Microscopic Slice Image Processing and Automatic Labeling - Biomedical Signal Analysis: Contemporary Methods and Applications - page 352

detected and counted using this method. In sections 13.6 and 13.7 we
give a more detailed comparison with other methods.
ZANE
We propose the following adaptive counting algorithm, which we call
ZANE (zell 1 analysis and evaluation). In the first step ZANE performs
image stitching of the various microscope images and manual ROI se-
lection to acquire the analysis image. The main counting step is based
on a method proposed by Nattkemper et al. [183, 184] for evaluating
fluorescence micrographs of lymphocytes invading human tissue; here,
however, it is applied to light microscope images, and classifier prepro-
cessing and training, as well as application, are different. The main idea
is first to construct a function mapping an image patch to a confidence
value in [0 , 1], indicating how probable it is that a cell lies in this patch
or not - we call this function the cell classifier . In the second step this
function is applied as a local filter onto the whole image; its applica-
tion gives a probability distribution over the image with local maxima
at cell positions. Nattkemper et al. call this distribution a confidence
map . Maxima analysis of the confidence map reveals the number and
the position of the cells. A flow chart of the ZANE algorithm is shown
in figure 13.2.
Regions of interest
In practice, the cell counting is to be performed not within the whole
image but only within a restricted region of the image called region of
interest (ROI) . For example, in the presented experiments we want to
count only cells from the dentate gyrus in the hippocampal formation of
the temporal lobe. So far, the selection of the ROI is done manually, but
we hope to automate this process in the future. However, precise criteria
for the ROI detection seem to be dicult to extract - we assume a joint
criterion taking both shape and image background texture into account
is needed.
The impact of manual ROI selection is rather low in our experiments
- the brain region of interest can be roughly identified manually by
brightness and, especially, shape. Small deviations in this identification
(given, for example, when comparing two experts who use implicit
1 German for “cell”.