Need help segmenting in co-culture


I have a large dataset of 21,120 co-culture images with over 4000 cells per image (so annotating seems impossible). I need help to figure out the best way to segment them.

I made a pipeline to IdentifyPrimaryObjects and to measure granularity, intensity & shape and sizes etc in CellProfiler. Then I created a properties file and used CPA, to separate the co-culture nuclei/cells in CellProfiler Analyst by selecting over 160 positive and negative examples.
But when I run classify, though it does decently identify the nuclei, it counts the same nuclei multiple times. I don’t know how I can improve the current results (screenshot attached).
My main objective is to get correct counts for each type of nuclei in the co-culture.

Any suggestions would be helpful right now.

Pipeline I used - coculture_Manual_overlay_CPA_2.cpproj (418.2 KB)

Best Regards,

Hi Swati,
The classifier is assigning a class (positive or negative) to each segmented object (nuclei).
Here your problem is probably due to an oversegmentation of your nuclei.
In your pipeline you don’t do any preprocessing of your images, and you’re using a manual threshold to segment the nuclei. The “negative” nuclei are bigger and dimmer, so they’re segmented into pieces.
You have to correct this by improving the segmentation first…
Then the classifier is indeed doing a good job (probably the intensity features are sufficient to separate your 2 classes apart)
It means that you could even avoid going through measurement of granularity and using CPA.
By just keeping size and intensity measurements, you might be able (with a good segmentation of individual nuclei) to use the filterObjects module in CP and use limits of size and intensity to directly count your 2 populations in cell profiler…
Good luck,
You can as well upload exemple picts so that we help you finetune the segmentation settings…

Hi Fabien,

Thank you for your help :smile:
I realized that today while seeing the measurements table for intensity, so I used filterObjects for the top 5 features extracted by CPA in the previous run.

The results are pretty good(screenshot attached)

Any suggestions to improve the results by solving underestimation problem for Leukemic Nuclei would be a huge help!
(Problem highlighted in the second screenshot)