Neural Networks versus Genetic Algorithms as Medical Classifiers - Foundations on Natural and Artificial Computation - page 394

The initial population is randomly generated to ensure the diversity of chro-
mosomes and represent the majority of the population. As in nature, not all
chromosomes are capable of reproduction. This will be decided by applying
over them a function called fitness, which assigns a score to each chromosome
depending on how close they are to the best possible solution.
There are several ways to get the selection of the next generation, the main
ones are the crossover, and mutation. In the crossover two chromosomes are
randomly selected to combine their genes and obtain two new descendant chro-
mosomes, composed of the parental characteristics combination. This is the main
operation in genetic algorithms. On the other hand, there is the mutation opera-
tor that is far less common. This operator randomly selects some genes from one
chromosome and changes their values with their complementaries. For example,
if we were talking about a genotype, if the gene value is 0 mutation will change
it to 1.
As we can see, genetic algorithms are systematic methods for solving search
and optimization problems, by means of applying to them the same methods
of biological evolution: population-based selection, reproduction and optimizing
mutation. We can find some researches focused in the application of genetic algo-
rithms to the general problem of classification [9], but very few researchers have
used this soft computing tool for medical diagnosis [10]. Nevertheless, recent
innovations have introduced new algorithms that overcome traditional methods
and are more likely to be accepted by the medical community, which is an impor-
tant fact, since the lack of acceptance of this tools by healthcare professionals
is one of the most critical problems hindering the development of automatic
diagnostic systems.
In 1998 it was made one of the first research projects where genetic algorithms
were used in medicine known as Galactica. In this project, a learning system that
used GA to discover knowledge rules from a clinical database was developed. This
expert system used precise rules for the diagnosis of female urinary incontinence
[11]. Subsequently, GA have also been used for the diagnosis of heart diseases
using samples from tests' results stored in a huge database [1]. More recent
examples of GA used in clinical field are the diagnosis of breast cancer through
processing 3D images [12], and to show if there is a relationship between DNA
and breast cancer [13].
2.2 Neural Networks for Diagnosis
Neural networks are not the only method of machine learning used for automatic
classification tasks, but rather a widely used tool for this purpose [14]. In general,
an ANN is able to model complex biological systems, revealing relationships
between the input information that cannot always be recognized by conventional
analysis.
To have a set of examples representing previous experiences is essential to
build an ANN-based classifier that ensures a good performace of learning and
generalization processes. These previous experiences are the ANNs design inputs.