Towards automated enhancement, segmentation and classification of digital brain images using networks of networks

An adaptive image processing algorithm (IEEE Trans. Neural Networks 8 (1997 ) 169), based on biological principles, is extended and applied to enhance, segment and classify digital brain images acquired using magnetic resonanc e imaging (MRI). The algorithm is based on the network of networks (NoN) ne ural computing theory (World Cong. Neural Networks 1 (1995) 561). The initi al algorithm, developed by Guan and Sutton (GS), uses vector connections am ong model neurons to delineate regions of pixels that have dynamic boundari es, corresponding to critical features of images. The boundaries subdivide large regions into smaller regions, where the smaller regions operate as gr adient descent networks to enhance local aspects of blurred images. In this report, the authors develop and test two new algorithms that build upon th e success of the GS algorithm. The first algorithm, termed the segmentation -variance (Sv) algorithm, maps pixels into discrete groups using a segmenta tion function of the local variance. Boundaries formed at the transition zo nes among groups are subsequently tuned to automatically delineate brain re gions of interest. This is demonstrated using MRI scans of the human brain. In a second algorithm, pixel groupings formed using the SV algorithm are c onsolidated into larger groups to generate histograms. Dynamic maps and his tograms capture features of object relationships and are used to classify s tructures. The resultant algorithm is called the dynamic segmentation and c lassification (DSC) algorithm. It is tested on MRI scans of four geometric patterns under various noise conditions. The SV and DSC algorithms illustra te proof of principal concepts in the evolution towards automated segmentat ion and classification of digital brain images. (C) 2001 Elsevier Science I nc. All rights reserved.