Watershed Segmentation

[765ooo]

Hi
I am trying to segment stained nuclei that are in close contact/proximity.  However the watershed command segments the negative space between the nuclei instead of the nuclei.  

I have inverted the image, thresholded, played around with eroding the mask and more, but the segmentation of the nuclei is still faulty.

[Nathaniel Ryckman]

Haha.

I feel like this is the big issue in using the Gaussian blur, threshold, watershed, particle analysis pipeline. I frankly do not know how people have been getting away with using this method since it seems to produce horrible results in real life conditions (uneven light distribution, cell touching, background noise, etc.). I myself am particularly frustrated because I was hired to improve the results of the implemented pipeline (I get the impression the professor and graduate student that I am working with think this should be easy). What would you do if you were in my situation?

ANYWAY, I am working on improving the pipeline to produce more accurate results, but, until then, you might want to try the method that I found described in this paper (2009):

http://genome.cshlp.org/content/19/11/2113.full

The paper claims accuracies into the 98 percentile, although I am skeptical and haven't been able to test the method out myself yet.

[Nathaniel Ryckman]

Oh...by the way, one other thing you might want to try is to invert the image....

The watershed algorithm might be trying to separate the dark spots, not the light spots.....

[Nathaniel Ryckman]

Sorry, sorry...one last thing. You also should probably binarize the image before you run the watershed algorithm.

Here is a really good explanation of how the Watershed program works. Also, you could obviously look up the paper if you were interested in more of the details:
http://imagejdocu.tudor.lu/doku.php?id=video:segmentation:how_the_watershed_filter_works

Programming isn't magic. If the method can't be done by hand, then the method can't be done by a computer.

[765ooo]

Thank you.
Have done inversion and yes the image is binary - but i will read and watch your recommendations!