Implementation of Dual Registration Network for 3D Volume Alignment

[s-shahpouri]

Hello MONAI Community,

I am currently working on a project that involves the registration of 3D medical images and am interested in implementing a dual 3D CNN architecture that can perform this task by predicting the xyz translation parameters needed to align two volumetric images (fixed and moving). The network structure I'm considering ends with a fully connected layer outputting the xyz coordinates. This network takes two volumes as input and outputs 3 parameters (xyz translation) that would help in aligning the moving image to the fixed image.

I am wondering if there is an existing network or module within MONAI that closely matches this architecture or any guidance on adapting MONAI components to achieve this functionality.

Any advice or pointers towards relevant parts of the MONAI library or similar implementations would be greatly appreciated.

Thank you in advance for your help!

[KumoLiu]

Hi @s-shahpouri, thanks for your interest here.
We have some existing tutorials about registration work. Hope you can find useful information from it
https://github.com/Project-MONAI/tutorials/tree/main/3d_registration
https://github.com/Project-MONAI/tutorials/tree/main/2d_registration

Thanks.

[dyollb]

The GlobalNet is probably the closest to what you are looking for. It predicts an affine matrix.

Note: The affine is then converted to a dense displacement field ddf to warp the moving image to the fixed space.
This network is used here: https://github.com/Project-MONAI/tutorials/tree/main/2d_registration

You can get the affine (excluding the trivial last row) like this:

    ndim = 2 # image dimension (2 or 3)
    moving = batch_data["moving_hand"].to(device)
    fixed = batch_data["fixed_hand"].to(device)
    ddf = model(torch.cat((moving, fixed), dim=1))
    print(f"affine {model.output_block.theta[0].reshape(-1, ndim, ndim + 1)}")

[aylward]

You may also want to look at this ongoing discussion: #8236