Helper Module for Deep Learning.

Volume Tweening Network (VTN) and Affine and Dense Deformable Network (ADDNet) for Unsupervised medical Image Registration.

class pynet.models.vtnet.ADDNet(input_shape, in_channels, kernel_size=3, padding=1, flow_multiplier=1.0)[source]¶

ADDNet.

Affine and Dense Deformable Network (ADDNet): affine registration subnetwork predicts a set of affine parameters, after which a flow field can be generated for warping.

Reference: https://arxiv.org/pdf/1902.05020. Code: https://github.com/microsoft/Recursive-Cascaded-Networks.

__init__(input_shape, in_channels, kernel_size=3, padding=1, flow_multiplier=1.0)[source]¶

Init class.

Parameters

input_shape: uplet

the tensor data shape (X, Y, Z).

in_channels: int

number of channels in the input tensor.

kernel_size: int, default 3

the convolution kernels size (odd number).

padding: int, default 1

the padding size, recommended (kernel_size - 1) / 2

flow_multiplier: foat, default 1

weight the flow field by this factor.

affine_flow(affine, size, without_identity=False)[source]¶

Generates a flow field given an affine matrix.

Parameters

affine: Tensor (N, 4, 4)

an affine transform.

size tuple

the target output image size.

without_identity: bool, defaul False

set to true if the identity matrix has already been substrated to the affine matrix.

Returns

flow: Tensor

the generated affine flow field.

debug(name, tensor)[source]¶

Print debug message.

Parameters

name: str

the tensor name in the displayed message.

tensor: Tensor

a pytorch tensor.

forward(x)[source]¶

Forward method.

y = Ax + b the model learns W = A - I

Parameters

x: Tensor

concatenated moving and fixed images (batch, 2 * channels, X, Y, Z)

class pynet.models.vtnet.ADDNetRegularizer(k1=0.1, k2=0.1, eps=1e-05)[source]¶

ADDNet Combined Regularization.

In addition to the correlation coefficient as our similarity loss, the orthogonality loss and the determinant loss are used as regularization losses for the affine network.

k1 * DetLoss + k2 * OrthoLoss

DetLoss: determinant should be close to 1, ie. reflection is not allowed. OrthoLoss: should be close to being orthogonal, ie. penalize the network for producing overly non-rigid transform. Let C=A’A, a positive semi-definite matrix should be close to I. For this, we require C has eigen values close to 1 by minimizing k1 + 1/k1 + k2 + 1/k2 + k3 + 1/k3. To prevent NaN, minimize k1 + eps + (1+eps)^2 / (k1+eps) + …

__init__(k1=0.1, k2=0.1, eps=1e-05)[source]¶

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

debug(name, tensor)[source]¶

Print debug message.

Parameters

name: str

the tensor name in the displayed message.

tensor: Tensor

a pytorch tensor.

class pynet.models.vtnet.VTNet(input_shape, in_channels, kernel_size=3, padding=1, flow_multiplier=1.0, nb_channels=16)[source]¶

VTNet.

Volume Tweening Network(VTN) consists of several cascaded registration subnetworks, after each of which the moving image is warped. The unsupervised training of network parameters is guided by the dissimilarity between the fixed image and each of the warped images, with the regularization losses on the flows predicted by the networks.

It follows an encoder-decoder architecture.

Reference: https://arxiv.org/pdf/1902.05020. Code: https://github.com/microsoft/Recursive-Cascaded-Networks.

__init__(input_shape, in_channels, kernel_size=3, padding=1, flow_multiplier=1.0, nb_channels=16)[source]¶

Init class.

Parameters

input_shape: uplet

the tensor data shape (X, Y, Z).

in_channels: int

number of channels in the input tensor.

kernel_size: int, default 3

the convolution kernels size (odd number).

padding: int, default 1

the padding size, recommended (kernel_size - 1) / 2

flow_multiplier: foat, default 1

weight the flow field by this factor.

nb_channels: int, default 16

the number of channels after the first convolution.

debug(name, tensor)[source]¶

Print debug message.

Parameters

name: str

the tensor name in the displayed message.

tensor: Tensor

a pytorch tensor.

forward(x)[source]¶

Forward method.

Parameters

x: Tensor

concatenated moving and fixed images (batch, 2 * channels, X, Y, Z)

pynet.models.vtnet.normal_transform_pixel(shape)[source]¶

Compute the normalization matrix from image size in pixels to [-1, 1].

pynet.models.vtnet.normalize_homography(affine, shape_src, shape_dst)[source]¶

Normalize a given homography in pixels to [-1, 1].

Reference: https://discuss.pytorch.org/t/

affine-transformation-matrix-paramters-conversion/19522/13

Parameters

affine: torch.Tensor (N, 4, 4)

homography/ies from source to destiantion to be normalized.

shape_src: tuple (3,)

size of the source image.

shape_dst: tuple (3,)

size of the destination image.

Returns

affine: torch.Tensor (N, 4, 4)

the normalized homography/ies.

pynet.models.vtnet.to_homography(batch_affine)[source]¶

Convert batch of affine matrices of size (N, 3, 4) to (N, 4, 4).