Using Ilastik to quantify intestinal epithelial cells

Hey fellas,
I’ve been spending the last couple of days reading Ilastik’s manual and playing around with it. I’m really impressed and I’m rather optimistic that it can help me, but I’m still not sure if and how it can, so I wondered if you could help me out here :slight_smile:

The image I’ve attached is of an intestinal section; red is the nuclei, blue marks the intestinal epithelium, and green marks a subset of the intestinal epithelium. My goal is to quantify the number of cells in green and normalize it to the number of all epithelial (blue cells), or at least to the length/area covered by the epithelial cells, and to do that over a large amount of images.

I have three pressing questions at the moment:

  1. Is it possible to do that with Ilastik? So far I’ve been able to use the pixel identification and the segmentation applets on a few images to mark the green cells individualy and accurately, as well as to mark the area covered by the blue cells, though, for now, not as individual cells. Do you think it would be possible to use Ilastik to count the number of individual blue cells with a reasonable accuracy?
  2. Let’s say that I have 20 images and I wish to use 5 of them to train “Ilastik” to identify my cells of interest, and then let the program automatically identify these cell types in the remaining 15 images based on the training I’ve provided in the first 5 images. How do I that, technically?
  3. How do I get the program to count specifically the green cells which are also blue (i.e. green epithelial cells), but not (or separately from) the green cells which don’t fluoresce in the blue channel (i.e. green cells which are not epithelium)?

Looking forward to hearing from you :slight_smile:
562651750824016 - without EGFP.tif (10.7 MB)

Hi @omers

  1. It looks to me that pretty much all the cells are expressing your blue marker at the plasma membrane - is that correct?
  2. I’m not sure I understand the question? You add your images in the ilastik interface to train your classifier, then run the resultant trained classifier (in batch mode) on the remainder of your images.
  3. This sounds like a job for the object classification workflow. If you can somehow segment all your cells, you can then classify them according to differing expression levels:


Hey there, djpbarry, thanks a lot for the quick response!

  1. I need to divide your question and my answer in two: the cells which do express the blue marker (which is EpCAM, btw) express it at the plasma membrane indeed; however, not all the cells in the image express EpCAM (most cells underneath the blue-marked layer do not express EpCAM (there is some autofluorescence though, so some of them may end up marked in blue, but at a lower intensity than the truly EpCAM-positive cells).
  2. Let me see if I understood your response correctly: so, I should load some training images in “input data”, train the classifier in the next steps and then, when getting to “batch processing” I should load the test (previously-unseen) images (without re-loading the training images)? And in the future, when I have new images (from the same issue and with the same kind of stainings) to analyze, do I load the project with the same (original) training images, but erase the testing images and replace them with the new (unseen) ones?
  3. The problem with the user guide you’ve linked to is that it describes classification of objects in any image with just one color in the forground (and, of course, one background color). So, it doesn’t apply to when you need to count cells showing colocalization. In other words, the linked manual would help me to detect/classify all the blue cells and all the green cells, but I can’t figure out how to detect/classify specifically those green cells which are also positive for the blue marker (i.e. each cells which is marked both by the blue and by the green marker, as opposed to ones marked only by the green marker).
  1. So what strategy are you employing to segment cells? Are you attempting to segment nuclei? Are you attempting to segment the membrane and maybe employ a graph-cut strategy?
  2. You don’t “load” test images when it comes to batch processing - you just tell #ilastik where the images are stored. You only load images in the ilastik interface for training purposes.
  3. No, the intensity measure features used for object classification are multi-channel.
1 Like