Is it possible to quantify histopathology structure using Qupath?

Dear All

I have couple of H&E histopathology slides and each slide has several glandular structures. I would like to quantify these glands. My initial strategy is to train Qupath to recognize the stroma from gland that contains columnar cells and goblet cells (cell has bubble appearance without nucleus). Once all cells are detected, I plan to run delaunay cluster feature to quantify these glands. Unfortunately, Qupath seemed to have difficulty to detect goblet cells. Does anyone have the solution or alternative strategies?

Thank you.

H%26E [Detection.tif|attachment]Detection

I have had trouble with non-standard cell types like that before. QuPath doesn’t really have a lot of options as far as cell shapes are concerned, and a deep learning or pixel classifier approach may work better (once the pixel classifier is scriptable).

Most software will detect the nucleus and then expand outward blindly by a set amount. The goblet cells just aren’t made like that. I would generally use some sort of area classifier to identify the golbet cell cluster areas, and then count nuclei within them.

One thing you could try to improve your results is the Add Intensity Features command. Rather than ROI as the area option (which searches the cell itself), use a circular tile with a set area. Run this a couple of times at different expansions (possibly using the Haralick features) to get some contextual information about the area around the cell.

A quick shot at using SLICs to classify areas, which could then be used for cell detection:

Obviously not great, though I would have needed many more images for data for the training set to make it better, given that my number of measurements was greater than the size of my training set! You may also be able to do better with the original image using accurate pixel metadata. Always easier to work in microns :slight_smile:

*Hmm, actually playing with the pixel classifier in 0.2.0m2 I wasn’t able to do all that much better. It might be better to solve this problem using a stain for the goblet cells, though if the stain chosen is too dark, that might make it difficult to actually count the goblet cells… FastRed with a hematoxylin counterstain maybe?

I followed the guidance and came out with an acceptable result. What I did was

  1. Generate region identification by SLIC superpixels segmentation
  2. Add intensity features
  3. Add coherence textual features
  4. Create cell specific annotation including foveolar cells, goblet cells, stroma and blank area.
  5. Train the classifier using random forest.

I am interested in the number of these glands (area mixed red and blue pathes/cells). I could simply count these features by eyes but wonder if there is any better solution? Thanks again for all the help.

I suspect not without a bit more coding. It might be possible to take an image like that and write a custom ImageJ script to locate the objects, or you could look into the pixel classifier in 0.2.0 (I didn’t have a lot more luck than what you are looking at there). The tiles to annotations would, ideally, allow you to count annotations, but even with a size threshold, there are too many annotations that touch. Even the smallest shared border would result in one larger object, and I don’t know of any way to watershed a bunch of created annotations.

It definitely looks like the sort of thing even a relatively simple deep learning classifier would make short work of, but machine learning tools will be stretched to get close.