search

marianocabezas / pytorch-code

Repository to work on pyTorch
GNU General Public License v3.0
0 stars 1 forks source link

Does voxelmorph work? #2

Open John1231983 opened 5 years ago

John1231983 commented 5 years ago

Interesting that you are the first person reimplement voxelmorph in PyTorch. I wonder what is reproduce result? Does it work?

marianocabezas commented 5 years ago

To be fair, all I did was porting the SpatialTransformLayer (which is easy enough, since most functions are the same for tensorflow and pytorch) and create a u-net. In terms of results... I used it for different purposes than what they did on their part with temporal images that don't really have large deformations between them... So, I am not sure how it compares to their implementation.

John1231983 commented 5 years ago

@marianocabezas : Thanks. Could I ask you some question about the implementation?

  1. The implementation of SpatialTransformer takes moving image and deformation field as input and compute the new location of the moving image, so, does deformation filed is same as theta in the original spatial transformation network example? So, your function will be the same as four lines below, right?
        theta = self.fc_loc(xs)
        theta = theta.view(-1, 2, 3)
        grid = F.affine_grid(theta, x.size())
        x = F.grid_sample(x, grid)

https://github.com/marianocabezas/pytorch-code/blob/b7c297b9658115185328da9c0cc14fd847d64365/layers.py#L40

  1. The theta is 2x3 for 2d image and 3x4 for 3D image. but the voxelmorph paper shows that the deformation filed is a feature maps of unet network size of 3xHxW (H and W is size of input moving images). What is the relationship between theta and the deformation field-df?

Thanks