I am wowed at the images published recently in BioArxiv in the Zero-shot deconvolution networks paper. It will be interesting to see the peer-reviewed paper.
I would be interested in visualizing the amount of errors made by the network. A simple way to do it is to acquire a short time lapse (eg 10 images) of a fixed sample, run it through the network and see which structures are stably identified and which change from frame to frame. 🙂