De-clumping Cells

I have been working on the following pipeline (attached) and found that I had to keep adjusting the settings (IDPrimaryObject) for each round of experiments. The minimum desired final output is CellCount and I am trying to create a pipeline for the lab to use repeatedly.

My guess is that the variation in settings is due experimental variation ranging from shRNA transfection efficiency (changing exposure which thus changes the signal:noise ratio), microscope lamp intensity variation, plate/well shape (shadows), number of cells plated, and any other numerous variables. Additionally the fluorescent scope that is being used is not the best for imaging.

To account for uneven lamp and false positives, I added in an illumination correction calculation and apply (done for each image). This helped greatly, but then I found that I was missing some cells despite how much I played with the IlluminationCalculation or the IDPrimObjects module.

I just added a RescaleIntensity to help me pick up the cells that were not being counted. I found that this was worse in the images that needed a longer exposure. I am not 100% sure how the module works, however after some trial and error have gotten much improved results.

However, I am still struggling with clumping. I am wondering if there is anything else that can be done to help with this? Or, would it be better to fix the lamp issue and the missing objects issue by ways that would make de-clumping easier? Or, is it not feasible for CellProfiler to declump that well?

Note: in the images attached below there are small uniform dots. Those are not cells, those are the fluorescent light shining through the pores of the well insert. Thus, I have set the bottom size limit in the IDPrimary to not measure them.

Thanks in advance!

BloodBrainBarrier16X_05272010.cp (8.07 KB)

Image with “missed” cells before the rescale was added

Image with clumping (extreme case)

Have you tried reducing the smoothing filter size? You can also apply several rounds of identify prim with different thresholds and sumeup the images, that does wonders for me.

1 Like

Hi Meg,

I’ve attached a pipeline which may help, but some comments follow:

Changing the exposure from run to run will probably be one of the prime causes of the issues you are having. If you are using a multi-well plate, it’s probably a good idea to calculate an illumination correction function using all the images on the plate (i.e., using “All” in CorrectIllumCalc) in order to compensate for some of this. Still, try to lock in whatever variables you can as far as the imaging system goes; that way, the degree of freedom in terms of variables to deal with will be biological.

Including the illumination correction is a good idea in your case even if it’s only for detection. I’ve included a revised pipeline in which the illumination correction is implemented using the Regular method which is then divided into the original image. This has the effect of normalizing the image intensities to a degree, making it easier to find consistent parameters in IDPrimObjs.

The pipeline I’ve attached may do away with the need for this module, but you are the best judge. :smile:

This is a difficult case since the objects are so variable in intesnity even within the image. However, using Laplacian of Gaussian (LoG) as I’ve done in the pipeline seems to help. This method basically finds maxima in the intensity regardless of the local intensity itself, allowing you to pick up the cells whether they are dim or bright. These maxima are then used as “seed” points to find the blob that comprises the cells. The main thing is at least threshold the image so that it picks up the dim cells relevively well and then let the LoG do the rest, which the illumination normalization should help with.

A tip wit\h LoG: Setting the filter diameter can be tricky, but the smaller the filter, generally the more objects will be found. However, adjusting the smoothing filter size and maxima suppression settings will most likely be your best bet in adjusting the segmentation quality.

Hope this helps!
2010_05_31.cp (7.34 KB)

1 Like

as always thanks for all the help!

sadly this isn’t large scale plate imaging like i used to do at the broad. this is an experiment where we actually make an “in vitro” blood brain barrier and look at tumor cell migration. as im sure you know, when there are multiple cell lines involved (3 - 6) despite consistency in procedure, variation occurs.

im getting ready to sit for my last final of the spring quarter, but as soon as i am done Wed AM i will start working with the pipeline you sent. i will be sure to let you know the outcome.

best regards,