Challenging clumped DAPI cells



Hello everybody,

I´m glad to have the opportunity to use cellprofiler programm.

After more than one week trying to build a pipeline to count Ki67 positive cells, I´m now desperate and I really need your help!!

The most challenging problem ist to de clump Ki67 cells. I tried different ways in the IdentifyPrimaryObjects modul how I have read in diffrent posts. I´m not sure, if somebody with more expertise could help me. Maybe with adding another module to preprocess the image?!

Thanks for every help!!


Ki67_DAPI_DAPIposKi67_Counts_ObjectsMaskedObjects_final.cpproj (659.7 KB)


@cljel, these images will be a challenge to cleanly segment given the density and overlap between nuclei. That said, I think you can segment nuclei with enough confidence to meaningfully quantify the phenomenon, i.e. Ki67 expression, you’re interested in.

I looked at your pipeline and would make a couple changes to simplify the image analysis. Primarily, I would segment nuclei solely on the DNA channel. From here you can then measure Ki67 and use the distribution of Ki67 across the nuclei to classify your cells.

One feature generated by CellProfiler is recording the upper quartile intensity of an image, such as your Ki67 channel. If you were to make your classification decision on a single feature, this one is somewhat robust to segmentation errors: integrated intensity is highly correlated to the area of nuclei and mean intensity can lower signal-to-noise if Ki-67 isn’t uniformly expressed throughout a segmented object.

In the figure below, it looks like there are two populations defined by the upper quartile intenstiy of Ki67


After measuring this distribution in a CellProfiler pipeline, a follow-up pipeline can be created that classifies cells using the FilterObjects module. Doing so leads to the following result:

The above image is the Ki67 channel with outlines of positive nuclei in green and negative nuclei in purple.

We can use single cell data to describe the population of cells captured in this image. The nuclei counts for this image are 847 total nuclei and 662 Ki67 positive nuclei. This means 78% of the population is Ki67 positive.

Alternatively, if segmenting single nuclei proves too challenging you can still quantify the entire image. If we threshold an entire DAPI image and Ki67 image, and then calculate the percentage of total Ki67 area with DAPI area we find a similar number, 77%; note Ki67 is nucleus localized, so we expect the segmented area to be closely related. To support this thought, if we add up the DAPI signal in the areas occupied by DAPI area and DAPI signal within Ki67 area we also find a similar number, 74%

Please check out the pipelines I tweaked in the zip file: (4.3 MB)