search

hci-unihd / antibodies-analysis-issues

Issue tracker for problems in the antibodies analysis workflow.
0 stars 0 forks source link

Plate titration_plate_20200403_154849 Site F08-0003 #13

Open tischi opened 4 years ago

tischi commented 4 years ago

{"plateName":"titration_plate_20200403_154849","siteName":"F08-0003","pixelLocation":[22132.557453701334,15205.579169115894,0.0]}

Another example of a wrongly (infected) classified cell. I do not really get why, because the cell segmentation looks good, i.e. no overlap with the surrounding infected cells.

tischi commented 4 years ago

image image

imagirom commented 4 years ago

The method to detect infected cells is really simplistic at the moment: Subtract the mean background value of the marker channel, then compute 50th brightest pixel in that channel per segment. If that intensity is greater than 250, the cell is declared as infected. This can obviously fail if

  1. The background segmentation is bad
  2. The Image is not uniformly illuminated (barrel correction did not work perfectly)
  3. The Image is illuminated a lot stronger / weaker, changing the scale of noise and signal (This can be solved by dividing by the background instead of substracting it; however it lead to slightly worse results when I last tried it)
  4. Due to noise there are 50 bright pixels in a non-infected cell
  5. there are less than 50 bright pixels in an infected cell (maybe if it is very small? I think this point is not as relevant as the others)

We could adress 4. by denoising the image, e.g. simply by gaussian smoothing. But I think we just really need to come up with a better method that is also robust to the other points.

tischi commented 4 years ago

There are also cases of weakly infected cells (the one in the middle of below screenshot) image Interestingly the serum response is quite strong in this cell: image

tischi commented 4 years ago

Here is a region which is quite noisy and indeed there are many false positives: image

tischi commented 4 years ago

@imagirom Where does the threshold of 250 come from?

imagirom commented 4 years ago

Grid search and evaluation on the plates vibor has provided ground truth for (excluding the titration plate). So it might very well be overfitting to these plates.

tischi commented 4 years ago

Ok, let's say the fluorescence lamp that Vibor uses on his microscope gets a bit dirty and everything is a bit dimmer, would that still be the right value? I mean is the 250 normalised to anything you just an absolute gray value?

imagirom commented 4 years ago

It is an absolute gray value. As I mentioned before, I totally agree that we need something better here.

tischi commented 4 years ago

Ok! I started an issue about it: https://github.com/hci-unihd/antibodies-analysis-issues/issues/17