I'm trying to automate the characterisation of diatom colonies sprayed onto agar plates. As you can see in the attached images, there are basically two types: stationary/round and swarming/ragged.
The recognition and counting of the colonies in general works quite well, based on edge-finding and contrast. To distinguish the colony types, the ImageJ macro uses "roundness" (length of circumference vs. the particle's area), but this misses many colonies of the swarming type.
Can you suggest other parameters that would be useful discriminators?
By looking at the Image I think it's worth taking the Color Distribution into account. The round ones are nearly fully brown, but the swarming ones have a large gray area compared to the brown center.
I have had success in the past discriminating colonies (and lots of other things) with the advanced weka segmentation in the Fiji distribution of imageJ. See here
http://fiji.sc/Advanced_Weka_Segmentation
Another thing to try could be the ratio of perimiter to area, it would be quite different in the ragged colonies
Thanks for the 3 suggestions, will try. "ratio of perimiter to area" I already tried, resulting in the above image, lacking good discrimination ;-)
Hi,
there are many morphometric features you can use for comparing cells or particles;
most generic ones can measure the "size" (area, perimeter, diameter of enclosing circle...). For comparing the shape, some common options include:
* "shape factro", as ratio of p^2 over area
* convexity, as ratio of area over convex area
* elongation, as ratio of Inertia ellipse axes lengths
* ratio of enclosing and inner discs diameters
* ...
If you are interested in the rugosity, you can try the convexity, or compare area over the area after performing morphological closing or opening.
Maybe in your case you can also be interested in comparing the relative areas of kernel to full cell, in this case you need more segmentation.
You might like to try some of these: Centroid Distance Function, Ellipse Variance,
Contour Curvature, etc.
More feature are described here:
http://math.uci.edu/icamp/summer/research/classification/shape_descriptors_survey_part3.pdf
Thanks a lot for the additional ideas! I'll be occupied for a while :-)
I Katrin,
Why don't you try with fractal analysis. Fractal exponent have been successfully used to discriminate morphologies for different cell phenotypes, tumor malignancies and neuronal diseases.
I have found this web that could help you...
http://imagej.nih.gov/ij/plugins/fraclac/FLHelp/FractalDimension.htm
I you have any query regarding this approach you can contact me. If you try it, please give a feedback, I am very curious if this approach works in your photos.
Good luck!
Carlos
Are all these ideas dependant on the availability of uncompressed images images (ZVI in my case)? Or do compressed ones like JPG work as well?
Uncompressed is preferable, but as Long as the compression does introduce artifacts (if the compression rate is set too high) or significantly changes colors, working with compressed images should work without issues.
Here are some good ideas how to deal with touching and overlapping objects: http://blogs.mathworks.com/steve/2006/06/02/cell-segmentation/
Touching & overlapping colonies are resolved with sufficient accuracy. By watershed even, as described in the article ;-)
Just the sorting into one or the other type is still problematic.
Dear Katrin:
I do not understand why it is supposed to be a colony of diatom, in some of the images I see a depression suggesting a cingulum (see specimen that contain the numbers 95.92-99.16 and 96.71-99.53 specimen).
It would be useful for me to understand what it is suggesting you that the images show colonies of diatoms.
Kind regards. Eugenia
- Badges
- Science method