White Balance / Background has infuence on OD measurement in Color Deconvolution?

[David]

Hi everyone,

does somebody know, whether the white balance / background intensity has influence on the result of OD measurements in one channel of the color deconvolution?

If yes and if I want to precisely measure the OD of an IHC Stain, should I measure the background intensity for each image to get propper results? Or is the mistake negligible?

Best
David

[Pete]

In principle it does / should.  However, as best I recall, within QuPath's primary Cell detection command the background values are always treated as 255 - regardless of whether these values have been set otherwise.

There may be some small improvement in performance if the Cell detection were changed to include the background values in its calculations.  This change might be introduced in the future, but I am reluctant to make any small modifications such as this to the existing cell detection command, since I don't want to do anything to impact backwards compatibility unnecessarily.  A future cell detection command, if it uses color deconvolution, would probably use the background values too - it depends how it is implemented.

[micros...@gmail.com]

I forget, does the same apply to the "Add intensity features?"

[Pete]

As far as I recall, the background values are used by Add intensity features.

[micros...@gmail.com]

Sounds like that might be David's best bet then.  

And as far as how important the background is, it probably depends most on how the image is being taken.  I have seen some backgrounds that were in the 190s, which I am sure had an impact on the overall OD measurements of a given sample.  Does it matter as long as all of your samples have the same background?  Not sure. Plus if your slide scanner is performing different adjustments per sample when it is auto adjusting the exposure/white balance (so that the background is the same), that could affect your stain OD measurements as well.

OD measurements are already non-linear and not terribly quantitative due to physical properties of most dyes (I think), so I would recommend being careful how you use them in any case.

[David]

ok good I asked. Thank you both for your answers and discussion. I will do some tests whether the OD result differs significantly, if the background values are measured for each image separately.
Since I want to get a rough estimation on different protein concentrations, a 10% difference in OD might probably be already an important result and could be strongly influenced by the background value.

[David]

I checked with the cell detection command and used different values for the background. The results vom "Make measurements" are allways exactly the same! So yes, probably cell detection uses the 255 255 255.

Then I did cell detection without "make measurements" and created statistics by "add intensity features" and added H and DAB and the four values MEAN, MIN MAX, Stdev, Median.

Also there I could not discover any impact on the OD results with different background values.

So there are two possiblities?
 1) The intensity features also use 255 255 255
 2) or the Background values does not have an impact on the calculation of the OD values in the Color deconvolution Stain 1 and Stain 2 results?

[Pete]

From looking at the code, the answer is 1).

I intend to change this for the next 'major' update, to improve accuracy.  But it ought to remain the same for any 'minor' update, for backwards compatibility.

[Pete]

In my more 'experimental', development branch of QuPath I've just simplified and unified much of the color deconvolution code (commit here).

As part of this, the white values are now used more consistently.  If this behaves itself whenever I start using and testing it more thoroughly, the changes will be in a later release.

[David]

Hi Pete,

thanks for looking it up.
Do you have a rough opinion about how big the "mistake" might be in DAB OD, if the real background value is varying between between 214 214 214 and 222 222 222  compared to a fixed value of 255 255 255?

[Pete]

From what I understand of DAB, the 'mistake' is to try to use it quantitatively in the first place :)

I have never once seen a suggestion that this is a good idea, and I have seen a lot of strong warnings against it.  This is the one I often refer to, but there are certainly others.

I don't feel I can/should give a more direct answer or opinion as it isn't really my area.  Personally, I have not worked on any projects with brightfield images (using any stains) where there was a need to do more than look at areas, or counts using broad categorisations (positive/negative, or sometimes 0/1+/2+/3+).

[David]

Thanks for the references, Pete!
Yes true. DAB is not a true absorber. About the other point that Antigen-antibody reactions are not stoichiometric is even new to me.

Anyhow, I was in the believe, that DAB is at least quite linear as long as it is not to high concentrated. And it is not about absolute protein concentration but only on relative comparison of epitope concentration.

So I was believing that the DAB OD contains a qualitative information on epitope concentration - at least as soon as it is evaluated accross many cells and many different images.

[micros...@gmail.com]

Yeah, like I mentioned, I would be a little nervous about relying too much on pure OD data.  Or if you really do want to try to use the OD semi-quantitatively, try and have a back-up method to try and correlate stain intensity with protein concentration.  The following paper suggested that it is at least possible in some cases.  Side note, I have not read the paper exhaustively!

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5318955/

Emphasis on this in the discussion:

"We apply our experimental framework in the IHC context to evaluate different image normalisation methods based on blind colour deconvolution. Our experimental analysis reveals that efficient colour deconvolution is an essential step which has to be applied on each staining batch. This means that colour vectors extracted from an IHC slide series cannot be applied as is to another series, even between successive IHC batches made in the same and well-controlled environment (such as illustrated by our results labelled “before”). Consequently, any IHC quantification tool based on colour deconvolution, provided in commercial or open-source software packages, must include such an adaptive step. This also means that the widespread use of the so-called “Ruifrok” method (i.e. the use of a single set of values, such as those given in ref. ) cannot be satisfactory without adaptation to each slide series left to analyse."