Augmentation Transform | InCut & OutCut Dropping (ETA 08/27/21)

[finalelement]

Dropping off randomly sized patches from inside the image and out the image where apart from them everything else gets dropped out.

[Nic-Ma]

Hi @finalelement ,

Thanks for the feature request, is it similar to this existing transform in MONAI: RandCoarseDropout ?
https://github.com/Project-MONAI/MONAI/blob/dev/monai/transforms/intensity/array.py#L1566

Thanks.

[finalelement]

Hi @Nic-Ma At the first sight, it does seem similar. This is quite similar to drop 'in' the image. However this does not support dropping off the rest of the image and just leaving the hole. This would be termed as out Out Cut.

Thanks for the pointer, Ill use this as reference and will also do some empirical testing on a couple of images.

[finalelement]

Hey @Nic-Ma I read and compared the code for In-cutout with the transform that you suggested me to look into RandCoarseDropout. There is only one difference that I see between the one available with MONAI and the one that I've been using. It is essentially allowing for the flexibility of filling the 'hole' with random values between 0 and 1 rather than a specified value.

Doing this would be more useful. Perhaps we could give the user an option to either specify the fill value, if it is not specified we could fill it in with torch.rand. Specifically by introducing an if condition at ~Line 1632 in this file

[Nic-Ma]

Hi @finalelement ,

I think you can add a mode arg in the RandCoarseDropout transform, mode can be "constant" or "random", similar to the numpy pad modes.

Thanks.

[Nic-Ma]

CC @wyli @ericspod

[Nic-Ma]

According to the online discussion, I will try to enhance the existing RandCoarseDropout transform to support random filling ASAP.

Thanks.