Comparison between two different methodologies to design neural classifiers

In this work we present the design of a neural classifier to be used in an automatic system aimed to recognize and count cells in different stages of development. The purpose of the neural classifier subsystem is to determine the kind of cell a pattern obtained from the digital image processing units belongs. However, the different cells presented in a preparation are not easy to discriminate, even to a human expert, due to the fact that the cell features that can be processed are not totally discriminatory. In this way, the neural classifier must provide a reliable classification of a set of overlapping patterns. In order to build a neural classifier, its structural and performance features must be adapted to the problem characteristics. Nevertheless, to find the best structure that produces the best classifier is a problem in itself that involves several factors. We had faced this problem comparing two different methodologies: one using classical neural network techniques to select the proper architecture and training process; the other using multi-evolutionary techniques to search optimal configuration of the classifier.