Multi-label diceloss

[AugustKRZhu]

Hi there,
Thanks to the MONAI team.
Recently, the multilabel segment on CT has been studied, where label 2 is a subset of label 1, and label 1 is the constrain of label 2. My question is whether Diceloss or other loss functions can calculate independent label1 (label1+overlap of label1 and label2) and label2.

[Nic-Ma]

Hi @AugustKRZhu ,

Thanks for your interest here.
I think DiceLoss can work well for your case, you can refer to the Brats segmentation tutorial:
https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/brats_segmentation_3d.ipynb
It's multi-labels task.

Thanks.

[AugustKRZhu]

Hi Nic @Nic-Ma ,

Thanks for your reply, I refer to the Brats segmentation tutorial and train my dataset now, but I am confused about whether the dice output is correct. as shown in fig, the dice losses are not mean values, are they suitable?

thanks.

[Nic-Ma]

Hi @AugustKRZhu ,

Actually, it's the mean value by reduction in the aggregate():
https://github.com/Project-MONAI/MONAI/blob/dev/monai/metrics/meandice.py#L96

Thanks.

[AugustKRZhu]

Awesome, you are right, I found that if I try to average them, it will be a problem. Thanks. @Nic-Ma