Segmentation - counting cells in 3D

Dear ImageJ community,

I am currently working with confocal stacks, in which the nuclei are stained in red and the signal of interest in green (GFP). There are a lot of nuclei in a single image, but I am only interested in those cells that express the green signal.

My current workflow is to perform a Difference of Gaussians, followed by a 3D Watershed on the red channel.

I perform background subtraction on the green channel and then apply a threshold to generate a binary mask, only accounting for the GFP-positive cells. Because of the image noise and due to cellular extensions the mask goes through a Particle Analyzer, excluding all particles that are below a certain size.

I then multiply the green mask with the 3D-Watershed of the red channel and then use the 3D Manager to count the cells.

The problem with this approach is, that I have to subjectively estimate the threshold for the green channel and the maximum size every particle can have each time I perform the count and it seems like this is causing major inconsistencies with the number of GFP-positive cells for each data set.

Does anybody have any suggestions on how this approach could be improved or are there any alternative approaches to count the number of GFP-positive cells in our image data you could recommend?

I have attached the raw GFP and RFP channels in a zip file and another zip file contains the final image that I use to count the GFP-positive cells. (78.8 KB) (19.2 MB)

Thanks in advance for your help!

Hi @disputator1991

Would it be enough to just count the signal in the GFP channel? Or are you interested in some GFP-signal/RFP-signal ratio?

I managed to get a GFP segmentation which looks like this:

But like you I use a fixed threshold and a fixed particle size to get rid of noise.

If you want to test the robustness of my segmentation you can download the KNIME workflow here (7.4 MB).

Normally we need some basic assumptions (like intensity, size of particles) to perform image analysis.

Hi @tibuch

thanks for the help, but unfortunately segmenting only the GFP channel is not sufficient as normally clusters of closely located cells normally express it. This segmentation approach vastly underestimates the number of GFP positive cells.