I noticed that the reported loss function of condition U-Net in the V2 paper is a combination of the soft dice loss and the cross-entropy loss. However, the anchor loss of the condition U-Net in the V2 implementation looks like a MSE loss instead of the Dice and BCE loss. Can you explain this loss function further? Why is this loss inconsistent with the reported loss in the V2 paper?
Thanks for your sharing about MedSegDiff V2.
I noticed that the reported loss function of condition U-Net in the V2 paper is a combination of the soft dice loss and the cross-entropy loss. However, the anchor loss of the condition U-Net in the V2 implementation looks like a MSE loss instead of the Dice and BCE loss. Can you explain this loss function further? Why is this loss inconsistent with the reported loss in the V2 paper?