Segmentation model with unbalanced loss function in part 1

Hello everyone, my name is Victor Arias, thank you very much for accepting me into this group, I would like to ask a question. My goal for this course is to segment microaneurysms into fundus images, I would like to know if in the first part of this course it is possible to do so, and if it is, if any of you can please tell me where I can read something to start.

I am not sure I fully understand but semantic segmentation is covered in the last lecture of part 2, lec 14 I believe. It is quite approachable and the fact that it is the last in the course is not indicative of how hard it is.

There are many losses you can use - binary cross entropy, focal loss, lovasz loss, bootstraped cross entropy. You can also use a combination of losses, etc. The problem of the background class being prevalent is what some of the losses attempt to address.

1 Like

Thank you very much @radek

1 Like