Thanks for looking into this @Nic-Ma. I notice an issue here that I hadn't thought about previously.

During training/validation/testing, there will typically be a ground truth label attached, and we'll want to use those transforms for the inverse.

At deployment, we'll probably only be using images (no labels). And typically for labels, we'll use nearest-neighbour interpolation, whereas for images we'll use something else (e.g., linear). So if we do the inverse as you've suggested in this PR, we'll probably be using non-nearest neighbour interpolation on our labels.

I wonder if this would be solved by simply doing AsDiscrete at the end. A little messy. cc @wyli @ericspod.

/black

Hi @rijobro ,

Thanks for your review, there are another 2 minor issues about inverse logic:

1. You didn't implement the inverse for ToTensorD transform, so it will run with Tensor data, fixed in this PR.
2. The BatchInverseTransform requires the transform must be InvertibleTransform while the allow_missing_keys_mode requires Compose or MapTransform, it will cause some typehints issue, I ignored in the PR.

About the sampling question you proposed, it's a good point, let me add another arg as_discrete.

Thanks.

/black

Hi @rijobro ,

Thanks for your review, I already tested it with the workflow example:
https://github.com/Project-MONAI/tutorials/blob/master/modules/engines/unet_evaluation_dict.py

And I will write another PR to solve the interpolation mode issue soon.

Thanks.

thanks @rijobro chat with Nic offline that we plan to update to store/use the inverse parameters such as mode and align_corners in EXTRA_INFO