Generalise DiceLoss never decrease about boundary-loss HOT 3 CLOSED

Hello, I have new something to share with you. In my experiments, I found that the GDL is hard to decrease. According to the Paper "AnatomyNet: Deep Learning for Fast...", the generalized dice loss uses squared volume weights. However, it makes the optimization unstable in the extremely unbalanced segmentation. I've replaced GDL with DICELoss and only focused on the classes of interest, which will decrease as expected. I said before that diceloss didn't work, maybe it was my little mistake. Is it still possible to use boundaryoss and diceloss together? Maybe I'll add FocalLoss, because diceloss is unstable.

from boundary-loss.

Hello,

Yes, you can use Boundary loss with any other loss; they are all compatible. It is about finding a combination that works for your specific problem.

My experience with the GDL is rather “chaotic”, and I usually prefer other losses (even the simple cross-entropy loss). The focal loss is indeed quite powerful, from my more recent experiments, so I encourage you to give it a go.

I've replaced GDL with DICELoss and only focused on the classes of interest, which will decrease as expected.

How many classes does your problem have? Indeed, supervising different classes differently (either with a different weight, or another loss) is not out of the ordinary. You can also decide to supervise the background explicitely (idc=[0,1,2...], or implicitely (idc=[1,2,...]).

Let me know,

Hoel

from boundary-loss.

Hello, I have new something to share with you. In my experiments, I found that the GDL is hard to decrease. According to the Paper "AnatomyNet: Deep Learning for Fast...", the generalized dice loss uses squared volume weights. However, it makes the optimization unstable in the extremely unbalanced segmentation.

Related to that general topic, there is that pre-print (still under revision) from colleagues that might interest you:
The hidden label-marginal biases of segmentation losses

from boundary-loss.