How to improve and finetune deepedit performance

[dan-david]

Hi!

I am exploring deepedit with some endoscopy images. Could you give me some guidance here?

• How do I make inference more or less sensible to positive or negative clicks?
• Work with multiple labels at the same time, is that as simple as adding new channels to my model?
• I did try to train the model with high-resolution images, but deepedit stopped working even with a high dice score. Why is that?

To give more context, here is a colab notebook where I am using the deepedit approach:

Colab Notebook

Last but not least, could someone judge if my implementation in the notebook is correct or optimal?

Thank you very much!