Re: question regarding Colocalization Threshold in Fiji - Coloc2

Posted by Daniel James White on Feb 17, 2011; 12:54pm
URL: http://imagej.273.s1.nabble.com/Re-question-regarding-Colocalization-Threshold-in-Fiji-Coloc2-tp3683647.html

Dear Judith,

On Feb 16, 2011, at 11:27 PM, Judith Lacoste, Ph.D. wrote:

> Hi,
>
> First, all my congratulations for FIji, I enjoy it a lot and recommend it all the time.  

We are very happy to hear that.
The Fiji project is very much about people like you - people that use it, and spread the word to others.
We all really appreciate the active members of the Fiji community.

> I have been comparing the Fiji colocalization Threshold and the Jacop plugin (v.2) and I am a bit puzzled now.

This is not surprising, sadly.

>  For the same images, the automatic threshold results are fairly similar.  

Indeed the methods used are basically the same, but the implementations do differ slightly, leading to slightly different thresholds in some cases.
This is a problem, as it casts doubt over the method that should not be in doubt actually, since it is pretty robust if you feed it sensible input data.

I strongly suggest you read:
http://pacific.mpi-cbg.de/wiki/index.php/Colocalization_Analysis
so you can understand the pitfalls and make sure you are doing the "right thing"

> However, the tM1 and tM2 values are 10 fold lower than the ones calculated with the Jacop plugin.  

That could be because the thresholds are very slightly different.
Depending on the kind of data that the images contain... how strong or weak any correlation is,
and if there is an uncorrected camera offset or pmt offset, and lots of noise,
a small difference in thresholds can easily make a very large difference in the values of tM1 and tM2

> I did see the thread in the listserv last june about a bug to be corrected and I guess it's been done since then.  Any clues?  Many thanks for your help!

We corrected a bug in the Colocalization Threshold plugin
that was pointed out by Andrew Barlow from Perkin Elmer (Volocity)
where the value of the Pearsons r for above the thresholds was calculated incorrectly.

There has been much discussion about wether Pearsons r above the auto calculated "Costes" thresholds is actually a sensible
and interpretable statistic, since Pearsons r is used to calculate where the threshold should be....

see
Microscopy and Microanalysis (2010), 16: 710-724
Colocalization Analysis in Fluorescence Micrographs: Verification of a More Accurate Calculation of Pearson's Correlation Coefficient

This suggest a slightly different approach to get around this problem and improve the "meaningfulness" of the Pearson's r
that does not use Pearson's r to calculate the thresholds in the first place...
but then I still wonder exactly how one is supposed to then set the thresholds in a robust, reproducible, objective manner.

From that paper, here is a quote from the discussion:
"Our method of calculating thresholded PCC is not the same as the PCC of pixels above the thresholds set by applications that implement the approach of Costes et al. (2004) such as JACoP. Software implementing the Costes' method often display Pearson's correlation coefficient for two subsets of pixels, those with intensities above the thresholds and those with intensities beneath (this differs from the original implementation of Costes et al. in which only PCC for the subset of pixels fainter than either threshold was calculated). This is done to reassure users that the thresholds have been set in positions that separate positively correlated intensities from uncorrelated intensities, and to set thresholds that give objective measurements of Costes' variants of M1 and M2. Given that PCC is used to determine these thresholds, the same thresholds cannot then be used to determine thresholded PCC objectively. Our approach requires thresholds that separate signal from background to be set, and then thresholded PCC, Mx, and My are generated using all pixels above both thresholds. While the values calculated for pixels above the automatically set thresholds are the same when calculated by JACoP and Volocity 5.3, the method of setting the threshold for the calculation of thresholded PCC must differ from the approach of Costes et al. in order to make an objective measurement of thresholded PCC. Additionally, the algorithm of Costes et al. (2004) is designed to set thresholds that separate positively correlated pixels from uncorrelated or negatively correlated pixels. Thresholded PCC, which requires thresholds that separate signal from background, is effective at describing uncorrelated and negatively correlated datasets."

It seems there is still much theoretical work to be done on this problem in order to find more robust solutions.

Further, the differences between the JACOP and Colocalisation  Threshold plugin's implementation of Costes' method for autothresholding
generates the confusion and uncertainty that you are currently experiencing. Welcome to my world.

In order to take the next step forward in dealing with these problems,
we have made a new Colocalization plugin - using new technology for storing and processing image data (imglib)
and a modular design with test driven development (industry best practice software development approach)

We took the different algorithms, and make them self sufficient modules of code that can be reused
and work on any bit depth of image data.
We built tests to make sure that the algorithms do the right thing when we make changes to the code.
We made a prototype graphical user interface for driving the different algorithms.
We started to design a standardized output style for coloc analysis,
with a standardized way of selecting which numbers to show and which scatterplots and images to show,
and putting it all in a PDF file output.
This way everyone doing the same analysis get easily comparable results,
that are already formatted and ready to go into supplemental info of a paper.

All this is to standardize the implementation of the maths of the algorithms, so everyone gets the same answer,
and re ove the chaos of different folks nor understanding what each other mean by colocalisation,
and arguing about what numbers and pictures to show and how.

This prototype is working and usable and hopefully industrial strenght (as opposed to little or ethan a hacked and buggy script) ,
and we called it Coloc2 ... while we find a better name.

It does :
Peasrsons r
Costes Autothreshold
Costes Significance Test (block randomisation, (not just white noise) and Peasrons)
Manders coefficients tM1 and tM2
Li Scatterplots and single global value.
2D histogram / Scatterplot / cytofluorogram - plus regression line
Region of interest and Masks (3D)
Checking image for suitability for analysis\
        detect problems that cause numerical problems for the algorithms
        detect regression  line intercept that is not close to zero.
and more.

It will be published at some point, but first we will ask the community for help in refining the graphical user interface and PDF output style
and also which algorithms and methods to use and exactly how.
The Coloc2 code should be easy to add new things/modules to, without breaking other bits of it (modular design - test driven development)

Soon it will be in the Fiji updater. I will announce it, and ask for feedback from general users
and from the authors of existing coloc algorithms and plugins.

You can see design ideas and discussion here:
https://docs.google.com/View?id=df66rgc7_2dtqkv3dx
If you want write access to that doc, just let me know your google account ID.
This is the place where we can argue and discuss about the right way to move forward.

The aim is to standardize and make more robust the whole game of colocalization analysis,
so people can speak the same language about it,
and get comparable results.
Then they can focus on the biology, instead of arguing about the details of the plugins and different methods etc.

Hope that stimulates some discussion....

happy coloc analysis!

Dan

>
> Best regards,
>
> Judith
>
>    
>
>
> --
> Judith Lacoste, Ph.D.
> [hidden email]
> Cell.: 514.916.4674
>
> MIA CELLAVIE Inc.
> PO Box 192, STN Anjou
> Montréal QC H1K 4G6
> Tel.: [phone]
> Fax: [phone]
>        
>         <smallLogo.tif>
> http://miacellavie.com
>

Dr. Daniel James White BSc. (Hons.) PhD
Senior Microscopist / Image Visualisation, Processing and Analysis
Light Microscopy and Image Processing Facilities
Max Planck Institute of Molecular Cell Biology and Genetics
Pfotenhauerstrasse 108
01307 DRESDEN
Germany

[phone] (German Mobile)
[phone] (Work phone at MPI-CBG)
[phone] (Fax MPI-CBG LMF)

http://www.bioimagexd.net  BioImageXD
http://pacific.mpi-cbg.de                Fiji -  is just ImageJ (Batteries Included)
http://www.chalkie.org.uk                Dan's Homepages
https://ifn.mpi-cbg.de  Dresden Imaging Facility Network
dan (at) chalkie.org.uk
( white (at) mpi-cbg.de )