First, you will need to tackle the second issue: a segmentation problem. It might very well be that there is no good or robust threshold value. Maybe I can help; can you upload an example image(s)?

The first issue can be solved by reducing the width to unit-length, i.e. make the width equal to one. Google 'thinning' or 'skeletonization'.

Thank you so much for the positive reply. Please find the attached images. The original image (in color with length scales) will give you other details.

Here is the gray-scale image from which you should extract the intensity values after converting it into gray-scale.

I agree that u need to do thinning or skeletonization from the binary image before you can count the pixel to improve the accuracy of the length measurement.

I did following :

- Crop the central part from the color image

- Convert to short

- Subtract Green from Blue channel

- Threshold at zero

- (Not done, but you should put a threshold to remove small blobs)

- Skeletonize

(Instead of skeletonization, you could also trace the blobs, calculate the area and then divide by two)

Good luck!

Great job Lambert. If you using Matlab, maybe u can search for blob analysis or connected component where u can access many blob properties without developing the algorithm from scratch.

It is very simple if you implement the code in Matlab, where exist particular function for adding the contour and afrerwards counting the pixels of the contour or of the features. I don't know that method you used in determination of the threshold, but a efficient method is Otsu's method.

Yes, Otsu's method works quite good, but on which color channel do you apply it? Or would you do the histogram separation directly in 3D vector space?

I recommend to play with the color channels first (e.g. subtraction), to see if one can get the signal stronger. Maybe also try another color representation like YCrCb.

Following are some of my results of skelonization using Mathematica:

Check this one also

I used a threshold value of 0.3 in both the cases.

Interesting!

Apparently I use a skeletonization algorithm that is much less sensitive.

There are many versions around, the reason being : while the skeleton of a closed contour is well defined in the continuous domain, this is not so in a binary image.

Either look for a less sensitive version or trace the DNA blobs and calculate the circumference. Because the DNA strands are thin and long, the length is approximately half the circumference.

Did you use MATLAB? If not, then what did you use?

Actually, I wrote it myself ! It is an iterative peeling algorithm. It is part of a huge proprietary image procession library of the Netherlands Cancer Institute.

From the documentation of the module :

Related to a 'medial axis transform' are terms like 'thinning' and 'skeletonizing'. The computational complexity of the medial axis transform is very high.

The algorithm used in this module was introduced by Zhang and Suen, and produces a fairly good approximation of the medial axis (it does not reproduce all details) in reasonable time...

(T.Y.Zhang and C.Y.Suen, 'A fast parallel algorithm for thinning digital patterns',

Communications of the ACM, 27,(3),236-239,1984.)

P.S. maybe put skeletonize threshold much higher?

@Simona:

Could you tell me what is that function in MATLAB for determining the contour?

I don't think any intensity based global thresholding will work in this case. The features differ from each other in terms of aspect ratio and perimeter.

If I have a nice image like this

Then I can do thinning very successfully

Result of binarization based on Otsu's global threshold.

On WHAT did you apply Otsu?

@Krishnashish : You write "I don't think any intensity based global thresholding will work in this case. The features differ from each other in terms of aspect ratio and perimeter".

Intensity thresholding has nothing to do with shape (aspect ratio, perimeter). I thought the threshold of 0.3 was a sensitivity parameter of skeletonize. Increasing the value will (probably) make it want to preserve less features, which is what you want in this case.

Probably you used Otsu on the monochrome image you supplied, and you threw away the color information. Try it on Blue minus Green!

Finally i am successful.

Thank you all.

Almost there, but the DNA suffers from double strand breaks!

Could you share your method with us?

I wonder why you have so many circles in them, and why the only one in the result I posted is absent. Do you have holes in the DNA segmentation?

I believe that your method is better, but I couldn't do that unfortunately.

Could you explain the following from your second reply:

- convert to short

- threshold at zero

- subtract green from blue channel

You had a color image. Such images have three color channels; in fact you have three images. Red, Green and Blue: RGB.

For each channel one byte is used per pixel.

I wanted to subtract the image Green from image Blue, because after a quick look at the three channels, the DNA appeared to stand out better from the background by doing that.

However, subtracting one byte value from another will lead to undesirable overflow or wrap around conditions.

To prevent that, one can convert the unsigned bytes to signed short integer (two bytes) or to signed integer (four bytes).

The threshold at value zero just worked for me and is kind of 'neat' (every pixel that has a higher Blue component than a Green component is DNA), but it is probably better to let Otsu find a good threshold.

Thank you. I understand you now.

If it's not too late, you have great functions under matlab in the image processing toolbox to do this kind of job (regionprop function=>shape measurements=>perimeter).

I did it in Mathematica. Thanks a lot Sebastien. I don't know MATLAB.

I am a pure biologist and hence completely useless when comes to writing codes to measure pixel intensity. What I use is simple tools in ImageJ where you can measure contour length of DNA by simply tracing it.