Fakulteta za elektrotehniko, računalništvo in informatiko Laboratorij za sistemsko programsko opremo Univerza v Mariboru
Ultrasound Image Recognition using Cellular Automata and Neural Networks

Author: Bogdan Viher
Mentor: Prof. Dr. Damjan Zazula
Co mentor: Prof. Dr. Andrej Dobnikar
Date: 1998

Ultrasound Image Recognition using Cellular Automata and Neural Networks

Keywords: cellular automata, cellular neural networks, parallel computation, image recognition, ultrasonic images, ovary, follicles

UDK: 621.391:534-8:007.52:681.3

Abstract: The presented work deals with possibilities of using two parallel computational models - cellular automata and cellular neural networks - for ultrasonic image recognition purposes. First we take a look at how exploiting of basic physical ultrasound properties - refraction and reflection - can create an ultrasonic image and what devices have been developed for this purpose so far. Then we discuss in detail the cellular automata and cellular neural network models, which are both based on a large number of identical building blocks, called cells, connected in a multidimensional grid. Mostly, cells adjust their state only according to connections to their nearest neighbors, which permits their massive parallel execution. Cellular automata model allows rule that changes the state of each cell to be specified, whereas cellular neural networks are more or less limited to the sum of cell's weighted inputs, which then determines the output of the cell through a sigmoidal function. After the theoretical overview of both parallel computational models we approach a real-world problem - recognition of follicles in the ultrasonic images of women's ovaries. With both models we have first tried to recognize the largest, dominant follicles only, and in their enhanced version also all of them. The cellular automata recognition algorithm is based ofn the ideal of the immune system, where in the first phase the objects we are looking for are allowed to establish a protection mechanism, which in the second phase prevents them from being destroyed. With cellular neural networks approach, we have used weight combinations known from the literature and achieved the recognition effect with interconnection of several networks. Effectiveness of recognition was tested on fourtyfive test ultrasonic images. Cellular automata were better in looking for real follicles, but have, at the same time, recognized more false follicles than cellular neural networks. At the end we have compared cellular automata recognition method with classical region-growing segmentational approach, since there are some similarities between them.