NEURAL-NETWORK APPROACH TO CHARACTERIZATION OF CIRRHOTIC PARENCHYMAL ECHO PATTERNS

We have classified parenchymal echo patterns of cirrhotic liver into f our types, according to the size of hypoechoic nodular lesions. Neural network technique has been applied to the characterization of hepatic parenchymal diseases in ultrasonic B-scan texture. We employed a mult i-layer feedforward neural network utilizing the back-propagation algo rithm. We carried out four kinds of pre-processings for liver parenchy mal pattern in the images. We describe the examination of each perform ance by these pre-processing techniques. We show four results using (1 ) only magnitudes of FFT pre-processing, (2) both magnitudes and phase angles, (3) data normalized by the maximum value in the dataset, and (4) data normalized by variance of the dataset. Among the 4 pre-proces sing data treatments studied, the process combining FFT phase angles a nd magnitudes of FFT is found to be the most efficient.