Seeking help for young scientist


I am writing on behalf of my daughter. Mikaela is quite a serious young scientist (14), having won gold and platinum medals (biotechnology) at last year’s national science fair competition with her study of algal species interactions in bioreactors. She is extending her research this year with an investigation of population dynamics - more specifically the robustness and performance of multi-species cultures under non equilibrium conditions.

Mikaela is looking to use Cell Profiler to track cell populations from samples taken from continuoulsy running chemostats. The hope is that CP will be able to count AND classify up to three morphologically distinct freshwater algal species:

Scenedesmus Obliquus (SO) - oblong cells connected in four-cell coenobia
Chlorella Vulgaris (CV) - round
Pseudokirchneriella subcapitata (PS)

The unstained cells are being imaged in cell count chambers. The CP challenge is to create accurate outlines of the cells. The interior detail is not important to the study.

I have been attempting to help Mikaela set up the pipelines and we’ve had success with getting accurate counts and cell outlines for two of the three species (the round CVs and oblong SOs). However, finding a CP configuration that can consistently distinguish between all three species continues to elude.

With regards to staining: As there will be over 1,000 samples taken, a lengthy staining procedure is not really viable - unless there is some essentially instant contrast enhancement technique that can be applied during the sampling. e.g. Filters? Thoughts?

With regards to the quality of the current imagery - Our ‘home lab’s’ generic microscope has illumination and spherical aberration limitations that became very apparent once working with CP. The halo effect around the cell bodies may turn out to be our main problem. Fortunately a microscope manufacturer has offered to loan Mikaela clinical grade equipment so image quality should not be an issue in the near future.

Despite the optical challenges, I believe we’ve made pretty good progress on pipelines and we’re anxious to get some early feedback as to whether we’re on the right track.

Please find attached:

a) An image series of a mixed algal species culture (SO and PS cells), and

b) A pipeline that can be tweaked in various directions (relative weight color channel, thresholds, etc) to do a good job counting and tracing the cell boundaries of the CV and SO cells. However the PS cells do not have the same level of interior contrast. Tuning the pipeline to do a fair job on tracing PS cell boundaries causes the SO four cell coenobia to trace as two cell blobs instead of four distinct cells. The provided pipeline tuning is a compromise. The pipeline includes further comments in the module notes.

Mikaela would greatly appreciate any advice that could be offered.

Thank you

Chuck b/o Mikaela

Generic_Brightfield_40x_PS-SO_V12.cp (19.6 KB) (7.1 MB)


In the process of composing last night’s blog, it occuured to me that perhaps the use of an illumination filter might offer a convenient alternative to staining. I was able try a blue filter today. Happy to say the filter greatly improved the contrast of the green algal cells against the background. The cell outline traces of the PS cells are very much improved. Also the there now appears to be far greater tolerance of background noise (i.e. less background artifacts). The central problem however still remains: The cresent shaped CV cells tend to become segmented into two cells. Increasing the thresholds to the point where this effect is minimized causes the 4-cell SO coenobia tend to blur into two cell blobs. There doesn’t seem to be an acceptable mean. So the end result is too high a count for CV cells and too low a count for SO cells. I’m hoping there are some options other than threshold factor to improve the situation. I’ve tweaked the settings a bit but I would very much appreciate getting someone’s feedback on the general approach before going too far. Hopefully you feel the use of an illumination filter was in fact a good idea and I that didn’t waste anyone’s time with the first set of data.

Please find attached a new set of images (with blue filter in place) and a new pipeline.

Chuck b/o Mikaela
Generic_Brightfield_40x_PS-SO_V13.cp (19.7 KB) (3.28 MB)

Hi Chuck,

I’ve revised your pipeline and attached it to this post. Some comments follow:
[ul][li]I’ve performed image division with the blue and green channels to try to enhance the contrast. This seemed to work, although I have a couple of caveats:
[list][]If the color balance changes in later images, this approach may not work[/li][li]I’m hard pressed to give a formula as to whether division works better than subtraction, wor which channels are appropriate for this purpose[/li][/ul][/:m][li]With the improved contrast, 3-class Otsu with the middle class assigned to background seemed to do well.[/li][li]My approach was to tweak the correction factor to better separate adjacent cells without any declumping. Once that was optimized, I could start on adjusting the declumping parameters.[/li][li]Using shape seemed to work over intensity for the declumping method. Since most of the cells were oblong, this makes sense to me.[/li][li]The hard part will be adjusting the smoothing and maxima suppression settings. One note is that the automatic values for both of these parameters are based on the limit bound for the size criteria, so you are better served by setting these manually. I took a first crack at it, but you will need to explore these further.[/li][/list:u]
2011_02_02.cp (6.38 KB)

Hello Mark,

Thanks so much for reviewing the pipeline we’re developing to support Mikaela’s population dynamics project. Since our original post we’ve made quite a few enhancements to the original design and by incorporating some of your suggestions are able to get CP to create fairly decent cell outlines - and importantly in a reliable fashion that doesn’t require a great deal of tweaking for each data set. here’s a summary of what’ve we done since last post:

We now run two pipes sequentially. We find this more efficient procedurally as if there any tweaks it is generally only on the 2nd pipe.

FIrst Pipeline

  1. Split color in to channels (as opposed to simply create grey scale)
    - this made a huge difference as we determined that most background noise was on red and blue channels which we smooth

  2. Invert each channel

  3. Correct for illumination each channel independently

    • also a major improvement as we saw differences in illumination issues on each channel
  4. Save Corrected Images for 2nd pipeline.

Second Pipe

  1. Smooth (median smoothing) the red and blue channels

    • the red and blue have the most backgoround noise)
  2. Combine

  3. IdentifyPrimary

    • used your suggestions on threshold tuning and using shape instead of intensity for declumping
  4. IndentifySecondary

  5. Create Database for CPA

Note we didn’t use the blue-green channel contrasting technique you tried because, as per your point, we found it made pipe too sensitve to input conditions.

So, we seem to be getting decent and consistent results and we expect those results to improve when we get use of clinical grade optics starting this week. I think we’re in pretty good shape on CP for now, but if you have any thoughts on our pipe design that would be great too.

Most impoprtantly we felt we were at the point where CPA should be able to accurately classify cells. We installed CPA this weekend and were absloutely thrilled with the results! (We do have a question on CPA which we will post shortly on that forum)

Thanks again for your help, and for making this great software available.

Chuck Preston
b/o Mickaela