torchreg

torchreg is a tiny (~300 lines) PyTorch-based library for 2D and 3D image registration.

Usage

Affine Registration of two image tensors is done via:

from torchreg import AffineRegistration

# Load images as torch Tensors
big_alice = ...    # Tensor with shape [1, 3 (color channel), 1024 (pixel), 1024 (pixel)]
small_alice = ...  # Tensor with shape [1, 3 (color channel), 1024 (pixel), 1024 (pixel)]
# Intialize AffineRegistration
reg = AffineRegistration(is_3d=False)
# Run it!
moved_alice = reg(big_alice, small_alice)

Features

Multiresolution approach to save compute (per default 1/4 + 1/2 of original resolution for 500 + 100 iterations)

reg = AffineRegistration(scales=(4, 2), iterations=(500, 100))

Choosing which operations (translation, rotation, zoom, shear) to optimize

reg = AffineRegistration(with_zoom=False, with_shear=False)

Custom initial parameters

reg = AffineRegistration(zoom=torch.Tensor([[1.5, 2.]]))

Custom dissimilarity functions and optimizers

def dice_loss(x1, x2):
    dim = [2, 3, 4] if len(x2.shape) == 5 else [2, 3]
    inter = torch.sum(x1 * x2, dim=dim)
    union = torch.sum(x1 + x2, dim=dim)
    return 1 - (2. * inter / union).mean()

reg = AffineRegistration(dissimilarity_function=dice_loss, optimizer=torch.optim.Adam)

CUDA support (NVIDIA GPU)

moved_alice = reg(moving=big_alice.cuda(), static=small_alice.cuda())

After the registration is run, you can apply it to new images (coregistration)

another_moved_alice = reg.transform(another_alice, shape=(256, 256))

with desired output shape.

You can access the affine

affine = reg.get_affine()

and the four parameters (translation, rotation, zoom, shear)

translation = reg.parameters[0]
rotation = reg.parameters[1]
zoom = reg.parameters[2]
shear = reg.parameters[3]

Installation

pip install torchreg

Examples/Tutorials

There are three example notebooks:

• examples/basics.ipynb shows the basics by using small cubes/squares as image data
• examples/images.ipynb shows how to register alice_big.jpg to alice_small.jpg
• examples/mri.ipynb shows how to register MR images (Nifti files) including co-, parallel and multimodal registration

Background

If you want to know how the core of this package works, read the blog post!

TODO

• Add 2D support to SyN, NCC and LinearElasticity
• Add tests for SyN