After thresholding and connected component labeling, we get:
If prior image labelling, watershed is performed, the chromosomes are broken into particles that corresponds mainly to the chromatids of the largest chromosomes:
We can see that most of the chromatids are well detected in large chromosomes but less accurately in small chomosomes such HSA 19 or HSA21:
Some pieces of chromatids are well detected, let's paint ("Human labeling") them by hand:
- in green or blue for the chromatids
- in red for the centromeric labeling
- in yellow, for the badly segmented chromatids.
Yellow regions needs to be reaffected to blue / green or red region. By hands, it gives:
Can we imagine an algorithm capable to perform that task?
A low level image opening finally yields: