Soft Dice Metric request

[pedrob37]

Is your feature request related to a problem? Please describe.
For segmentation tasks where the ground truth labels are continuous rather than binary, the DiceMetric fails.

Describe the solution you'd like
Simply put, a version of DiceMetric that allows for continuous labels to be input as y, instead of just binaries.

Describe alternatives you've considered
This would in essence function as MONAI's dice loss, so being able to use losses as metrics would be a suitable workaround.

[wyli]

Hi @yiheng-wang-nv do you have time to help review all the metrics? ideally I think we can remove raise ValueError("y should be a binarized tensor.") if it's not a hard requirement by the underlying algo.?

[yiheng-wang-nv]

Hi @wyli ,

For DiceMetric, the calculation in here requires the input are binarized, otherwise the logic is wrong.

For ConfusionMatrixMetric, the requirement is also necessary, please check the logic here.

[wyli]

Thanks, agree that the confusion matrix would require binary inputs at the moment. Perhaps for the Dice we can change it to a warning instead?