Robust segmentation

cellprofiler

#1

Hello,

I have been reading and thinking about the various primary and secondary object identification and thresholding options, in the IdentifySecondary module. It seems to me that the “Propagation” option for determining cell regions, and the Otsu method for distinguishing cells from background are both robust (if not independent of) absolute pixel intensities. Therefore if I use the “Propagation” and “Otsu Adaptive” options in IdentifySecondary module, this should segment cells in a way that would be robust against background intensity non-uniformities. Would this be a reasonable conclusion?

Thank you,

Ernest


#2

Hi Ernest,

Yes, I would say your conclusion is generally valid. However, if you are apprehensive about heterogeneities in the background intensity, you might be better served by pre-processing the images prior to thresholding and object identification.

The optimization of the thresholding region for the adaptive method takes place behind the scenes, and so is outside of your control. On the other hand, if you were to use a module like CorrectIllumination_Calculate/Apply or SmoothOrEnhance with top-hat filtering, you can fine-tune the settings and the optimize the input to the identification module of your choice.

Regards,
-Mark


#3

Hi Mark,

Thanks also for this reply. I have gone ahead and implemented an illumination correction algorithm, and corrected a fairly pronounced parabolic non-uniformity in my images. I agree that this is an important step that will allow better control over further processing steps.

But I have a further question concerning the ‘Adaptive’ option. It is a somewhat technical question, but I think it can be readily answered. In the manual entry for IdentifyPrimaryAutomatic, it states that “Adaptive: the threshold varies across the image”, without any further details. I would like to know if the “Adaptive” option implements a unique threshold for each cell, or if the segmentation thresholds between different cells are still related under the “Adaptive” algorithm.

One reason for this question is a few of my images have large, bright, saturated artifacts. I am wondering if using the “Adaptive” option will allow cells outside of the artefact region to be segmented properly, despite the presence of a large local distortion elsewhere. A second reason is that I am interested in studying the variability of metrics over cells. Thus I would like to have an idea of the possible effect of segmentation on the cell-to-cell variability.

Thanks a lot once again,

Ernest


#4

To clarify, the adaptive method divides the image up into equal-sized blocks and then computes the threshold for each block independently. The final image threshold is then taken as the mean of the per-block thresholds. So the algorithm operates in the latter sense, in calculating a unique threshold per block.

If you’re dealing with bright artifacts, you might have a better chance by creating a mask for the artifact using MaskImage, provided the artifact intensity is consistently above a certain value. Then you can use any thresholding method you wish, since they should respect the image mask (e.g, the threshold calculation will disregard the masked pixels).

Regards,
-Mark


#5

Hello Mark,

Many thanks for your latest reply.

To clarify, the adaptive method divides the image up into equal-sized blocks and then computes the threshold for each block independently. The final image threshold is then taken as the mean of the per-block thresholds. So the algorithm operates in the latter sense, in calculating a unique threshold per block.

Would it be possible to let me know what the size of these blocks are? Also, are they computed as a fraction of the total image size, or do they have absolute pixel dimensions?

Best,

Ernest


#6

In a manner of speaking, both. The block size is initialized as the 1/10 of the image size in each direction, with a minimum block size of 50 x 50 pixels. Then a range of acceptable block sizes is calculated as +/- 10% of the suggested block size. The best one is selected as the size that requires the least amount of image padding in order to cleanly divide the image.

Regards,
-Mark