For calculating the classwise dice metrics, we currently use torchmetrics.segmentation.GeneralizedDiceScore, which in this context should be equivalent to standard dice score. (Standard Dice so far did not support classwise calculation, because the average='none' option was broken (see issue).

However, one issue with the current implementation of GeneralizedDiceScore in torchmetrics (<1.6.0) is that it assigns scores of 0.0 to samples that don't contain a particular class, which leads to very low classwise dice scores, even if the predictions are of good quality.

Seems that this issue is fixed in the new torchmetrics.segmentation.DiceScore which has been released in torchmetrics 1.6.0, so we should update to that version and move away from GeneralizedDiceScore.