Posted by
karo03 on
URL: http://imagej.273.s1.nabble.com/Watershed-question-tp3685658p3685665.html
Hi,
Am 13.06.2010 um 13:34 schrieb NatashaW:
> Many thanks for your explanation; you have exactly mentioned the most
> problems that I have found during my work!!
> I’m really working to count/measuring some particles (foci) inside the
> nuclei. Daniel’s algorithm gives me good results when I work on the foci
> images (gray-level values).
Seemingly foci are mostly of the same size and compact.
>
> But during this work, I found some problems:
>
> 1-When I smoothed the input image, the oversegmentation problem has been
> solved.. But watershed didn’t separate all the desired objects. I’m still
> looking for a better threshold value. Indeed, I didn’t find yet the best
> criteria to control the MIN and MAX values..
With the MIN and MAX values, you will not improve the separation of connected objects but only the size of the basins, the connected components. Typically doing a full watershed from 0 to 255 gives you a partition of the whole image. Basins are delimited by the topology of the intensity of the input image. The borders of the basins, some sort of an exo-skeleton, allows you to cut the binary image won by any other segmentation "mask XOR exoskeleton". The MIN/MAX values are an implicit thresholding.
>
> 2-The detected particles are bigger than their real size. This may cause a
> problem during the measurement (like the mean and IntDen..) because the
> boundaries of the detected objects don’t match exactly their real size..
> Maybe because of the chosen MIN value??
(See above) If you need exact particles areas (I don't know what that might be!) you can perform a full watershed (0..255) and segment each basin with a separate (automatic) threshold.
>
> 3-Is Daniel’s watershed algorithm an implementation of this article:
> Vincent, Lee & Soille, Pierre, "Watersheds in digital spaces: An efficient
> algorithm based on immersion simulations", IEEE PAMI13 (6): 583-598
> If not, could someone tell me which one? I want to understand how the
> algorithm works Please..
I have not looked into Daniel's algorithm, but results looks like the Vincent algorithm. Differences in implementation are the result of different sequences to test neighborhood conditions. The algorithm is described in the Vincent paper. In fact the basin filling can be understood as an efficient method to find monotone path connectivity. Each point pair reachable by a monotone path (in terms of intensity) belongs to the same basin (or connected region). From this explanation (?) you might understand, that the separation expected or wished has not much to do with the separation got, although the algorithm delivers watersheds. Additionally the digital grid with only 4 or eight directions influence strongly the exo-skeleton, easily visible in watershed on distance maps with large connected objects and especially with objects of varying size.
If the foci are well detected possibly the number of cells, typically much less than foci, can be estimated by simpler tools (mean area, manual count). At least, automatic separation should never be used without carefull observation!
Regards
Karsten
[hidden email]