2
$\begingroup$

I am currently trying to segment cells from digital pathology images.

The method I use is an algorithm based on color distance. This works for most of the cases, however, when dealing with the images which cells are overstaining, then the segmentation results is pretty bad. Is there any golden-standard method in treating these poorly stained with highly adhered-cell images?

enter image description here

$\endgroup$
  • 2
    $\begingroup$ First, can you show an example of a good and bad segmentation? Also, there’s a variety of things you could do. A simple thing here you could do is take the pixels from a single image like above and do k-means with k=2 and use that to segment out the more visible cells from the marked areas (or vice versa). This may resemble your color distance approach. You could use something smarter that takes into account spatial information. For example, you could mark up a few photos like the one above and create a classification dataset that can be used to train a convolution neural net classifier. $\endgroup$ – spektr Apr 26 at 5:09
  • $\begingroup$ You could then use the CNN classifier to decide what part of the image should be segmented out. There’s also other models you can use for the classifier instead of a CNN if you desire. You could even use graphical models if you want, though that might be a pain compared to something like a CNN or SVM. $\endgroup$ – spektr Apr 26 at 5:11
  • $\begingroup$ Looks like their edges are still present (you mean the dark blue ones?), that could probably be used. $\endgroup$ – Emil Apr 26 at 5:38
  • 3
    $\begingroup$ What kind of Software are you currently using? How many Images do you want to segment? If it is a one-off thing, you might use the Matlab imaging toolbox. It seems there is quite a range of Segmentation algorithms available in Matlab: [de.mathworks.com/discovery/image-segmentation.html](Matlab: ImageSegmentation) $\endgroup$ – MPIchael Apr 26 at 13:41
  • $\begingroup$ I second the need for the question to have additional explanatory material. Also, check these: stackoverflow.com/a/53566610/752843 and stackoverflow.com/a/51717007/752843 $\endgroup$ – Richard Apr 26 at 17:37