How to choose the correct loss function (Image Segmentation)

Hey,
I am working on (binary) semantic segmentation and thinking about changing my loss function to improve my model (currently F.cross_entropy) and found this:


(source of image)

How do I choose which loss function is usefull in my case?

This is how my output looks at the moment (left GT, right model prediction):

Is there something like a guide to choose a loss function somewhere? :smiley:

Thanks for any help :slight_smile: