Microarray image analysis

[Kathmandu]

Hi there,

I am looking for a better image analysis for the evaluation of microarray pictures. The home-made software we are using at the moment has some major shortcomings, so I just started to find out if Vision could do the job or part of it. Maybe there is already existing something suitable. If you have any pointers, please let me know.

Task description:

We are developing microarrays for water quality management. The images to analyze are recorded with a CCD camera and have a black background with white spots on it. The spots consist of pixels with different intensities. Each spot corresponds to a sample. The spots are arranged on a grid, in columns and rows.

The ideal spot is circular, has a high signal-to-noise ration and an identical diameter. There should be a sharp border between the spot and the image background. In reality, all of these parameters vary to an extent: Intensity, shape (e.g. „donut effects“), size, the sharpness of the transition. Also, there can be additional impurities and smear on the image.

The old software was not flexible enough, e.g. it was necessary to define the exact position of each spot and each center. What the new software should ideally be able to do:

-  Automatically recognize all spots which have a higher signal intensity than x times the background signal
-          Automatically recognize the size of the spots, especially non-circular shapes, and its borders.
-          Distinguish spots from impurities.
-          Calculate a normalized sum signal for each spot.

Thank you for any comments.

Greetings from Munich!

Gerhard

[BruceAmmons]

This project sounds fairly straitforward, and there are several tools in NI Vision that could be used to do the job.  I think a fully automated system would be possible.

Can you post a sample image or two?  That would help clarify the difficulty of the task.

Bruce

[Kathmandu]

Please see a typical images below. The grey one is a raw picture. The other ones are brightened and flamed for better visualization.

A defined amount of analyte (bacterial cells) is placed on each grid position via microcontact printing. Each sample is printed in six parallels. Reagents are added and a chemoluminescence reaction is generated. The light is proportional to the concentration. A sum signal gives more accurate results than e.g. the maximum intensity of a pixel at the corresponding position. It is important in the project to quantify the signal. Up to now, the signals are still weak, and we need to improve the intensities. Depending on the viscosity and concentration, or because of necessary additives in the sample solution, different sizes are resulting.

[BruceAmmons]

This is very interesting.  Some of the spots look like they would be very easy to analyze.  Other spots are badly distorted, very uneven, or can't even be seen.  On your first sample, the top half just looks like random noise.

Some things I would do to improve this:

Use a larger image size with more resolution.  More pixels per spot will always make it easier to analyze the spot.

Take a series of images at different exposure times.  I would consider 1x, 2x, 4x, and 8x as long as the current exposure.  I would use an algorithm that would select the best image to analyze each spot.  Using more of the dynamic range of the camera will improve the results.

Bruce

[Kathmandu]

Thanks for the ideas, Bruce. We are are still at the relative beginning of this project. A larger image size with more resolution would certainly be an advantage. A better camera is also a question of money, though.

I modified the scripts for cell analysis and culture analysis from the solution wizard and played around a bit. The lookup table can be used to increase the dynamic. The principle is good, though far too insensitive. A big disadvantage is that the wizard script defines a threshold and reduces the data to a binary image. Rather, I would need an integrated signal which takes into account the different signal intensities for each pixel in the area of interest.

The signal intensity increases with higher exposure time, true. And so does the background signal. So a higher signal-to-noise ration is more important than absolute signal intensity.

Some background information on the experiment:
High reaction times are only useful, if generally small signals are detectable. Very high signal intensities result in saturation of the detector signals of the CCD chip. The maximum value that is measurable theoretically with a 16 bit CCD chip is 65536 a.u. per pixel. If the value of the signal relative to one pixel is at the limit of the detector, the electrons will be shifted to the next quantum well. The results are visually larger spot diameters (Wolter et al. 2007).