Multilayered GMDH-Type Neural Network with Radial Basis Functions and its Application to 3-Dimensional Medical Image Recognition of the Liver

Abstract

In this paper, a Group Method of Data Handling (GMDH)-type neural network algorithm with radial basis functions (RBF) is proposed. The proposed algorithm generates optimum RBF network architectures fitting the complexity of nonlinear systems using heuristic self-organization. The number of hidden layers, the number of neurons in hidden layers and relevant input variables are selected by minimizing prediction error defined as Akaike's Information Criterion (AIC). Various nonlinear combinations of variables are initially generated in each layer and only relevant combinations are selected based on AIC. Hence, the optimum RBF network architecture fitting the complexity of the nonlinear system is obtained. We apply the GMDH-type neural network algorithm with RBF to 3-dimensional medical image recognition of the liver, showing that this algorithm is very easy and useful in 3-dimensional medical image recognition of the liver because the neural network architecture is automatically organized to minimize prediction error based on AIC.