Any tips for choosing good intensity transform settings for a specific task?

[demcbs]

Hello all,

For my segmentation task I have found that using a good amount of nonspatial data augmentation (not exactly the same as nnUNet's default, but similar) lowers my validation performance, compared to not using nonspatial data augmentation. However, I have also found that including just a little bit of nonspatial data augmentation seems not to worsen, and possibly even improves, validation performance.

This brings me to ask if anyone has any tips on how to set an appropriate amount of data augmentation for a specific task (apart from using domain knowledge, which surely applies, but is of limited use to me in this case)?

Two subquestions:

• Am I right in assuming that, as long as validation performance does not suffer, more nonspatial DA should be considered better? I thought it would likely tend to improve the robustness of the trained model, hence the assumption.
• Is there some sort of hierarchy in the transforms of nnUNet, which could be considered more generally useful and which less so (as a starting point for me to try things out)?

(for Fabian, in case he answers):

• It will happen occasionally that multiple nonspatial transforms are applied to the same image or channel. It seems hard to me to control the appropriateness of the parameters of the transforms for such cases (e.g. maybe the combined transform is too strong to be helpful). Has this at any point been a consideration/problem, or, in fact, shown not to be a problem in choosing the data augmentation settings?
• In your experience, what can reasonably be expected, in terms of performance improvement, from choosing appropriate nonspatial DA?
• Is there some principle which decides whether a nonspatial transform is applied channelwise versus on whole images?