Some Questions on Your Iterative Process Mentioned in Paper

[riven314]

I appreciate your work. It is a field that not many deep learning researchers work on and I am glad to see an encouraging result from your work. You mentioned from paper that you tried various architectures and design of model before the final work. I am interested to learn more about it:

1. What other model architectures did you considered (e.g. ResNet, VGG … etc)? and how was their performance compared to your "modified UNet"?
2. What do you mean by "class weight" in the paper
3. Did you try to encode the two chromosomes in the same class (result in 3 classes in total) ? Or encode the two chromosomes and background in the same class (result in 2 classes in total)? If so, what are their performance?
4. I noticed you did many measures to reduce the computational cost (e.g. reduce image size, simplify UNet). How long did it take to train your "modified UNet"?