Selecting correct threshold algorithm

[tmurp002]

Hi all,

Hopefully this question hasn't been answered before (I've searched back quite a bit into the archives and found nothing). I'm performing a set of experiments looking at cell swelling in brain slices under certain conditions over short intervals of time. To visualize this I'm using patch-clamp to load the cells with a membrane-impermeant fluorescent dye and imaging via CSLM. My standard protocol involves taking rapid z-stacks through the soma (to capture all edges of the cell body), then using max intensity projections through these stacks to compare the soma area between time points. Through a lot of trial and error I've come to the strategy of using Fiji to filter, align, crop, and threshold the images for each cell, and perform measurements on the thresholded area.

My problem is essentially figuring out how to pick the appropriate thresholding algorithm. I've basically settled on the Mean threshold method, because it is the only one which shows the same features I observe in the original image. It also would stand to reason (at least to me) that this method would appropriately account for intensity fluctuations in the soma (for example, lowering the threshold if average intensity drops, as is often the case when the cell takes on water).

Am I thinking about this correctly, or is there some large issue I'm overlooking which invalidates this method?

[gankaku]

Hi "tmurp002",

First to your method. If the analysis in 2D seems appropriate or the only
possibility (which is feasible) then the MAX-projection is most likely the
best idea. Problem in such cases is that you neglect the 3-dimensionality
of your objects. Another possibility while filling the cells completely
with fluorescent dye could be the use of the 3D manager from Thomas Boudier
(http://imagejdocu.tudor.lu/doku.php?id=plugin:stacks:3d_roi_manager:start)
and an analysis in 3D space. A problem which often arises here is that
cells in close proximity which might visually fuse might hinder an
individual cell analysis.

Having said this back to the 2D analysis:

Finding "the correct" thresholding/binarization method is limited by the
fact that something needs to define the best result and you would need to
be able to evaluate this objectively. Since perfect ground truths either
don't exist or are also subjectively created by humans (example
thresholding) you will always suffer from subjectivity.

Figuring out a "good/suitable" auto thresholding algorithm suitable for
your images might be facilitated by applying the "Threshold Check" from the
BioVoxxel Toolbox (http://fiji.sc/BioVoxxel_Toolbox#Threshold_Check for
explanations on usage). This tool offers you a comparison of all auto
thresholding methods (by Gabriel Landini) present in Fiji. You will have
the possibility to visually compare those for one example image according
to a color coded system as well as get a relative quality quantitation of
all thresholds according to a user given reference intensity.
This makes the quest for the "optimal" threshold somehow easier, especially
when the results are fairly similar.
For further information I recommend readinf the publication "Qualitative
and Quantitative Evaluation of Two New Histogram Limiting Binarization
Algorithms<http://www.cscjournals.org/csc/manuscript/Journals/IJIP/volume8/Issue2/IJIP-829.pdf>"
(Brocher, IJIP 2014).
In this publication you will also find two new binarization algorithms
(MoLiM and DiLiM) based on the histograms' mean value which account for a
certain ammount of variation in the image and might also help you in
extracting your cell projections. These algorithms are also available from
Fiji's BioVoxxel update site (
http://fiji.sc/BioVoxxel_Toolbox#Mode_and_Differential_Limited_Mean_Binarization)
and are installed in the folder >Plugins >Filters.

Hope this helps you in your decision process for a suitable method.

regards,
Jan


2014-04-18 19:49 GMT+02:00 tmurp002 <[hidden email]>:

> Hi all,
>
> Hopefully this question hasn't been answered before (I've searched back
> quite a bit into the archives and found nothing). I'm performing a set of
> experiments looking at cell swelling in brain slices under certain
> conditions over short intervals of time. To visualize this I'm using
> patch-clamp to load the cells with a membrane-impermeant fluorescent dye
> and
> imaging via CSLM. My standard protocol involves taking rapid z-stacks
> through the soma (to capture all edges of the cell body), then using max
> intensity projections through these stacks to compare the soma area between
> time points. Through a lot of trial and error I've come to the strategy of
> using Fiji to filter, align, crop, and threshold the images for each cell,
> and perform measurements on the thresholded area.
>
> My problem is essentially figuring out how to pick the appropriate
> thresholding algorithm. I've basically settled on the Mean threshold
> method,
> because it is the only one which shows the same features I observe in the
> original image. It also would stand to reason (at least to me) that this
> method would appropriately account for intensity fluctuations in the soma
> (for example, lowering the threshold if average intensity drops, as is
> often
> the case when the cell takes on water).
>
> Am I thinking about this correctly, or is there some large issue I'm
> overlooking which invalidates this method?
>
>
>
>
> --
> View this message in context:
> http://imagej.1557.x6.nabble.com/Selecting-correct-threshold-algorithm-tp5007356.html
> Sent from the ImageJ mailing list archive at Nabble.com.
>
> --
> ImageJ mailing list: http://imagej.nih.gov/ij/list.html
>

--

CEO: Dr. rer. nat. Jan Brocher
phone:  [phone]
mobile: [phone]
e-mail: [hidden email]
info: [hidden email]
inquiries: [hidden email]
web: www.biovoxxel.de

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html

[tmurp002]

Thank you gankaku, those BioVoxxel tools really helped. Threshold check seems to confirm that the mean threshold is a good way to go. MoLim and DiLim do look very accurate, and in fact give similar/nearly identical results to the mean algorithm (which makes sense) but require more effort to use since they don't work on stacks, and I'm a bit wary of changing my images to 8-bit prior to thresholding. So I think I'll stick with the mean threshold for now.

As to your first point, I've considered 3D analysis (and I'll definitely check out that link), although it has proven too complicated to implement thus far. Perhaps I was simply using the wrong tools before.

Thanks again!