CNN with Complex Values

I’m dealing with MR Images.
So, we have 3D or 4D Volumes, which can be in Image Domain or Frequency Domain.
I’m designing a network that can work in Frequency Domain, which means it’s in Fourier space.
For being in Fourier space, all the values are complex numbers.
Can’t use just magnitude [abs(x)] because they can’t reconstruct the image back.
Using two channel tensor isn’t a solution I’m looking for.
Can you please add support for complex values in pytorch?

It’s what they are working on I think. For example,