NEURAL NETWORKS FOR X-RAY IMAGE SEGMENTATION



Darryl Davis, Su Linying & Bernadette Sharp



12th International Conference On Control Systems And Computer Science,

Romania, 1999


Abstract

The paper addresses the application and challenges of using neural networks to segment gray-level images; approaches we term direct perception. The work described here is part of Intelligent Multi-Agent Image Analysis System, which is being developed to promote the automated diagnosis and classification of digital images. In this paper we show how neural networks may be successfully segment medical X-ray images of the thigh. Weak use of  Back Propagation neural network, Counter Propagation neural network, Self-Organizing Feature Map, Bi-directional Associative Memory, and a hybrid network consisted of BP and SOFM. The comparisons among their performance are made, and some feature extraction techniques used here are presented. A kind of one layer neural network, known as WISARD, is used to validate the segmentation performance based on the segmentation results. A highly general validation information, the centroid curve of the segmented images, is proposed here.