Follicle segmentation in H&E image

[Pascal Lorentz-2]

Dear list

I got an H&E image from a user on which she wants to count follicle
number. I don't know how to properly segment them.

http://picfront.de/d/8sa3
http://picfront.de/d/8sa4

The first links is the original image and the second link guides you to
the image where I have roughly selected the follicles that we want to
count. (6 of them in this image)
Do you have an idea how this is easily done. We want to find a process
that we can then use in a macro for batch processing.

I already tried color deconvolution followed by trainable segmentation.
It somehow works but does not only recognise the follicles. It seems a
bit complicated and is not very convenient since it also leads to false
positive results.

If you have know any tricks to solve that problem I would be more than
happy if you share them with me.

Best regards

Pascal

Pascal Lorentz
BioOptics Facility
Department of Biomedicine
University of Basel
Mattenstrasse 28
4058 Basel
Switzerland

[Gabriel Landini]

On Friday 23 Dec 2011 13:03:46 you wrote:
> I got an H&E image from a user on which she wants to count follicle
> number. I don't know how to properly segment them.

You could try this: use colour deconvolution, take the H channel and then use
a large median filter (32 or so). Then threshold this image with a method that
isolates the dark blobs, binarise it, analyse the plarticles and retain only
those that are near circular and have a size in the range expected.

It might help if the images are corrected for background illumination (they
appear to be brighter in the centre). The IJ wiki has an article on how to do
this.

Regards

Gabriel

[Daniel Hoeppner]

Hi Pascal,

This macro works for your example image.  You will have to adjust the settings in analyze particles for batch processing.

selectWindow("2014879.png");
run("Split Channels");
imageCalculator("Subtract create", "2014879.png (blue)","2014879.png (red)");
selectWindow("Result of 2014879.png (blue)");
run("Gaussian Blur...", "sigma=10");
setAutoThreshold("Default dark");
//run("Threshold...");
setAutoThreshold("Default");
setThreshold(27, 255);
run("Convert to Mask");
run("Analyze Particles...", "size=3000-Infinity circularity=0.50-1.00 show=Nothing display exclude clear summarize add");

Dan

Daniel Hoeppner, PhD
Investigator
The Lieber Institute for Brain Development
855 North Wolfe Street, 3rd Floor
Baltimore Maryland 21205
T: 410-955-0984
E: [hidden email] 
http://libd.org



-----Original Message-----
From: ImageJ Interest Group [mailto:[hidden email]] On Behalf Of Pascal Lorentz
Sent: Friday, December 23, 2011 8:04 AM
To: [hidden email]
Subject: Follicle segmentation in H&E image

Dear list

I got an H&E image from a user on which she wants to count follicle number. I don't know how to properly segment them.

http://picfront.de/d/8sa3
http://picfront.de/d/8sa4

The first links is the original image and the second link guides you to the image where I have roughly selected the follicles that we want to count. (6 of them in this image) Do you have an idea how this is easily done. We want to find a process that we can then use in a macro for batch processing.

I already tried color deconvolution followed by trainable segmentation.
It somehow works but does not only recognise the follicles. It seems a bit complicated and is not very convenient since it also leads to false positive results.

If you have know any tricks to solve that problem I would be more than happy if you share them with me.

Best regards

Pascal

Pascal Lorentz
BioOptics Facility
Department of Biomedicine
University of Basel
Mattenstrasse 28
4058 Basel
Switzerland

[Arne Seitz]

Hi Pascal,

I managed to get the follicles segmented by using the colour segmentation plugin from D. Sage
http://bigwww.epfl.ch/sage/soft/colorsegmentation/

First I did a 6px mean filtering. Then I selected various points inside and outside the follicles. After running the colour segmentation (Hidden Markov Model), I separated the two channels with the "Color Threshold...". Via particle analysis you can count the follicles (excluding non-interesting regions via area and circularity thresholds).

I  hope that it works with the other images as well.

Have a nice day.

Cheers Arne

> -----Original Message-----
> From: ImageJ Interest Group [mailto:[hidden email]] On Behalf Of
> Pascal Lorentz
> Sent: vendredi 23 décembre 2011 14:04
> To: [hidden email]
> Subject: Follicle segmentation in H&E image
>
> Dear list
>
> I got an H&E image from a user on which she wants to count follicle number. I
> don't know how to properly segment them.
>
> http://picfront.de/d/8sa3
> http://picfront.de/d/8sa4
>
> The first links is the original image and the second link guides you to the
> image where I have roughly selected the follicles that we want to count. (6 of
> them in this image) Do you have an idea how this is easily done. We want to
> find a process that we can then use in a macro for batch processing.
>
> I already tried color deconvolution followed by trainable segmentation.
> It somehow works but does not only recognise the follicles. It seems a bit
> complicated and is not very convenient since it also leads to false positive
> results.
>
> If you have know any tricks to solve that problem I would be more than
> happy if you share them with me.
>
> Best regards
>
> Pascal
>
> Pascal Lorentz
> BioOptics Facility
> Department of Biomedicine
> University of Basel
> Mattenstrasse 28
> 4058 Basel
> Switzerland

[Pascal Lorentz-2]

Dear Gabriel, dear Daniel and dear Arne

Sorry for my late response. It took me some time to test your
recommended solutions.
They all worked pretty well and I am really thankful for your help.
We will now further test the methods with other images and see which
method performs best.

Thanks again

Pascal


Am 23.12.2011 15:00, schrieb Gabriel Landini:

> On Friday 23 Dec 2011 13:03:46 you wrote:
>> I got an H&E image from a user on which she wants to count follicle
>> number. I don't know how to properly segment them.
> You could try this: use colour deconvolution, take the H channel and then use
> a large median filter (32 or so). Then threshold this image with a method that
> isolates the dark blobs, binarise it, analyse the plarticles and retain only
> those that are near circular and have a size in the range expected.
>
> It might help if the images are corrected for background illumination (they
> appear to be brighter in the centre). The IJ wiki has an article on how to do
> this.
>
> Regards
>
> Gabriel