That result doesn’t look IDEAL, but it certainly does look REASONABLE, especially given the clumpiness of the objects and the fact that there aren’t any particularly large intensity changes at the borders. What aspect of the cell are you trying to measure? Depending on the exact metric, that small of an error may not make an issue (and should be present in any conditions you have) - check our blog post on deciding when a segmentation result is good enough.
The only thing that I can think you might try is since you’re using brightfield, and since you have some cells that are out of focus, you could try to use ilastik to solve both problems at once; train a classifier that learns classes something along the lines of “background vs out-of-focus-nuclei vs in-focus-nuclei vs cell-middles vs cell-borders”. Depending on the metric you care about though and how often each of these classes of mistakes are happening, I think you might be spending a lot of time for a minimal improvement; your best thing may just be to go forward.