Re: Is manual thresholding methods accepted by scientific journals?

Posted by gankaku on Dec 08, 2014; 1:39pm
URL: http://imagej.273.s1.nabble.com/Is-manual-thresholding-methods-accepted-by-scientific-journals-tp5010814p5010840.html

Dear Anders, dear all,

potentially I can also add some personal opinion to this indeed interesting
and moreover important question, since many people face the problem on
deciding for a "proper" method (however you want to define this) as I often
figure out talking to students about these topics.

Your question actually has two parts which I think need to be addressed
(thresholding and co-ocalization):

1.) Thresholding:

Generally, I would not (never) decide on a method just because you will
"get away with it" or because a specific journal would accept it. Just
because it will be accepted by a few people does still not necessarily mean
that it is a suitable or appropriate analysis method. That said, here some
considerations.

As Dimiter pointed it out already.... thresholding is not easy and
sometimes might not lead you to a satisfyingly accurate result.
Nevertheless, I had so far mostly good experiences using automatic
thresholding methods. But in many cases you will only get a good feature
extraction if you pre-process the images. This in turn might/will lead to
an alteration of feature outlines, forms and sizes. So, you first need to
figure out such processing steps and check if this would still be
acceptable regarding the features you need to extract. A possible help in
deciding for a suitable filtering and thresholding might be the "Filter
Check" as well as the "Threshold Check" (using the available auto
thresholds in ImageJ and Fiji) from the BioVoxxel Toolbox (
http://fiji.sc/BioVoxxel_Toolbox).
I further agree with Michael Schell that it is worth investing the time in
finding suitable auto-thresholds if possible, or one of the methods Dimiter
mentioned. Because you reduce user bias and improve the extraction result.

Manual thresholds are not really suitable if you try to analyze a bigger
set of data for several reasons.
In terms of comparability you need to apply the same threshold to all
images in your experiment to keep the user bias at least a little lower
(besides your decision for the initial threshold). Due to natural
variability in your images a manual threshold with one or even  two (an
upper and lower) cut-off value(s) will not work on a full set of data under
most conditions. An automatic threshold might also fail to achieve this but
since those algorithms consider the image histogram they account for those
variabilities and the chance to find suitable cut-offs is way higher.
If you manually threshold each image with different cut-off values you
actually loose comparability in your experiment completely due to massive
user bias.

Another problem, since you where talking about separating signal from
no-signal, is identifying a suitable separation of those two parts. Your
background in a intensity based fluorescent image is in most cases very
dark. Our vision unfortunately is very prone to mis-interpret different
intensities which is getting especially difficult the darker those are.
Additionally, we perceive different colors with different visual
sensitivities. Thus, it is also important if you look at a grayscale image
(with a gray LUT) or at the same image in one of different false colors.
Gray is always preferable in this context because you will see differences
best.

Nevertheless, manual thresholding is very subjective and should be avoided
whenever possible. Even without those it is already difficult enough to
achieve objective analyses for many studies (my opinion).


2.) Co-localization
If I got it correctly, your initial aim is a co-localization study. In this
context, I would not rely only on a pixel based overlap determination. This
might give you a hint and is partially used during object-based
co-localization studies. Nevertheless, you should consider additional
parameters like resolution limit and your actual image resolution and the
intensity distribution in your images/features. Therefore, I would pay
attention first to a proper imaging setup, with a good nyquist-sampled
image and usage of the full dynamic range (besides other parameters). To
this end the following paper might be helpful:

Jennifer C. Waters, J Cell Biol. Jun 29, 2009; 185(7): 1135-1148. Accuracy
and precision in quantitative fluorescence microscopy.

Analysis wise there are two very excellent tools available in Fiji which is
the Coloc2 and the JACoP. The latter implements two object-based methods
including a thresholding. In this context it might also be a possibility to
use binary images as result of an auto-thresholding and mask your original
images with them (e.g. with >Edit >Paste Control or the >Process > Image
Calculator). This is similar in applying a ROI as possible in the Coloc2
plugin.
As a suggestion for co-localization studies... I would not rely on a single
output method only, but rather combine several suitable ones as is possible
in Coloc2 and JACoP to get a better confidentiality about a potential
co-localization.

So, to not create more confusion here some interesting reading and
important literature regarding co-localization:

There was an interesting discussion about different co-localization
analyses on the list a few month ago with some papers suggested already (
https://list.nih.gov/cgi-bin/wa.exe?A2=ind1403&L=IMAGEJ&P=R31566&1=IMAGEJ&9=A&I=-3&J=on&d=No+Match%3BMatch%3BMatches&z=4
)

Recommendable, especially if you use the JACoP tool: Bolte and Cordelieres,
J Microsc. 2006 Dec;224(Pt 3):213-32. A guided tour into subcellular
colocalization analysis in light microscopy

Furthermore: Dunn et al. Am J Physiol Cell Physiol. 2011
Apr;300(4):C723-42.  A practical guide to evaluating colocalization in
biological microscopy

There are many more papers regarding the topic e.g. from Elise Stanley's
lab, Ingela Parmryd and Jeremy Adler. And many more I might not be aware of.

So, in the worst case I misunderstood your question and overloaded you with
unnecessary answers (but they might be helpful to others). In the best case
you have a lot of reading suggestions and potentially a clearer picture on
how you want to start your analysis.

Kind regards,
Jan



2014-12-06 16:00 GMT+01:00 Anders Lunde <[hidden email]>:

> Dear mailing list,
>
> I have developed a nice macro for identifying colocalized signals for
> z-stack confocal images with multiple channels/colors. However, my
> advisor/professor has now come to question my method for setting a
> threshold for signal/no-signal in the infividual channels.
>
> My manual method has been to simply raise the threshold above what I
> relatively confidently can see is background, like large areas with no
> apparent staining. The reason I did it manually is because when I played
> around with the automatic thresholding methods in ImageJ I decided that
> they were not any better than manual and could be subject to mistakes.
>
> My supervisor now feels that this sounds too subjective and would not look
> good in a paper. He therefore asked me to try to find a way that was more
> guided e.g. by the histogram or something, anything that is less subjective
> (not sure if he is worried about accuracy or how it sounds in a paper).
>
> What is the current standard for this kind of analysis in scientific
> journals, in particular with regards to the acceptability of manual
> thresholding of immunofluorescent brain sections stained with various
> antibodies (and nuclear markers and neuron trancers)? Is there a preference
> for automated, manual or some hybrid methods? Could I "get-away" with
> something like this:  "Thresholds were set manually at a level that
> excluded most pixels in assumed background areas. Inspection of the
> assigned threshold level in the ImageJ intensity histogram showed that the
> thresholds were set at where the main peak (background pixels) started to
> or had reached a minimum value."
>
> Image set that Im working on:
>
> I am working with images of brain sections with 4 colors/channels: nuclear
> stain, two immunofluorescence staining for transciption factors (nuclear
> localization), and a retrograde nerve cell staning (nuclear + cytoplasm
> staining).
>
> Greateful for any advice!
>
> --
> 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