How can I segment micro aneurysms and hemorrhages from fundus images in case of diabetic retinopathy based on thresholding?Please suggest me threshold to segment this.Thanks in advance

In this image the dark lesions has to be segmented only which excludes retinal blood vessels .

[KALYAN ACHARJYA]

What is your question exactly (specific)?

For both (microaneurysms and hemorrhages), you have to go for the different classifier.

[supriya suman]

Yes sir,I have to detect microaneurysms and hemorrahages both from the image.I am using adaptive threshold but along with microaneurysms and hemorrahages,I am getting blood vessels which is not required.Is there any thresholding method to detect microaneurysms and hemorrahages.Thanks

[Star Strider]

There aren’t any microaneurysms in that image. There are only hemorrhages and exudates (infarcts).

[supriya suman]

Yes,but it can be present in other image.Basically I want to segment both

Answer 1

[Image Analyst]

Thresholding alone is most likely no good unless there is substantial pre- and post-processing.

binaryImage = rgbImage(:,:,3) < someValue;

See how people have done it successfully here in published papers: Retina Image Processing

[supriya suman]

yes sir, u are right but I want some easy code,as I am beginner in this field.

[Image Analyst]

Then try the code I gave you. It won't work very well though.

[supriya suman]

Thanks

[supriya suman]

Can anyone provide top-hat transform code to supress the blood vessels for accurate detection ?

[Image Analyst]

There is an example in the help for imtophat(). Chances are it won't be that great. At the very least you'd have to post process the result to clean it up and remove false positives.

[supriya suman]

I am getting image like this after applying thresholding but I don't want vessels here.Please suggest me some method to suppress blood vessels.

[Image Analyst]

I already did a suggestion. In my answer, I said this:

"See how people have done it successfully here in published papers: Retina Image Processing"

Personally I have not reviewed all those methods and don't know the "best" one. I imagine all the authors think theirs is the best. Nor do I have any code to do them. Perhaps you can look for an overview/survey paper.

[supriya suman]

Is there any code to eliminate those blood vessels after thresholding?

[Image Analyst]

What is your definition of eliminate? Blacken? Just make a 3-D mask with repmat() or cat(),

mask3d = cat(3, mask, mask, mask);

and use the mask as an index to blacken those pixels.

rgbImage(mask3d) = 0;

[supriya suman]

I want to substract blood vessles from the image using line operator But i know how to do it.

[Image Analyst]

Subtracting the vessel image pixel values from the original image is the same as blackening the pixels. The value will be zero there.

[supriya suman]

Here how can we recognize blood vessels through mask

[Image Analyst]

I don't know what that means. "through" the mask??? You've already recognized the vessels - that's how you created the mask in the first place. Right? How did you get the mask if it was not from finding/recognizing the vessels?

[supriya suman]

No,i have not recognized vessels.I need code to detect blood vessels to eliminate them after subtracting from image .

[Image Analyst]

I still don't understand. You say you want to eliminate them "AFTER" subtracting them from the image. So to subtract them, you must have identified them already, right? If you're already subtracted them, then you must have a mask for them. And if they've already been subtracted, what does the image look like now? What is there that is left to "eliminate" AFTER subtraction has been done? After subtraction, the pixels will be black where the vessels are. So what does "eliminate" mean to you? You want to eliminate the black (subtracted) pixels??? Do you want to use regionfill() to fill in the black/subtracted areas with local background?

[supriya suman]

I have not detected vessels till now

[supriya suman]

I need a code to detect blood vessels first

[Image Analyst]

I already gave my answer for that above. Perhaps you overlooked it so here it is again:

See how people have done it successfully here in published papers: Retina Image Processing

So pick one and code it up. Or ask the authors if you can get (free or buy) their code. But I don't have the code myself.