Threshold methods, how to choose suitable one?

[Sarah Mohamed]

Hi,
I would like to ask about the suitable choice for threshold method to quantify the area fraction of brown color.
I have H/DAB images for treated groups. I need to use a method that is least subjective, gives consistent results meaning if i use it several times gives the exact same area fraction.
There are several methods:
1, Image>>adjust>>threshold, moving brightness and hue sliders to get 2 areas then divide them to get area fraction.
2. Image>>color>>split. choose green chanel. Image>>adjust>>threshold>>clt+M
3.www.med.upenn.edu/cellbio/documents/imagej_colorsegmentpdf.pdf
Image.. Plugins..threshold color...move saturation slider to select ROI..close plugin..image..type...8bits..image..adjust...threshold..clt+M.
I know there may be other methods for color segmentation but thats all i know.
I cant use color deconvolution as i cant do background correction calculation ...i am just a pharmacologist.
Thanks in advance.
Sarah M.Mosaad, M.Sc.
--
Sent from myMail app for Android

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

[gankaku]

Hi Sarah,

Some comments to extraction methods possible in your case:

1) You first mentioned something about uneven illumination if I am not
mistaken. In respect to that, your images actually look fine to me. I do
not notice any drastic uneven illumination which would influence your
analysis.
Nevertheless, in the case you wnt to chech out the following....? You could
for example use the Pseudo-flat-field correction (from the BioVoxxel
Toolbox; http://fiji.sc/BioVoxxel_Toolbox). This basically implements one
of the methods described by Gabriel Landini among other methods on the wiki
page (
http://imagejdocu.tudor.lu/doku.php?id=howto:working:how_to_correct_background_illumination_in_brightfield_microscopy)
and in this paper [1]. In this case you need to use (as described) a big
value for blurring the image content (using a big gaussian convolution
kernel). I tried it and did not see a marked difference in the outcome for
your image analysis.

The extraction:

2) You mentioned already splitting it into R-G-B channels. This is a good
starting point but in my experience this seldomly leads to a satisfying
result. I would try to convert the image into different color spaces. This
you can achieve in Fiji by e.g. >Image >Color >RGBtoCIELAB or >Image
>Adjust >HSB Stack. There are more available but mostly those do the job
already. Then you will see the image also separated into 3 channels which
are actually the basis for what you would manually do while using the color
thresholding dialog.

2a) I would avoid using manual thresholding whenever possible! There are
automatic thresholds (available in ImageJ/Fiji) implemented by Gabriel
Landini, which will give you a non-subjective reproducible outcome.
Nevertheless, subjectiveness is introduced since you initially need to
decide for one of those methods which in your eyes performs best with
different repressentative test images from your series.
If I find the time in the following weeks, I will also post a more detailed
information on the Imagej wiki page regarding those binarization methods.
These thresholds work only on grayscale images, this means you first need
to convert your images in a manner suitable for your analysis. This you do
with the described color space conversion.

2b) In those image stacks I would check if any of the channels depicts your
staining of interest in a more pronounced way allowing a better
differentiation from the rest of the image. On this channel I would then
apply an automatic threshold. Sometimes further post-processing might be
necessary to improve the outcome, but in your case it works already fine
like this. To check which threshold does a better job than another one, you
might want to try the "Threshold Check" [2] from the toolbox (
http://fiji.sc/BioVoxxel_Toolbox#Threshold_Check)

2c) Once you figured out a suitable threshold by testing different images
(best with some variance in staining to be sure that the threshold you
choose will also reliably recognize your feature of interest under
different conditions). Then you run >Image >Adjust >Auto Threshold (in
rather special cases potentially also Auto Local Threshold). This results
in a binary image (only black and white pixels) which you can use to
determine the area fraction (%Area) by measuring the image or a ROI.
Obviously, your measurement is influenced by the choice of the field of
view during your imaging which introduces again some bias. Imaging many
field of views per sample might therefore be recommendable.

Below you find a short macro which basically does what I just described. I
figured out that the auto threshold "Default" or "IsoData" (from the >Auto
Threshold function in ImageJ) do this fairly good for both of your images.
This will give you reproducible measurements but once again, needs to be
checkt if the outcome reflects the reality of your staining when applied to
variable staining qualities which might occur during one experimental
series.

//--------------------------------
run("RGB to CIELAB");
run("8-bit");
run("Delete Slice");
run("Delete Slice");
run("Auto Threshold", "method=Default white");  // you might want to
exchange the thresholding method here if necessary
run("Set Measurements...", "area_fraction display redirect=None decimal=3");
run("Measure");
//--------------------------------

3) another possibility would be to use machine learning based tools like
SIOX (http://fiji.sc/SIOX). This is very intuitive and after good training
also gives you good results. This might be applicaple for more tricky
images. Additionally, there is the Trainable WEKA Segmentation (
http://fiji.sc/Trainable_Weka_Segmentation).


Hope this helps!
Kind regards,
Jan



[1] (Leong et al. Correction of uneven illumination (vignetting) in digital
microscopy images. Journal of Clinical Pathology 2003;56:619-621)
[2] Qualitative and Quantitative Evaluation of Two New Histogram Limiting
Binarization Algorithms. Brocher, 2014, International Journal of Image
Processing - IJIP 8(2), pp30-48.




2015-01-15 10:07 GMT+01:00 Sarah Mosaad <
[hidden email]>:

>
>
> Hi,
> I would like to ask about the suitable choice for threshold method to
> quantify the area fraction of brown color.
> I have H/DAB images for treated groups. I need to use a method that is
> least subjective, gives consistent results meaning if i use it several
> times gives the exact same area fraction.
> There are several methods:
> 1, Image>>adjust>>threshold, moving brightness and hue sliders to get 2
> areas then divide them to get area fraction.
> 2. Image>>color>>split. choose green chanel.
> Image>>adjust>>threshold>>clt+M
> 3.www.med.upenn.edu/cellbio/documents/imagej_colorsegmentpdf.pdf
> Image.. Plugins..threshold color...move saturation slider to select
> ROI..close plugin..image..type...8bits..image..adjust...threshold..clt+M.
> I know there may be other methods for color segmentation but thats all i
> know.
> I cant use color deconvolution as i cant do background correction
> calculation ...i am just a pharmacologist.
> Thanks in advance.
> Sarah M.Mosaad, M.Sc.
> --
> Sent from myMail app for Android
>
> --
> 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

[Sarah Mohamed]

Hi Cameron,
Thanks for your reply.

In order to do color deconvolution i must do background correction & enter my own vectors; my knowledge in imaging is little, i find difficulty to go through all these steps (bright field, dark field calculation,..etc).

so  i am trying to find another segmentation method that is consistent and not subjective.

Regards,
Sarah M. Mosaad, M.Sc.

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