Change hovernet losses to MaskedLoss?

[OmarAshkar]

Hi,

I have a challenge in my data of a lot of missing labels. I am wondering if I can use the concept of MaskedLoss here? What I understand that 0's will not be accounted for in the loss.

I was thinking to change these 2 lines of DiceLoss and CrossentopyLoss

MONAI/monai/apps/pathology/losses/hovernet_loss.py

Line 62 in 2daabf9

self.dice = DiceLoss(softmax=True, smooth_dr=1e-03, smooth_nr=1e-03, reduction="sum", batch=True)

I need some guidance on how to change those lines please.

Thanks alot!

[JohnMasoner]

hey @OmarAshkar ,

MaskedLoss may not be reasonable as you describe it. The MaskedLoss def:

region with mask 1 will keep the original value, region with 0 mask will be converted to 0.

but your data of a lot of missing labels.

For the problems with your data, my suggestion you can either set the label to 0, or remove the data with missing labels, or use a semi-supervised training.

Mason

[JohnMasoner]

I'm not sure that methods is a better fit for your data. But my suggestion is that you can use labeled data for training. bc the smei-supervised need a base model has a base dice score.

[OmarAshkar]

Thanks @JohnMasoner for your quick and comprehensive response. I didn't get exactly the idea of weakly supervised. Is there is a monai tutorials and implementation of it?

[JohnMasoner]

As far as I can tell monai does not currently implement this feature. Maybe you can try Unimatch(CVPR2023). (I recommend this repo because I have implemented their work, the repo had a good performance.)