Adjustment of cell detection tool (ki67)

Hi all,
I want to count the percentage of ki67-positive tumor cells in tumor tissue that also contains stroma. For this I plan to use the (1) cell detection and then (2) create annotations in order to train a classifier based on those annotations.
My problem is the cell detection of ki67-stained tumor cells. Many tumor cells have a dot-like, punctate nuclear stain patten, which necessarily isn’t to be considered as positive, however, many of them are not identified at all by the detection tool. Also many negative tumor cells are not identified adequately. Stroma cells are picked up more easily.
I read that optical density sum is recommended in situations with punctate nuclear staining patterns. I have tried to adjust the nucleus-, cell- and general parameters extensively but just can’t get it identify all tumor-negative and tumor-positive cells. I intend to use a single threshold. I use default stain vectors.
Here is part of the original image. Tumor cells are morphologically easy to distinguish from stromal cells.

Here the detection. Many tumor cells (positive and negative) are not encircled properly.

Threshold seems right as far I can see

I’m using 0.2.0-m8 version.

Thanks in advance!
Pentala

1 Like

If you are able to share a copy of an image on GoogleDrive or similar, I would be interested in trying “a few tricks,” but that looks fairly challenging for QuPath’s current cell detection algorithms. Maybe @petebankhead or @melvingelbard will have a straightforward answer, but I can’t think of anything simple given that staining pattern.

My first suggestion would be to reduce the stain intensity in the KI67 staining step, either AB concentration (titration) or otherwise.

Not necessarily - it really just depends upon which works better. I recommend trying both hematoxylin and optical density sum, and using the ‘Brightness/Contrast’ window to visualize how both appear.

I think the challenge here is that the positive nuclei are extremely dark relative to the negative nuclei. This dramatic change in intensity is something that QuPath’s cell detection algorithm cannot really cope with (technically, it uses a laplacian of Gaussian filter for a key step and this can sometimes result in anything immediately adjacent to a very bright/dark structure being ‘lost’; if nuclei are ‘similarly dark’ this isn’t such an issue).

You might actually find using hematoxylin for detection works better here than optical density sum - but I’m not sure.

I will add that if you are using haematoxylin for cell detection, I found that getting the stain deconvolution values right first is extremely important, otherwise very dark areas of DAB staining will show up in the haematoxylin deconvoluted channel.

You may need to get a good number of areas with varying levels of haematoxylin and DAB staining, creating a composite (On M9, Classify > Extras > Create combined training image), and then automatically Estimating stain vectors using the composite.

Good luck!

1 Like