Microscopic Slice Image Processing and Automatic Labeling - Biomedical Signal Analysis: Contemporary Methods and Applications - page 379

detector was then applied onto the image to get a confidence map, and
maxima analysis yielded the cell locations. Experiments also showed
good performance of the classifier when compared against more tradi-
tional segmentation techniques.
In future work, one goal is to face the problem that in typical brain
section images, some cells not lying directly in the focus plane are
blurred. In order to count those without counting them twice in two sec-
tion images with different focus planes, a three-dimensional cell classifier
can be trained for fixed focus plane distances. A different approach for
accounting for non focused cells is simply to allow “overcounting”, and
then to reduce doubles in the segmented images according to location.
This seems suitable, given the fact that cells do not vary greatly in size.
We also plan on automating ROI selection in the future by classifying
separating features of these regions.