Improving results

[I-CANT-CODE]

So I've gotten pretty good preliminary results with segmentation on my data set. I was wondering does anyone have any ideas or experience with different things to try getting better and better results and improve the model? Data I am working with is pretty noisy/blurry, but it segments ok. Just wondering if anyone had any advice, particularly with getting extra separation between individual units.

I have read a few things like classifying the borders as a separate class, which I did, and using watershed as a preprocessing stage, which didn't seem to work for me. If anyone has some literature they can refer me to let me know!