Intensity normalization in nnU-Net

[arnaudberenbaum]

Hello everyone,

I have some interrogations about the Normalization process for intensity-based 3D imaging like PET imaging:
If I understand correctly, :

• when I run "nnUNet plans_and_preprocess", the library use the segmentation file to check "foreground" voxels (>1 in the seg), then calculate the mean, median etc... and write it to the Plans.json file (so this is only the characteristics of a small portion of the image).
• After that, the normalization scheme is applied to all images using data from the Plans.json

But what I dont understand is:

• I'm trying to predict some tumor segmentation (like in the AutoPET challenge) so my seg file only contains 1 = Tumor label / 0 = non tumor label.
• But 0 does not mean background and 1 foreground.
• The median / mean / max / min is therefor false and the CTnormalization won't work as intended on PET imaging ?
• So is there a possibility to define how is calculated the intensity characteristics in plans_and_preprocess?

Thanks a lot for you help !! :)