RuntimeWarning: Mean of empty slice means[r][m] = np.nanmean([i['metrics'][r][m] for i in results]) ； Mean Validation Dice: nan

[lucky653]

Hi,
I am encountering the following issue during runtime:

My training dataset is multi-class; however, not all images contain every labelled region/class. Is this partial presence the reason why the Dice values for certain classes are showing as NaN? Please note that while these corresponding classes may be sparse, they do definitely exist in the dataset.

Is it possible for the training process to handle cases where only a subset of the regions/classes are present in a given image?

Below are screenshots of the dataset.json and summary.json files.

Thank you for your reply!

[vmiller987]

I have datasets that are similar and not all images contain all labels.

What I have noticed is that the random splits that gets generated isn't ideal for this type of scenario and it's best to manually create split folds that contain all labels in the validation. This will avoid the NaN in mean output.

If your fold doesn't contain the label in validation, you'll be training a model that won't learn that label at all.