- Segementation example from synthetic image
- Bibliography of related papers and texts
Discussion: Application of this technique to scoring a cancer panel of images.
Current scoring techniques are based on loose cell suspensions,
and discard overlapping cells or nuclei as just being noise.
For fixed tissue,
especially with high-proliferation, that technique selectively
discards the data
you are trying to measure! That technique is used
by the CAS-200 system.
The figure above illustrates how discarding the overlapping regions
gives
a false reading of the purple-stained features as being evenly distributed,
and completely misses the high-concentration in the lower half of the
image.
The V-algorithm can "see-through" overlapping features, figure out what
unique combination of features could produce that color, and reconstructs
both original stains as if the other were not there. No data
are discarded.
If you want to measure proliferation, it makes more sense to use a technique
that doesn't discard proliferative regions as being "too hard to measure."
Example:
In the example below, a synthetic image was created with a linear ramp
horizontally of green stain,
overlaid with a linear ramp vertically of blue stain. Segmentation
algorithms that are forced to assign
each pixel to one or the other category result in the segmentation
shown on the top, which is simply
incorrect. According to this type of algorithm, there is ZERO
co-expression of these stains!
The bottom set of images show the result of analyzing this image with
the V-Algorithm.
The actual analysis can
be viewed with this link, which includes the original
jpg you can
download and try your own algorithm on.
* pan-and-paint - multiple image automated alignment and montage
Reticula Corporation
* V-algorithm® dual-stain color-correction and scoring
The blue/green ramps image shown above
Ki-67 example
BioGenex 4A11 example
BioGenex DAB example
H&E example
example 3
Important note: Images digitized from film or otherwise captured
in a non-linear method produce poor results.
Theory behind the V-algorithm (pdf document, requires Acrobat)
* autoscore -- stained nuclei counting
Link to Bibliography of related papers