Segmentation does not have good separation between multiple instances of cells

[MBronars]

I have trained U-Net to detect instances of stem cells in bright-field images. Current performance is fairly good, with a final IoU of around ~.7. However, in scenarios where the cells are touching, I am not able to get the network to detect this as two seperate cells. Do you have any suggestions about how I can fix this? I went back and redid my annotations to ensure that I left boundary between any cells that were touching. I also looked into creating a custom model and changing some of the parameters there, but I didn't see much improved results. Thanks for the help, and here is a picture of that I am talking about

(Recognizes the right cells as two separate cells, but not the top cells)