[Question] Is the 'CT' label for PET channel in AutoPET II conversion intentional?

[ZhanghaiFA]

Description Hi, thanks for providing this great framework!

I was reviewing the dataset conversion script for AutoPET II (nnunetv2/dataset_conversion/Dataset221_AutoPETII_2023.py) and noticed something that I would like to clarify regarding the normalization strategy.

In the script, the PET channel (index 1) is labeled as "CT" when generating the dataset json:

Python

nnunetv2/dataset_conversion/Dataset221_AutoPETII_2023.py

generate_dataset_json(out_base, {0: "CT", 1:"CT"}, ...) # Index 1 is PET (SUV), but labeled as CT
My Question Based on my understanding of nnU-Net's preprocessing logic, labeling a channel as "CT" triggers CTNormalization, which clips intensity values to the [0.5, 99.5] percentiles.

I am wondering if this behavior is intended for PET images? My concern is that high SUV values (which are crucial for tumor detection) might fall into the top 0.5% and get clipped by the CTNormalization scheme, potentially losing important signal.

Clarification Should this channel be labeled as "PET" (or another non-CT name) to use ZScoreNormalization without clipping, or is there a specific reason to use CTNormalization for this dataset?

Thanks for your help!