Consulting about HD loss about hausdorffloss HOT 2 CLOSED

I believe that's a common problem for distance based losses. They tend to behave unstable when net returns random results (as it is for the first few steps) and label everything as background when classes are highly imbalanced.

I think the problem may lay in your alpha change being dependent on loss value thus being very unstable in early stages. I would try to define loss as: loss = dice * alpha + Haussdorf * (1 - alpha), and start for example from alpha=0.95 and define it to be 0.5 or even lower at the training end - alpha will progressively go down during training time. This way you allow dice do the heavy lifting early to establish some reasonable prediction and not allow early Haussdorf anomalies confuse the network that much. Later on when net outputs better predictions Haussdorf loss should be stable.

from hausdorffloss.

Thanks a lot for your advice. I have used the trick that you provided. The initial alpha is set to 0.00001, and the alpha increases by 0.00002 for every iteration. Then the network can find the vessels in the training and testing. The loss curve is attached. The loss is a little unstable, and I think I need to give a better initial alpha and a better strategy to change the alpha.

from hausdorffloss.