Repository for sharing well-performing classifiers?

Hi,

Is there a place where QuPath users can deposit, perhaps with a brief performance description, the classifiers that they have built? There must be much duplicated effort in this community with everyone training their own classifiers to perform the same task. It’d be great if I could upload my mitosis counter with some brief performance description so that the next person doesn’t have to build one from scratch.

Thoughts?

I could totally see some use for a description of how classifiers were created for particular projects, what problems you ran into, what methods you used to determine which measurements to use to isolate a particular population, that sort of thing. Could even start a thread on it here in the forum (I used subcellular detections of channel X, calculated the intensity sum for channel Y within those subcellular detections, and then created a threshold on the intensity sum per cell to distinguish population A from population B).

The classifiers themselves, though, might only be useful within an institution, and likely not even then. A classifier built for confocal images taken of the same exact sample would likely not work for widefield images simply due to the differences in background when you image all the way through the sample. The thresholds would be too different, and I don’t know of an easy way to manually edit the classifiers… or if that would even be a good idea. Not everyone will order their channels the same, have the same detectors, set the same exposure times/gains/laser powers.

These sorts of problems might be slightly mitigated if the channels were named as in version m4, and the classifier were looking at channel names instead of channel positions (most versions)… but that brings up another thought, that as channel names change for fluorescent channels between versions of the program, and that could add an additional complication.

That doesn’t even take into account differences in fixation protocol, staining times, new batches of stain or antibody, etc. Great talks on those at the QuPath workshop at LJI!

If you have a script you want to share, then creating a Gist like this one here, might be a good way to go. Though I don’t think there are any central Gist repositories for scripts at the moment.