Patches of noise on the edge of the generated atlas images about atlas-gan HOT 6 CLOSED

I tried setting the threshold to 1e-1 and it is now working perfectly. Thank you for the insights and thanks for making such an amazing model available!

from atlas-gan.

Hi @kvttt, thanks for reaching out! Yup, setting clip_bckgnd=True (which just zeroes out the voxels outside of some intensity threshold of the linear average input in training) should resolve this issue.

For context, those atlas voxels aren't being moved there by some deformation as they're being synthesized by the template generator. Surprisingly, almost all of the ablations and baselines tested in the paper can do weird things in the empty background (I've speculated that this is because the LNCC registration loss is scale-invariant and something like LNCC + 0.1*L1 might do better), although it is definitely worse when GAN training is unstable.

Just in case setting clip_bckgnd=True brings some artifacts into the foreground, try increasing losswt_gp slightly and/or decreasing d_ch.

Good luck and let me know how it goes :)

from atlas-gan.

Hi @neel-dey, thanks for the clear explanation. I will adjust the hyperparameters you suggested and get back to you as soon as I have a result! Thanks!

from atlas-gan.

Hi @neel-day, I tried setting clip_bckgnd=True. I am seeing less noise visible from a coronal view but still some decent amount of noise from a sagittal view.

At this point do you think it would be a good idea to try replacing the LNCC loss with MSE or MAE loss and start from there?

from atlas-gan.

Interesting. If you look at these lines, you'll see that it's basically just clipping atlas_inputs (which is the linear average input) with a predefined heuristic threshold of 1e-2. I would just increase the threshold to something higher that basically only keeps the brain in the mask.

But trying LNCC + MAE should be really interesting! I wouldn't do MSE/MAE only because that typically doesn't do brain registration as well. Let me know how either option goes!

from atlas-gan.

Fantastic! I'm very happy to hear that. Feel free to reach out again if you have any other questions.

from atlas-gan.