Ultrasonic tissue characterisation using neural networks

Ultrasound imaging is an important diagnostic tool in medical practice and research. It can be used to scan soft tissues, to characterise and to classify these according to possible diseases. In this paper diffuse liver diseases are studied, the available database consists of healthy livers and four kinds of diseases. From the echographic measurements five parameters are calculated for tissue characterisation. To obtain a sufficiently large training set artificial data is generated using an optimal kernel estimate of the probability density function of the original data. Tissue characterisation is then performed using different kinds of neural networks: feedforward networks with error back propagation, self-organising feature maps and the ARTMAP network. The obtained results are given, compared and discussed. The results are also compared with a classification based on discriminant analysis.<
>