search

cwmok / LapIRN

Large Deformation Diffeomorphic Image Registration with Laplacian Pyramid Networks
MIT License
122 stars 22 forks source link

Upsampling dense displacement field #15

Closed Lancial closed 2 years ago

Lancial commented 2 years ago

Hi! I'm trying to use the displacement version of your method but found something that's a bit confusing.

When you take the result from a higher level and pass it to a lower level for refinement, you only trilinear unsample the dimension of a displacement field but leave its magnitude unchanged. Doesn't this affect the grid sampler behavior as the sampled displacement no longer has the same physical scale as before?

lvl1_disp, _, _, lvl1_v, lvl1_embedding = self.model_lvl1(x, y) lvl1_disp_up = self.up_tri(lvl1_disp) warpped_x = self.transform(x_down, lvl1_disp_up.permute(0, 2, 3, 4, 1), self.grid_1)

If above description is confusing, I guess my question is simply why not lvl1_disp_up = self.up_tri(lvl1_disp) * 2?

Maybe I have missed some other things in your code.. Can you help me out?

Lancial commented 2 years ago

a follow-up question, why do you scale a new displacement field predicted at each level with range_flow=0.4?

cwmok commented 2 years ago

Hi @Lancial,

Because we are predicting the normalized displacement field in LapIRN. We follow the "grid_sample" function in Pytorch to normalize the spatial dimensions.

Specifically, consider the displacement field in x-direction, a magnitude 1 in the displacement field will equal to shifting 1 pixel, while a magnitude 1 in the normalized displacement field will equal to shifting (x's dimension - 1) pixels.

This is also the reason that we need to transform the normalized displacement field into the displacement field before calculating the smoothness regularization, see code for more details.

range_flow=0.4 means the range of magnitude of the predicted normalized displacement field is in [-0.4, 0.4].

Lancial commented 2 years ago

Thank you! I understand it now. Is there any reason you picked 0.4?

cwmok commented 2 years ago

No particular reason. Empirically, 0.4 is large enough for most of the medical image registration tasks. If you are doing intra-subject registration/problem with small initial misalignments, feel free to lower the number.

cwmok commented 2 years ago

I will close the issue. If you have any further questions, feel free to open a new one.