Cell Detection in tissue - Cell size

I am trying to go through the multiplex analysis of immunofluorescence images.
However, I am not really happy with the cell segmentation.
I am quite confident the DAPI detection is (more or less) good, but when it comes to cell boundaries, I don’t think it gets it right at all.
I am mostly looking at CD3 positive cells (round with cell boundaries very close to the nucleus) and CD68 positive cells (big macrophages, with mostly irregular shapes). This means that the CD3 cells look bigger than I would think and the CD68 cells are either underestimated or fragmented.
The consequence is that if I have the two closely associated, they are often marked as double positive when I try to run a classifier.

Any idea of how I could change the settings to make it work?
Thank you

Hi @APellicoro, I’m skeptical that a change of settings would be enough… you may need a change of strategy / something custom developed for your images. If you can post example images then perhaps someone will have more specific ideas.

I spoke a bit about the limits of the existing cell detection in QuPath - and work to address it - at the workshop hosted at the La Jolla Institute for Immunology, now online at http://tiny.cc/QuPath

Yep, the closest I could recommend is using a very small cell expansion and only asking the question “what is the staining in the area immediately surrounding the nucleus” rather than “how big are my cells and what is an accurate outline.”

Macrophages are almost impossible to get 100% right in 2D images though. Between multiple nuclei and very weird shapes, a 2D plane through one can look very weird. The staining can appear to show up all over the place in random little bits. You might be better off measuring an amount of CD68 stain per area, rather than a cell count, to get an accurate representation of the CD68 positive population.

Thank you both
yes, that’s what I am trying now, hardly any cell expansion, we’ll see. It probably will work for T cells.
I agree for macs: the staining is indeed all over the place in random blobs, the problem is that ideally I would like to get to do some image cytometry, so I need something discrete rather than an area