Evolutionary Modular Neural Network Approach for Breast Cancer Diagnosis

Knowledge Discovery paradigms especially Soft Computing techniques like Artificial Neural Networks have been at the fore front of research aimed at solving the problem areas involved in many diverse fields of application. Automated diagnosis of deadly diseases is one of such fields that have seen much effort from researchers in the last few years. One area where this effort has been most felt is the diagnosis of breast cancer in women. However, development of a computationally efficient, detection-wise effective and robust framework for the diagnosis of breast cancer has still not materialized. The major problem here is the presence of a number of decision variables involved that makes this problem of diagnosis much more complex and intricate. This makes it difficult to be tackled by traditional computing paradigms efficiently. In this paper, we explain how the paradigms of modularity and optimization using evolutionary technique could be used to solve the aforesaid problem with significant success. Here, to take benefit of modularity, we make of use modular neural network instead of the traditional monolithic neural network for the recognition of input vectors implying breast cancer. Also, to make the architecture more optimal, we make use of genetic algorithms to achieve optimal connections (weights) among the neurons in each of the individual experts of the modular neural network. Experimental results show that the proposed approach has been significantly successful in dealing with aforesaid problem of breast cancer diagnosis with a training accuracy of 95.97% and testing accuracy of 96.5%. That is well above what shown by traditional approaches as described later on.