Positive pixel count detecting negative pixels?

Hi,
I am using QuPath v0.2.0 and I created annotations of tiles that I want to determine the positive pixel count but something odd is happening along the top border of some of the tiles(Please see the fourth image below). The images below show the annotation and detection overlays.




I don’t think that the detections along the top border of the tile are accurate, but I was wondering if anyone else has seen this before? It didn’t happen to all the tiles, but only some of them… The files were scanned with an Aperio scanner so they are .svs files. I would greatly appreciate any advice anyone has!

Thanks so much!

1 Like

Is this the command Positive pixel count (deprecated)?

If so, I’m not sure what precisely has gone wrong here but the command has a numerous limitations which mean it will be removed in a later version. It has been replaced by thresholding and pixel classification – see https://qupath.readthedocs.io/en/latest/docs/tutorials/measuring_areas.html for example.

Thanks for your quick response! Yes, it was the Positive pixel count (deprecated) command :roll_eyes:. I did read that it would be replaced in later versions, but I have not used pixel thresholding yet - but I will certainly check it out!

If you don’t mind me asking, what are the limitations of the Positive pixel count. I read about some of them in the documentation and other posts about the command and I understand the limitations of using pixel count as a metric - especially with DAB - but this was something that I have never encountered so I’m uncertain as to what is going on? If this gets in the “weeds” too much, I understand! But just thought I would try!

Thank you again for your reply and time!

Hi @petercla, it’s mostly the implementation in QuPath rather than the underlying idea. The command was put together rather quickly as an alternative to cell detection some years ago (in 2016…), but I never actually applied it in any research project myself. It then lingered on too long because creating a ‘better’ alternative proved quite a big task…

A fairly fundamental limit of it is works just in a single region – without any clever tiling built-in – and so it is limited in the size of regions it can handle (when developing it, I was just using TMAs so this wasn’t so troublesome). Additionally, the output is one or two ‘detection objects’; if these are large, they can contain a huge number of vertices and make the software performance bad. Finally, it has no preview – so tuning it with sensible thresholds and other parameters is difficult.

Those are the problems that immediate come to mind, but there might be more – your post above has potentially found another one :slight_smile: