Giant cells - titration_plate_20200403_154849 - G04-0002

[tischi]

{"plateName":"titration_plate_20200403_154849","siteName":"G04-0002","pixelLocation":[11811.997148196311,17397.64991544256,0.0]}

More giant cells

Screenshot

[tischi]

@constantinpape @wolny

I think it may be worth at some point to train something that predicts a mask for these blurry dirt regions (as in above screenshot) such that we can specifically mask out these regions from the images and still use the rest of the image. Because really many images have such a region, of varying size, and it is a shame to throw away all the other cells in such an image by classifying the whole image as having a problem.

If you agree on this we could make an issue in the batchlib repo.

[constantinpape]

In which channel is this? Only in dapi or also in the other channels?

Because really many images have such a region, of varying size, and it is a shame to throw away all the other cells in such an image by classifying the whole image as having a problem.

If you agree on this we could make an issue in the batchlib repo.

Yes, we should do something about it, but we might be able to do this in the cell level qc:

• add a classifier (RF) that predict whether a given cell contains such a defect or not based on the cell features we extract. this would need much less training data then training something pixelwise.

And feel free to make a batchlib issue to brainstorm more ideas.

[tischi]

In which channel is this? Only in dapi or also in the other channels?

Often also in the serum channel, sometimes also in DAPI. As said I think it is not super critical now, because excluding the giant cells based on the single size feature would already help a lot. We can add more features (and an RF) later.

[wolny]

I agree that it should be handled via cell level qc and I'd rather have some hand engineered features to find those cells: having too many of those classifiers in the pipeline will lead to maintenance issues in the long run.