Reticular Background Problems (False negatives/positives)

Hello,

I’ve been working with CP for about a week, trying to identify/count clusters of my fluorescently-tagged protein. I love the software, but I’m having problems with false negatives and false positives in one of my critical assays.

My goal is to quantify the number of foci in the cytoplasmic/ER regions of my cells. This overlaps with the nucleus for my protein, so the “Identify Primary Objects” tool does a great job isolating this region by tracing around the fluorescent signal. No need to exclude the nuclear region with a mask, so I limit the analysis to a single channel from my images.

The general phenotypes of the labeled structures are either:

  1. Reticular/webbed (consistent with an ER-localized protein),
  2. Discrete/punctate (9-14 pixel circular foci)
    or
  3. A mix of these two phenotypes.

The foci show up more with some mutant forms of my tagged protein than others, and I’d like to quantify those differences. The eventual idea is to move this toward a high-content screening operation, so sensitivity and reliability are critical.

The problem is that no matter what pipeline and image enhancement/masking strategy I’ve devised, CP ends up dividing the areas with nice ER morphology into little spots, and counting them. So when I have a control cell with no visible foci, the program ends up counting upwards of 30, 70, or sometimes hundreds of foci. This happens in about 20% of the cells I analyze (the others yield a correct “0” or small number of foci).

To minimize the false positives, I’ve had to increase the stringency of the analysis, and thereby introduce false negatives. Quite a challenge so far. I’d like the numbers to tightly track what I can count manually.

In this pipeline I do the following steps:
Find the object outline >> Suppress small features >> enhance speckles larger than the suppressed size (near foci size) >> detect foci based on size and thresholding above background >> classification, relation, math, etc. >> export

What could I do better with my approach? I’m aware of the varying intensity in several of my images being a possible problem, but this is the nature of the samples, unfortunately.

I’ve included the basic pipeline here, as well as a set of images that illustrate the issue.

Any help would be appreciated - thanks for the software and community support!

Cheers








100X_Foci_Auto2.cppipe (13.1 KB)

Hi,

Yes, this is a frustrating one. I made an attempt starting with your pipeline, but I didn’t have much better success. We need a better particle vs. line filter, which is the distinguishing characteristic here between your samples, it seems to me. We have added one recently (ilastik.org/documentation/counting/counting.html . Note that this version of ilastik does not yet work with CellProfiler, so take this suggestion only as far as it goes.

Hope that helps,
David
100X_Foci_Auto2.cppipe (15 KB)