In practical medical image segmentation, it is very common to have label values that should be ignored (e.g. padding, unlabeled regions, auxiliary classes).
Currently, MONAI losses and metrics (e.g. DiceLoss, DiceMetric, MeanIoU) do not natively support ignoring specific classes or label values. Users must manually mask predictions and targets, which is error-prone and inconsistent.
Request:
Add native support for an argument such as ignore_index or ignore_classes to losses and metrics.
Example:
DiceCELoss(include_background=False, ignore_index=2)
DiceMetric(include_background=False, ignore_index=2)
Expected behavior:
Voxels with ignored labels are excluded from numerator and denominator
Works consistently for binary and multi-class segmentation
Compatible with reduction="mean" and get_not_nans
This would improve correctness, reproducibility, and align MONAI with common PyTorch APIs (e.g. CrossEntropyLoss(ignore_index)).
Thanks for considering!
In practical medical image segmentation, it is very common to have label values that should be ignored (e.g. padding, unlabeled regions, auxiliary classes).
Currently, MONAI losses and metrics (e.g. DiceLoss, DiceMetric, MeanIoU) do not natively support ignoring specific classes or label values. Users must manually mask predictions and targets, which is error-prone and inconsistent.
Request:
Add native support for an argument such as ignore_index or ignore_classes to losses and metrics.
Example:
DiceCELoss(include_background=False, ignore_index=2)DiceMetric(include_background=False, ignore_index=2)Expected behavior:
Voxels with ignored labels are excluded from numerator and denominator
Works consistently for binary and multi-class segmentation
Compatible with reduction="mean" and get_not_nans
This would improve correctness, reproducibility, and align MONAI with common PyTorch APIs (e.g. CrossEntropyLoss(ignore_index)).
Thanks for considering!