Echocardiogram structure and tissue classification using hierarchical fuzzy neural networks

Abstract

A system to automatically classify structures and tissues in echocardiogram images is presented. Structure classification is the first step required for any system that is designed to measure cardiac parameters. Described here is a multiple feature, hierarchical, neural network fusion solution to the problem. The system 'learns' to classify tissue types by examination of image training data. Classification assigns each image pixel a membership measure for each structure or tissue type. Final hard classification, if required, is delayed until the system's output stage. This allows important fuzzy information to be retained throughout the system. The first layer in the hierarchy of networks determines gross spatial relationships and texture classes. The second layer fuses the spatial and textural net outputs to make the final classifications. Examples of processing real data are presented. >