MOGA design for neural networks based system for automatic diagnosis of Cerebral Vascular Accidents

Design of a neural network classifier involves selection of input features and a network structure from a very large search space, preferably respecting the problem's constraints. Most published methods just focus on the feature selection aspect and do not consider any approach for determining a model structure that best fits the application at their hand. Moreover, the design criteria usually include multiple conflicting objectives which may not be handled simultaneously. The proposed method aims maximization of classification precision while reducing Neural Network (NN) model complexity. A Multi Objective Genetic Algorithm (MOGA) based approach is used to determine the architecture of the classifier, its corresponding parameters and input features subject to multiple objectives and their corresponding restrictions and priorities. This classifier is part of a computerized automatic diagnosis system for identification of Cerebral Vascular Accident (CVA) through analysis of Computer Tomographic images (CT). Comparison with Support Vector Machine (SVM) results shows that the number of False Detections (FD) in both validation and test sets in the model obtained by the proposed work is lower than that of the SVM; even when large number of support vectors is used.