Counting nuclei with chromatin speckles

Hi all,

I found this post which has a similar title with no answer so I didn’t know if I should make a new post or not, let me know.

As you can see from this image, I have densely packed nuclei with bright chromatin dots which are making the segmentation very difficult since most algorithms find these dots. Do you have any ideas ?

I even tried with pretrained StarDist models but it finds some nuclei very nicely while others only the bright spots.

1 Like

Hey @lguerard,

not sure if this is the perfect answer, but: In the age before StarDist, we used to process these kind of datasets by estimating an average cell area (one could do this by outlining some which are differentiable) and measuring the area of the whole area of touching cells. The whole area divided by the estimated cell area gives you the number of cells - estimated. Maybe that’s a starting point?

Let me know what you think.


I would suspect training a dedicated stardist model has a good chance to work pretty well in this case. For that you would have to annotate some crops (e.g. 5-10 images containing 10-20 nuclei) and train a new model with our example notebooks. You can as well try to add the data from the pretrained network. Additionally, I would recommend having a look at the nice stardist workflow descriptions by @oburri that contains images similar to yours.
Hope that helps!


Hey all

@haesleinhuepf : Thanks a lot for your input, this is the way we thought about it if we can’t find anything better but I still wanted to try :wink:

@mweigert : Thanks a lot for your reply ! Training a new model is indeed what we thought would be best, but it still requires some work by the user and he’s not sure yet how much time he wants to spend on this project so I was looking for a quick and dirty method if it existed ! :slight_smile:

1 Like