Binary Operators

Affine Intensity Filter

class AffineIntensity(similarity, dim=2, init_affine=None, init_translation=None, device='cpu', dtype=torch.float32)

static Create(similarity, dim=2, init_affine=None, init_translation=None, device='cpu', dtype=torch.float32)

Object for registering two structured grids with an affine transformation. The affine and translation are optimized independently. This Affine Intensity filter must be on the same device as the target and moving structured grids. The affine and translation attributes have requires_grad=True so they can be added to a torch optimizer and updated with autograd functions.

Parameters

• similarity (Filter) – This is the similiarty filter used to compare the two structured grids to be registered. This filter is usually a Binary Operator (ie. L2 Image Similarity)

• dim (int) – Dimensionality of the structured grids to be registered (not including channels).

• init_affine (tensor, optional) – Initial affine to apply to the moving structured grid.

• init_translation (tensor, optional) – Initial translation to apply to the moving structured grid.

• device (str) – Memory location - one of ‘cpu’, ‘cuda’, or ‘cuda:X’ where X specifies the device identifier. Default: ‘cpu’

• dtype (str) – Data type for the attributes. Specified from torch memory types. Default: ‘torch.float32’

Returns

Affine Intensity Filter Object

forward(target, moving)

Apply the forward affine operation applied to the moving image and calculate the resulting similarity measure between the target and moving images. The gradients on the affine and translation attributes are tracked through this forward operation so that the gradient update can be applied to update the affine and translation. This function is meant to be used iteratively in the registration process.

Parameters

• target (StructuredGrid) – Target structured grid. Does not get updated or changed.

• moving (StructuredGrid) – Moving structured grid. Affine and translation are applied this structured grid before the similarity calculation.

Returns

Energy from the similarity evaluation (usually a single float).

Compose Grids Filter

class ComposeGrids(padding_mode='border', device='cpu', dtype=torch.float32)

static Create(padding_mode='border', device='cpu', dtype=torch.float32)

Object to compose StructuredGrid look-up table fields into one grid.

Parameters

• pad_mode (str) – padding mode for outside grid values - one of ‘zeros’, ‘border’, or ‘reflection’. Default: ‘zeros’

• device (str) – Memory location - one of ‘cpu’, ‘cuda’, or ‘cuda:X’ where X specifies the device identifier. Default: ‘cpu’

• dtype (str) – Data type for the attributes. Specified from torch memory types. Default: ‘torch.float32’

Returns

Object to compose a list of look-up tables.

forward(L)

Given a list of StructuredGrid look-up tables L = [L0, L1, L2] returns a composed look-up table comp_field = L2(L1(L0(x))) of type StructuredGrid.

Parameters

L (list, tuple) – List of look-up tables. All fields in the list must be on the same memory device.

Returns

Composed look-up tables Comp_filed

L2 Image Similarity

class L2Similarity(dim=2, device='cpu', dtype=torch.float32)

static Create(dim=2, device='cpu', dtype=torch.float32)

Compare two StructuredGrid objects using an L2 similarity metric.

Parameters

• dim (int) – Dimensionality of the StructuredGrid to be compared (not including channels).

• device (str) – Memory location - one of ‘cpu’, ‘cuda’, or ‘cuda:X’ where X specifies the device identifier. Default: ‘cpu’

• dtype (str) – Data type for the attributes. Specified from torch memory types. Default: ‘torch.float32’

Returns

L2 comparision object.

c1(target, moving, grads)

First derivative of the L2 similarity metric.

Parameters

• target (StructuredGrid) – Structured Grid 1

• moving (StructuredGrid) – Structured Grid 2

• grads (StructuredGrid) – Gradients of the moving image.

Returns

forward(target, moving)

Compare two StructuredGrid with L2 similarity metric. This is often used for registration so the variables are labeled as target and moving. This function preserves the dimensionality of the original grids.

Parameters

• target (StructuredGrid) – Structured Grid 1

• moving (StructuredGrid) – Structured Grid 2

Returns

L2 similarity as StructuredGrid

Normalized Cross Correlation Filter

class NormalizedCrossCorrelation(grid, window=5, device='cpu', dtype=torch.float32)

static Create(grid, window=5, device='cpu', dtype=torch.float32)

c1(target, moving, grads)

forward(target, moving)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.