I am wondering what is the most appropriate way to report a use of pixel classifier and its quantify accuracy, if it has been used to analysed silicon particles (no cells) in chromogenic stained sample?
From the QuPath side, the main thing is to mention the software by name and its version (e.g. v0.2.3) and then to cite the Sci Reports paper – see https://qupath.readthedocs.io/en/latest/docs/intro/citing.html for details (the pixel classifier didn’t actually exist in 2017, but that remains the current reference). Then you can mention any specific commands you used if they are important to understand your steps.
I’m afraid quantifying accuracy is much trickier, and dependent on the exact application – these comments on object classifier would also apply
You might consider using something like the Dice coefficient, as long as you are comfortable manually annotating a test set. You might be able to create a training image montage from sections of several images, manually annotate them, and then generate a Dice coefficient from the overlap between the manual and automatic segmentation.
Just a thought, and something I saw in some papers recently when they discussed the accuracy of their machine learning algorithms for area segmentation.