Microscopic Slice Image Processing and Automatic Labeling - Biomedical Signal Analysis: Contemporary Methods and Applications - page 370

as other experimental settings such as [191]. ZANE resolves this problem
by using the previously learned shapes; the better counting performance
can be observed in figure 13.9(c). Finally, the advantage of ZANE lies
in the fact that after training, image analysis is basically performed by
a combination of filtering (to generate the confidence map) and maxima
analysis, which is enough for counting but not for segmentation. Hence,
ZANE can be expected to outperform the above methods.

13.7 Results

Brain section image acquisition

The analyzed brain sections were taken from the dentate gyrus of mice.
BrdU given systemically is integrated into the dividing DNA instead
of thymidine during the S-phase of the mitosis [69]. Using a specific
antibody against BrdU, labeled cells can be detected by an immunohis-
tochemical staining procedure. The nuclei of labeled cells on 40 μm -thick
brain sections are dense dark brown or black. To determine the number
of BrdU-positive cells in the granular cell layer of the dentate gyrus,
they were counted on a light microscope (Olympus IX 70; Hamburg,
Germany) with a 20 × objective. Digital images with a resolution of

1600 1200 pixels were taken by a color video camera adapted to the

analySIS software system (Soft Imaging System, Munster, Germany).

×

Manual counting

Before analyzing ZANE counting performance, it should be noted that
the number of labeled cells counted by an expert varies considerably
between observers. This is not due only to inaccuracies of the experts,
but mainly depends on the interpretations of whether a labeled cell is
present at a location or not (i.e., on the personal cell classifiers of the
experts). In order to quantitatively compare the proposed algorithm with
manual counts, we therefore need an estimate of the expert cell count
deviation. Such a deviation has been mentioned in the literature [30],
but finding actual values is dicult.
From our knowledge, a typical value of the inter observer variability
in expert cell counts ranges from 10% to 15%. In order to confirm this,
we compare the cell counts of two experts given 18 different animals with
five to seven slices each. Each slice is counted using the microscope, so