Microscopic Slice Image Processing and Automatic Labeling - Biomedical Signal Analysis: Contemporary Methods and Applications - page 375

Table 13.1
Mean square error of automatic counting methods applied to the five section images
from figure 13.11.

Algorithm Counting error Counting error

(vs. manual+focusing) (vs. manual)

ZANE 3.4 3.5

clustering 5.0 6.8

watershed 5.8 5.4

preferable: using the data from figure 13.11(c), we get mean square errors
of the three counting algorithms as shown in table 13.1. Clearly ZANE
considerably outperforms the other two algorithms, especially in terms of
counting performance given only the digital image data. A comparison
of the computing times T (figure 13.11)(d), shows that ZANE (mean
T = 24sec) and Benali's clustering method (mean T = 29sec) perform
similarly, with the watershed taking roughly twice as long (mean T =
68sec).
Finally, we wanted to test ZANE on images with different numbers
of labeled cells and to give some neuro biologically interesting results.
We therefore analyzed neurogenesis in the dentate gyrus of mice, and
compared a control group of animals with mice that had been treated
with pilocarpin to induce a status epilepticus. It is known that this
condition raises the number of proliferating cells in the hippocampus
one week after the treatment, which can be shown by an increase of
BrdU-labeled cells [31, 196]. Our experiment confirmed these findings
by showing that proliferation of cells in the dentate gyrus was 340%
stronger in the epilepticmice than in the control group (see figure 13.12).

Counting images with multiple markers

The advantage of the presented method lies in the fact that it can
be readily extended to the detection, localization, and identification of
other kinds of cells in microscopic images. For example, images marked
by multiple markers such as neuronal nuclei antigen (NeuN) ,BrdU,
doublecortin (DCX) ,or S100β allow the differentiation of various types
of cells. By adapting the cell classifier we can identify the desired type of
cell, and multiple cell classifiers can then identify the various cell classes.
In the following we will demonstrate this by applying ZANE to a slice