RandomGhosting algorithm removes zero-frequency component

🐛Bug

When the ghosting algorithm removes the center of k-space, the image is too distorted.

To reproduce

I’m using this image and the TorchIO Slicer module.

8 ghosts along axis 2:

7 ghosts:

The shape of the image along axis 2 is 176, which is a multiple of 8.

Expected behavior

An image that looks more like the second.

TorchIO version

0.17.0.

[fepegar]

Note that since #204 you don’t need to pass the subject through the dataset:

suj = tio.data.Subject(t1=tio.Image('/data/t1.nii.gz'))
t = tio.RandomGhosting(axes=2, num_ghosts=(8, 8), intensity=(1, 1))
s = t(sun)
s.t1.save('/tmp/trans.nii')

[fepegar]

I prefer the MRI artifacts at the beginning as well, but users might, for whatever reason, have some negative values in the input.

Here’s a notebook that I’m playing with. You can see that there’s not much difference: