Feature selection for MR image classification

Abstract

Elements of the feature vectors are searched to increase the classification performance of MR images and to reduce the number of nodes of the neural network. Elements of a feature vector are determined by dynamic programming. This algorithm uses divergence analysis and orders the elements of the feature vector to give maximum divergence. The classification performance of new feature vectors is compared with features formed by the gray values at one neighborhood of the center pixel. The MoRCE network, which gave satisfactory results in the previous study (Z. Dokur et al., 20th Annual Int. Conf. of the IEEE-EMBS, vol. 20, no. 3, p. 1418-21, 1998), is used as the classifier. MoRCE gives 97% classification performance with 7 nodes by using the new feature vectors.