Is your feature request related to a problem? Please describe.
If deep supervision is set to true the DynUNet model returns a tensor with the second dimension containing the upscaled intermediate outputs of the model, whereas the new MedNeXt implementation returns a tuple with the tensors of the intermediate outputs in their original low res.

Describe the solution you'd like
For compatibility purposes and ease of use, I would suggest using the DynUNet behavior also in MedNeXt. I think having the intermediate outputs all in the same output resolution makes them easier to use for loss function calculations. Plus, the DynUNet is already stable for some time whereas the MedNeXt is not.

Describe alternatives you've considered
Obviously, the DynUNet behavior could be changed to act identical to the MedNeXt, which might be slightly more memory efficient, but that would certainly break more existing code and it is less convenient if references need to be downscaled for loss calculations.