How to compute metrics directly without resizing predictions to original size?

[Mo-Ments]

Hi team,
I hope you're doing well. I noticed that in the current setup, the model first restores the predicted segmentation/image to the original input size, and then computes the evaluation metrics based on this resized output.
I was wondering if there is a way to compute these metrics directly on the raw prediction without performing this size restoration? Any guidance or suggestions would be greatly appreciated.
Thank you so much for your time and help!

[FabianIsensee]

This is not possible and I would not recommend you do that. In any realistic scenario you expect predictions in the same geometry as the original image. If your model must downsample a lot, then it needs to pay the price (errors through interpolation). You wouldn't be able to make meaningful comparisons if each model would be allowed to measure scores in their internal resolution.
If you really want to do this, resample all data and segmentations to your target spacing BEFORE putting them into nnU-Net