Change HoVerNet Model Output from Dict to Tensor

HoVerNet model has three branches and currently each of them are being return with an associated key and all together as a dictionary. This approach is nice since we know what branch we are dealing with with the Enums that are defined; however, it is creating some challenges since this kind of outputs is not common in MONAI, and requires some changes:

1. HoVerNet Loss function needs to accept dictionary of Tensors (instead of one Tensor) for prediction and target (https://github.com/Project-MONAI/MONAI/pull/5199)
2. a new PrepareBatch is need to extract all the keys for the target and return a dictionary for target (instead of a tensor)
3. SuperviseTrainer needs to be updated to accept dictionary as the model outputs
4. Post process Transforms should implement new additional keys (https://github.com/Project-MONAI/MONAI/pull/5186#discussion_r981308312)

Considering that with this simple change, there is no need for any of the above changes, I am leaning toward making this backward incompatibility and making HoVerNet outputs a tensor where each channel represents output of one branch:

• first channel: nuclei segmentation (NP branch)
• second channel: horizontal map (HV branch)
• third channel: vertical map (HV branch)
• fourth channel (and any additional channel): classification output (NC branch)

@wyli @Nic-Ma @ericspod @rijobro @JHancox What do you think?

[drbeh]

Hi @Nic-Ma, thanks for your questions.

1. yes, they are all have the same spatial shape
2. postprocessing can use three of the channels and create a new tensor for instance segmentation.

However, IMO we need more motivation to make a backward incompatible change, so I close this issue, and plan for the afterward changes.

[drbeh]

@wyli @JHancox Thanks for your feedback. I agree with @JHancox, that dictionaries are more elegant [and readable] than tuples in this case. If there is not enough motivation to make it tensor, we continue with the current implementation (dictionaries) and I will close this issue.