Hi Sik-Ho (Please forgive me if I'm typing your name wrong).

It was a great article. There are a few clarifications that I wanted :

1 - How is w_c(x) computed? Is it multiplied to the actual class value? For calculating cross entropy? Or are the class values modified to form w_c(x)?

2 - By Overlap-Tile Strategy, do you mean padding but instead of zero padding, they use mirror padding?

3 - In the paper, they mention using a Drop-Out layer after the contracting path. Do they use it anywhere else in the network?

4- Can you elaborate more on the data augmentation part of the training? The paper makes it very vague for me.