Segmentation of Cells in a Colony

I would greatly appreciate any feedback or advice on obtaining better segmentation. I am attaching a picture of my colony and the two pipelines with which I have found the best segmentation. Any help is appreciated as there may be something very simple that I overlooked (or don’t know about). From looking at my image I feel that it may benefit from Illumination Correction. However, the best illumination correction that I can come up with doesn’t really look too much better than the uncorrected and it doesn’t segment as well, or possibly I am not optimizing it well enough?


19Aug11_No_Illumination_Correction.cp (3.33 KB)
19Aug11_Illumination_Correction.cp (7.38 KB)

Hi Matt,

I suggest using the EnhanceOrSuppressFeatures module with the method set to enhance and the filter size set to the approximate nucleus size (20 seems to work). This is basically a top-hat filtering method which removes non-uniform intensity heterogeneities larger than the filter size, which can be used as a background subtraction method.


Hi Mark,

Thanks for the hints!! When I apply the EnhanceOrSuppress module to my images the segmentation isn’t improving. I have seen some improvement using FFT low cut filtering. I have attached the two pipelines that I have configured using Suppress and without. Am I using it properly? Is there a better or more straightforward way to come up with a better or at least the same result?

FFTprocessing_without_suppress.cp (6.5 KB)
FFTprocessing_with_suppress.cp (6.81 KB)

I am hard-pressing to find a procedure that successfully finds all nuclei given that they are overlapping and of widely varying intensity. But I’m attcahing a pipeline that gives it a shot. My thought process is the following:

  • Illumination correction to remove the high flourescenet back within the nuclei mass. For some reason, using the Regular/division method didn’t work, so I used the Background/subtraction approach to bring the intensity of the nuclei within the cluster down to the same level as those outside.
  • Smoothing the image helps with the removing the dark dimples within each nuclei. I assume this is what you were attempting the feature suppression? If that’s the case, I think your feature size was too big, on the order of the nuclei, rather than the size of the dimples which are throwing off segmentation. It also gives the same net effect as the FFT approach in ImageJ (which I couldn’t get working on my computer). Some further adjustment here might be necessary.
  • Thresholding by intensity seems to give somewhat better results, although ordinarily I’d prefer Laplacian of Gaussian for cases like this. Also i think the correction factor needs some adjustment, I’m just not sure by how much.

Hope this helps!
2011_09_09.cp (5.11 KB)