Tumor grading in GUI or ...?

I have micro tumor images(9) and there are areas of dlbcl tumor (Nd DAB-stained and Nh H-stained) Proliferation index for this is a solution of the equation PIc=Nd/(Nh+Nd). I have to find PIc for all pictures knowing RGB intervals for Nh= (<45, 180>, <50, 185>, <160, 215>) and for Nd= (<40, 115>, <6,80>, <10,75>). How AM i supposed to do it? Please help!!

[Jan]

The question contains a lot of details, which do not matter the problem. As far as I understand you have RGB images and want to find regions in RGB intervals? Then what about:

img   = rand(64, 480, 3);  % Test data
match = (img(:, :, 1) >  45 & img(:, :, 1) < 180) & ...
        (img(:, :, 2) >  50 & img(:, :, 2) < 185) & ...
        (img(:, :, 3) > 160 & img(:, :, 3) < 215);    

Looks straight.

But now "PIc=Nd/(Nh+Nd)" is not clear. What do you have to calculate exactly? How is Nh and Nd defined? Remember that your notation "Nh= (<45, 180>, <50, 185>, <160, 215>)" is not explained anywhere but a new invention. My above idea is a pure guess only.

[Aneta Chwala]

Sorry i made it unclear. This nine pictures are specimens and there are dab-stained nuclei (Nd) and H-stained nuclei (Nh) and for them I have to calculate index proliferation PIc. Those intervals are for Nh nuclei ( 45<r<180, <50<g<185, <160<b<215) and for Nd nuclei (40<r<115, 6<g<80, 10<b<75). As intervals are given i need to count the number of Nd and Nh pixels from the pictures.

thank you very much!

[Image Analyst]

And the equations in my answer below don't do it? Please explain why not. And attach your image.

[Aneta Chwala]

Image Analyst I am sorry I am new to MATLAB and I don't really know. But since the Jan's comment isn't ok (?) then how to modify it to your code and what it actually means? It's confusing for me. I didn't get images but they look similar to this:

Answer 1

[Image Analyst]

Use the code in Jan's comment above, but modify it to get two binary images: matchNh and matchNd. Then I think you want the sum (count) and ratio

Nd = nnz(matchNd)
Nh = nnz(matchNh)
Plc = Nd / (Nh + Nd)

Answer 2

[Image Analyst]

Try this:

K-means segmentation

[Aneta Chwala]

Thanks! But I actually can't clustering, also i need to use rgb intervals for my code. I think that your idea was better but i have no idea how to start with that.

[Image Analyst]

Why can't you use clustering??? That's strange.

Have you tried the Color Thresholder on the Apps tab?

As you can see from the RGB gamut below, it will be virtually impossible in rgb color space.

You'd be better off in HSV color space.

[Aneta Chwala]

I need to compute PIc, so i need the number of Nd and Nh pixels from rgb. So I have to find a code which will check a photo then based on rgb intervals gave Nc Nh and PIc. I guess lol. How is clustering helping here idk. Also I don't need new images occur.

sorry if I confused more now.

And thank you for helping me!!

[Image Analyst]

Look at the gamuts - they're essentially a 3-D histogram where there is one point for every (r,g,b) triplet (every pixel value/color).

Point 1: note how there are no well contained clusters - it's basically a continuum - so the dividing plane will essentially be sort of arbitrary. You could almost put it wherever you want but at least kmeans has some sort of logic behind where it splits the gamut into the 2 colors.

Point 2: thresholding essentially divides up the gamut by slicing the color could with planes aligned with the axes, and you can see that there are no planes perpendicular to the R, G, and B axes that will do even a half way decent job of dividing the image up into different colors that are meaningfull. With HSV color space, your dividing surfaces are basically cones, sectors, and planes perpendicular to the V axis. By doing that, you can definitely carve out different regions because the hue, saturation, and brightness are different between the pink, purple, and white. The R, G, and B are also different, but you can see that planes perpendicular to the axes will not carve that cloud up into pink, purple, and white sub-clouds.

Answer 3

[Aneta Chwala]

I decided to use the code below but i get wrong PIc from it. Why is this happening?

Let say i have a picture cat.tif:

%

img = imread('cat.tif');

matchNh = (img(:, :, 1) > 45 & img(:, :, 1) < 180)

(img(:, :, 2) > 50 & img(:, :, 2) < 185)

(img(:, :, 3) > 160 & img(:, :, 3) < 215);

matchNd = (img(:, :, 1) > 40 & img(:, :, 1) < 115)

(img(:, :, 2) > 6 & img(:, :, 2) < 80)

(img(:, :, 3) > 10 & img(:, :, 3) < 75);

Nd = nnz(matchNd);

Nh = nnz(matchNh);

PIc = Nd / (Nh + Nd);

imagesc(img);

title(PIc);

%

[Image Analyst]

You forgot to attach the image.

Why do you say it's wrong?

[Aneta Chwala]

My outcomes are different than the examples I got. Also I have specimens but can't upload here.

Image Analyst, may I have another question??

[Image Analyst]

Yes, you may.

If it's unrelated to your first one way at the top, then create a new thread.

Answer 4

[Aneta Chwala]

That method was based on color filtration pixel by pixel of the whole image (globally) so how to mix it to local method where we would analyse the image with sliding window?