`dmax`	`Sampler or float`	Sampler or Upper bound for maximum displacement	`0.1`
`shape`	`Sampler or int`	Sampler or Upper bound for number of control points	`5`
`unit`	`(fov, vox)`	Unit of `dmax`	`'fov'`
`bound`	`(zeros, border, reflection)`	Padding mode	`'zeros'`
`steps`	`int`	Number of scaling-and-squaring integration steps	`0`
`order`	`int`	Spline order	`3`
`zero_center`	`bool`	Subtract its mean displacement to the flow field so that it has an empirical mean of zero.	`False`
`nearest_if_label`	`bool`	By default, if a tensor has an integer data type, it is deformed using label-specific resampling (each unique label is extracted and resampled using linear interpolation, and an argmax output label map is computed on the fly). If `nearest_if_label=True`, the entire label map will be resampled at once using nearest-neighbour interpolation.	`False`

Other Parameters:

Name	Type	Description
`returns`	`[list or dict of] {'input', 'output', 'flow', 'controls'}`	'input': The input image 'output': The deformed image 'flow': The displacement field 'controls': The control points of the displacement field
`shared`	`(channels, tensors, channels + tensors, '')`	Whether to share random parameters across tensors and/or channels
`shared_flow`	`(channels, tensors, channels + tensors, '', None)`	Whether to share random field across tensors and/or channels. By default: same as `shared`

AffineTransform

AffineTransform(translations=0, rotations=0, shears=0, zooms=0, unit='fov', bound='border', nearest_if_label=False, *, dtype=None, device=None, shared=True, **kwargs)

Bases: NonFinalTransform

Apply an affine transform encoded by translations, rotations, shears and zooms.

The affine matrix is defined as: A = T @ Rx @ Ry @ Rz @ Sx @ Sy Sz @ Z with the center of the field of view used as center of rotation. (A is a matrix so the transforms are applied right to left)

Parameters:

Name	Type	Description	Default
`translations`	`[list of] float`	Translation (per X/Y/Z)	`0`
`rotations`	`[list of] float`	Rotations (about Z/Y/X), in deg	`0`
`shears`	`[list of] float`	Translation (about Z/Y/Z)	`0`
`zooms`	`[list of] float`	Zoom about 1 (per X/Y/Z)	`0`
`unit`	`(fov, vox)`	Unit of `translations`.	`'fov'`
`bound`	`(zeros, border, reflection)`	Padding mode	`'zeros'`
`nearest_if_label`	`bool`	By default, if a tensor has an integer data type, it is deformed using label-specific resampling (each unique label is extracted and resampled using linear interpolation, and an argmax output label map is computed on the fly). If `nearest_if_label=True`, the entire label map will be resampled at once using nearest-neighbour interpolation.	`False`

Other Parameters:

Name	Type	Description
`returns`	`[list or dict of] {'input', 'output', 'flow', 'matrix'}`	'input': The input image 'output': The deformed image 'flow': The displacement field 'matrix': The affine matrix
`shared`	`(channels, tensors, channels + tensors, '')`	Apply same transform to all images/channels

Final

Final(flow=None, matrix=None, bound='border', nearest_if_label=False, dtype=None, device=None, **kwargs)

Bases: FinalTransform

Apply an affine transform encoded by an affine matrix

Parameters:

Name	Type	Description	Default
`flow`	`([C], D, *spatial) tensor`	Flow field	`None`
`matrix`	`([C], D+1, D+1) tensor`	Matrix	`None`
`bound`	`(zeros, border, reflection)`	Padding mode	`'zeros'`
`nearest_if_label`	`bool`		`False`

RandomAffineTransform

RandomAffineTransform(translations=0.1, rotations=15, shears=0.012, zooms=0.15, iso=False, unit='fov', bound='border', nearest_if_label=False, *, dtype=None, device=None, shared=True, shared_matrix=None, **kwargs)

Bases: NonFinalTransform

Affine Transform with random parameters.

Parameters:

Name	Type	Description	Default
`translations`	`Sampler or [list of] float`	Sampler or Upper bound for translation (per X/Y/Z)	`0.1`
`rotations`	`Sampler or [list of] float`	Sampler or Upper bound for rotations (about Z/Y/X), in deg	`15`
`shears`	`Sampler or [list of] float`	Sampler or Upper bound for shears (about Z/Y/Z)	`0.012`
`zooms`	`Sampler or [list of] float`	Sampler or Upper bound for zooms about 1 (per X/Y/Z)	`0.15`
`iso`	`bool`	Use isotropic zoom	`False`
`unit`	`(fov, vox)`	Unit of `translations`.	`'fov'`
`bound`	`(zeros, border, reflection)`	Padding mode	`'zeros'`
`nearest_if_label`	`bool`	By default, if a tensor has an integer data type, it is deformed using label-specific resampling (each unique label is extracted and resampled using linear interpolation, and an argmax output label map is computed on the fly). If `nearest_if_label=True`, the entire label map will be resampled at once using nearest-neighbour interpolation.	`False`

Other Parameters:

Name	Type	Description
`returns`	`[list or dict of] {'input', 'output', 'flow', 'matrix'}`	'input': The input image 'output': The deformed image 'flow': The displacement field 'matrix': The affine matrix
`shared`	`(channels, tensors, channels + tensors, '')`	Whether to share random parameters across tensors and/or channels
`shared_matrix`	`(channels, tensors, channels + tensors, '', None)`	Whether to share matrices across tensors and/or channels. By default: same as `shared`

AffineElasticTransform

AffineElasticTransform(dmax=0.1, shape=5, steps=0, translations=0, rotations=0, shears=0, zooms=0, unit='fov', bound='border', patch=None, order=3, zero_center=False, nearest_if_label=False, *, dtype=None, device=None, shared=True, **kwargs)

Bases: NonFinalTransform

Affine + Elastic [+ Patch] transform.

Parameters:

Name	Type	Description	Default
`dmax`	`[list of] float`	Max displacement per dimension	`0.1`
`shape`	`[list of] int`	Number of spline control points	`5`
`steps`	`int`	Number of scaling-and-squaring integration steps	`0`
`translations`	`[list of] float`	Translation (per X/Y/Z)	`0`
`rotations`	`[list of] float`	Rotations (about Z/Y/X), in deg	`0`
`shears`	`[list of] float`	Translation (about Z/Y/Z)	`0`
`zooms`	`[list of] float`	Zoom about 1 (per X/Y/Z)	`0`
`unit`	`(fov, vox)`	Unit of `translations` and `dmax`.	`'fov'`
`bound`	`(zeros, border, reflection)`	Padding mode	`'zeros'`
`patch`	`[list of] int`	Size of random patch to extract	`None`
`order`	`int`	Spline order	`3`
`zero_center`	`bool`	Subtract its mean displacement to the elastic flow field so that it has an empirical mean of zero. This has no effect on the affine component.	`False`
`nearest_if_label`	`bool`	By default, if a tensor has an integer data type, it is deformed using label-specific resampling (each unique label is extracted and resampled using linear interpolation, and an argmax output label map is computed on the fly). If `nearest_if_label=True`, the entire label map will be resampled at once using nearest-neighbour interpolation.	`False`

Other Parameters:

Name	Type	Description
`returns`	`[list or dict of] {'input', 'output', 'flow', 'controls', 'matrix'}`	'input': The input image 'output': The deformed image 'flow': The displacement field 'control': The control points of the nonlinear field 'matrix': The affine matrix
`shared`	`(channels, tensors, channels + tensors, '')`	Apply same transform to all images/channels

Final

Final(flow, controls, affine, bound='border', nearest_if_label=False, backend=None, **kwargs)

Bases: FinalTransform

Determinstic affine+elastic transform

Parameters:

Name	Type	Default
`flow`	`(C, D, *spatial) tensor`	required
`controls`	`(C, D, *spatial) tensor`	required
`affine`	`([C], D+1, D+1) tensor`	required
`bound`	`str`	`'border'`

make_final

make_final(x, max_depth=float('inf'))

Parameters:

Name	Type	Description	Default
`x`	`(C, *shape) tensor`		required
`max_depth`	`int`		`float('inf')`

Returns:

Name	Type	Description
`warp`	`(C, D, *shape) tensor`
`controls`	`(C, D, *self.shape) tensor`
`affine`	`(D+1, D+1) tensor`

RandomAffineElasticTransform

RandomAffineElasticTransform(dmax=0.1, shape=5, steps=0, translations=0, rotations=0, shears=0, zooms=0, iso=False, unit='fov', bound='border', patch=None, order=3, zero_center=False, nearest_if_label=False, *, dtype=None, device=None, shared=True, shared_flow=None, **kwargs)

Bases: NonFinalTransform

Random Affine + Elastic transform.

Parameters:

Name	Type	Description	Default
`dmax`	`Sampler or float`	Sampler or Upper bound for maximum displacement	`0.1`
`shape`	`Sampler or int`	Sampler or Upper bounds for number of control points	`5`
`steps`	`int`	Number of scaling-and-squaring integration steps	`0`
`translations`	`Sampler or [list of] float`	Sampler or Upper bound for translation (per X/Y/Z)	`0`
`rotations`	`Sampler or [list of] float`	Sampler or Upper bound for rotations (about Z/Y/X), in deg	`0`
`shears`	`Sampler or [list of] float`	Sampler or Upper bound for shears (about Z/Y/Z)	`0`
`zooms`	`Sampler or [list of] float`	Sampler or Upper bound for zooms about 1 (per X/Y/Z)	`0`
`iso`	`bool`	Use isotropic zoom	`False`
`unit`	`(fov, vox)`	Unit of `translations`.	`'fov'`
`bound`	`(zeros, border, reflection)`	Padding mode	`'zeros'`
`patch`	`[list of] int`	Size of random patch to extract	`None`
`order`	`int`	Spline order	`3`
`zero_center`	`bool`	Subtract its mean displacement to the elastic flow field so that it has an empirical mean of zero. This has no effect on the affine component.	`False`
`nearest_if_label`	`bool`	By default, if a tensor has an integer data type, it is deformed using label-specific resampling (each unique label is extracted and resampled using linear interpolation, and an argmax output label map is computed on the fly). If `nearest_if_label=True`, the entire label map will be resampled at once using nearest-neighbour interpolation.	`False`

Other Parameters:

Name	Type	Description
`shared`	`(channels, tensors, channels + tensors, '')`	Apply same hyperparameters to all images/channels
`shared_flow`	`(channels, tensors, channels + tensors, '', None)`	Apply the same random flow to all images/channels. Default: same as shared

MakeAffinePair

MakeAffinePair(transform=None, *, returns=('left', 'right'), **kwargs)

Bases: NonFinalTransform

Generate a pair made of the same image transformed in two different ways.

This Transform returns a tuple: (transformed_input, true_transform), where transformed_input has the same layout as the input and true_transform is a dictionary with keys 'flow' and 'affine'.

Parameters:

Name	Type	Description	Default
`transform`	`RandomAffineTransform`	An instantiated transform.	`RandomAffineTransform()`

Other Parameters:

Name	Type	Description
`returns`	`[list or dict of] {'left', 'right', 'flow', 'matrix'}`	'input': Input image 'left': First transformed image 'right': Second transformed image 'flow': Displacement field that warps right to left 'matrix': Affine matrix that warps right to left (i.e., maps left coordinates to right coordinates)

Final

Final(left, right, **kwargs)

Bases: FinalTransform

Deterministic affine pair transform

SlicewiseAffineTransform

SlicewiseAffineTransform(translations=0, rotations=0, shears=0, zooms=0, bulk_translations=0, bulk_rotations=0, bulk_shears=0, bulk_zooms=0, slice=-1, unit='fov', bound='border', spacing=1, subsample=1, *, shared=True, **kwargs)

Bases: NonFinalTransform

Each slice samples the 3D volume using a different transform

Parameters:

Name	Type	Description	Default
`translations`	`list of [list of] float`	Translation per slice (per X/Y/Z)	`0`
`rotations`	`list of [list of] float`	Rotations per slice (about Z/Y/X), in deg	`0`
`shears`	`list of [list of] float`	Shears per slice (about Z/Y/Z)	`0`
`zooms`	`list of [list of] float`	Zoom about 1 per slice (per X/Y/Z)	`0`
`bulk_translations`	`[list of] float`	Bulk translation (per X/Y/Z)	`0`
`bulk_rotations`	`[list of] float`	Bulk rotation (about Z/Y/X), in deg	`0`
`bulk_shears`	`l[list of] float`	Bulk shear (about Z/Y/Z)	`0`
`bulk_zooms`	`[list of] float`	Bulk zoom about 1 (per X/Y/Z)	`0`
`slice`	`(0, 1, 2)`	Slice direction	`0`
`unit`	`(fov, vox)`	Unit of `translations`.	`'fov'`
`bound`	`(zeros, border, reflection)`	Padding mode	`'zeros'`
`spacing`	`int`	Spacing between thick slices (as a number of high-res voxels)	`1`
`subsample`	`int`	Additional isotropic subsampling	`1`

Other Parameters:

Name	Type	Description
`returns`	`[list or dict of] {'input', 'output', 'flow', 'matrix'}`	'input': First transformed image 'output': Second transformed image 'flow': Displacement field 'matrix': Stacked affine matrices (one per slice)
`shared`	`(channels, tensors, channels + tensors, '')`	Apply same transform to all images/channels

Final

Final(flow, matrix, slice=-1, spacing=1, subsample=1, bound='border', **kwargs)

Bases: FinalTransform

Precomputed slicewise transform

RandomSlicewiseAffineTransform

RandomSlicewiseAffineTransform(translations=0.1, rotations=15, shears=0, zooms=0, bulk_translations=0.05, bulk_rotations=15, bulk_shears=0, bulk_zooms=0, iso=False, slice=-1, spacing=1, subsample=1, shots=2, nodes=8, unit='fov', bound='border', *, shared=True, shared_matrix=None, **kwargs)

Bases: NonFinalTransform

Slicewise3DAffineTransform with random parameters.

Parameters:

Name	Type	Description	Default
`translations`	`Sampler or [list of] float or (*, 1\|D) tensor`	Sampler or Upper bound for translation (per X/Y/Z)	`0.1`
`rotations`	`Sampler or [list of] float or (*, 1\|D(D-1)/2) tensor`	Sampler or Upper bound for rotations (about Z/Y/X), in deg	`15`
`shears`	`Sampler or [list of] float or (*, 1\|D(D-1)/2) tensor`	Sampler or Upper bound for shears (about Z/Y/Z)	`0`
`zooms`	`Sampler or [list of] float or (*, 1\|D) tensor`	Sampler or Upper bound for zooms about 1 (per X/Y/Z)	`0`
`bulk_translations`	`[list of] float`	Sampler or Upper bound for Bulk translation (per X/Y/Z)	`0.05`
`bulk_rotations`	`[list of] float`	Sampler or Upper bound for Bulk rotation (about Z/Y/X), in deg	`15`
`bulk_shears`	`l[list of] float`	Sampler or Upper bound for Bulk shear (about Z/Y/Z)	`0`
`bulk_zooms`	`[list of] float`	Sampler or Upper bound for Bulk zoom about 1 (per X/Y/Z)	`0`
`iso`	`bool`	Use isotropic zoom	`False`
`slice`	`Sampler or int`	Slice direction. If None, a random slice direction is selected.	`-1`
`spacing`	`Sampler or int`	Spacing between thick slices (as a number of high-res voxels)	`1`
`subsample`	`Sampler or int`	Additional isotropic subsampling	`1`
`shots`	`Sampler or int`	Number of interleaved sweeps. Typically, two interleaved sequences of slices are acquired to avoid slice cross talk.	`2`
`nodes`	`Sampler or int`	Sampler or Upper bound for the number of nodes in the motion trajectory (encoded by cubic splines). If None, independent motions are sampled for each slice.	`8`
`unit`	`(fov, vox)`	Unit of `translations`.	`'fov'`
`bound`	`(zeros, border, reflection)`	Padding mode	`'zeros'`

Other Parameters:

Name	Type	Description
`returns`	`[list or dict of] {'input', 'output', 'flow', 'matrix'}`	'input': First transformed image 'output': Second transformed image 'flow': Displacement field 'matrix': Stacked affine matrices (one per slice)
`shared`	`(channels, tensors, channels + tensors, '')`	Apply same parameters to all images/channels
`shared_matrix`	`(channels, tensors, channels + tensors, '')`	Apply same affine matrix to all images/channels. Default: same as `shared`.