Augmentation Transform | Patch Shuffling (ETA 8/27)

[finalelement]

Local pixel/voxel shuffling is a an augmentation that can be used during training, its essentially a way to add noise to the image. In brief the a small sub-block of pixels/voxels are selected from the entire image and then local pixels are shuffled to generate a noisy image

The expectation is to send an image and the transform should be able to shuffle pixels given the hyper-parameters of how many blocks should be shuffled

It should also be noted that pytorch also offers a Pixel Shuffle (https://pytorch.org/docs/stable/generated/torch.nn.PixelShuffle.html). That should not be confused with this one, as it shuffles the dimensions of the image and not sub-blocks/local block etc.

[Nic-Ma]

Hi @finalelement ,

Thanks for the feature request.
I am not very clear about the shuffle progress: do you mean to shuffle the pixels within selected block or shuffle the blocks in the entire image?

Thanks.

[finalelement]

Hi @Nic-Ma I mean to say the voxels/pixels are shuffled within a local patch. Please also share if you already see something similar already existing.

[Nic-Ma]

@finalelement , OK, we don't have such transform so far I think.
Please go ahead to implement your ideas.

Thanks in advance.

[Spenhouet]

There might also be a method similar to RandomSwap coming.
To differentiate I would propose the following naming:

• For the RandomSwap equivalent: RandPatchSwap
• For the patch shuffling proposed here: RandPatchShuffle

[Nic-Ma]

Hi @Spenhouet ,

Thanks for your explanation, naming sounds to me.
@finalelement Do you have any other comments?

Thanks.

[finalelement]

I've created a pull request. The code has been added in this pull request.

This transform needs testcases and some more local testing, as well as how it has to be added to the dictionary.py

Updating the ETA to 08/20

[finalelement]

Follow this in future for making this N-d

Also shifting the transform from spatial to intensity, due to it being a better fit as an intensity transform.

[finalelement]

@wyli Thank you for all the support with this Issue. This current one supports 3D.

We will re-iterate on a new on a new PR with updated dev-branch for dictionary version and N-dimensional version of this transform.