Quantifying tumor tissue and non-tumor tissue

Hello, dear colleagues,

I need help with some really simple task like for QuPath:
I have an experimental lung from mouse with metastatic melanoma. I wanna measure how much µm² and also “%” tumor and normal tissue i have.
I tried to do it with pixel classifyer, but the resuöts didn’t made me happy, because:

  1. alveoli in the lung are transparent and dont have nucleated cells - thais way it wasn’t detected like tissue;
  2. with pixel classifyer i have a lot of cell parameter in “measure”, but i didn’t found “totally area”

This reason i marked tumor areas and lung manually with “wall” and try to calculate this with help of “measure” - “show annotation measurements”… but i have the feeling, that % with QuPath are too low, for example on the microscope i give 25% and QuPath numbers give me just 10%. I think that this is big differences.
Qould you please tell me, or QuPath measure just pixel size or some more parameter? Is it possible to perform my measures in another way?

Thank you in advance - i add pictures in comment

Sincerelly, Eugen

I’m not really clear on your questions as the images certainly show tissue and annotations, but not what is wrong with them or your desired result.
I would recommend reading the documentation on how to use the pixel classifier, as it doesn’t sound like you are looking for the result in the right place.
https://qupath.readthedocs.io/en/latest/docs/tutorials/pixel_classification.html

This problem is generally why we use image analysis. Humans are bad at measuring things in general. When looking at your images, you can’t use a feeling that the values are wrong; you need to provide examples of how it is wrong, especially if the classifier is systematically wrong. And once you know how it is incorrect, then you fix it with better training examples, if possible.

Not everything can be pixel classified. In some cases creating SLICs and adding more features will give you greater control. I just went through this earlier in the year when the pixel classifier couldn’t catch both the very fine texture detail and the very large structures at the same time, having only from 0.5 to 8x the chosen pixel size. Finding the right collection of variables is fairly tricky though, and it would be much better if you can get your pixel classifier to work.

An example of the pixel classifier measurements within a constrained area, in the lower left, with the annotation selected.