In vivo liver differentiation by ultrasound using an artificial neural network

A pattern recognition algorithm and instrumentation for in vivo ultrasound human liver differentiation are presented. An available 16-MHz microprocessor-based data acquisition and analysis system with 6-bit resolution is used to capture, digitize, and store the backscattered ultrasound signal. The algorithm is based on a multilayer perception neural network using the backpropagation training procedure. The network is implemented to differentiate between normal and abnormal liver. Data earlier obtained from 18 volunteers with normal liver history and from 12 volunteers with liver abnormalities are used to test the algorithm. The power spectra of the backscattered signal from depths of 5, 6.5, and 8 cm in the liver are calculated. The acoustic attenuation coefficient is calculated by the log spectral difference technique over the frequency range from 1.5 to 4.5 MHz. The change of speed of sound with frequency (dispersion) is estimated over the 3-MHz bandwidth. The attenuation and velocity dispersion are used as differentiation features. The results show that of the 22 tested cases, the system differentiated correctly 19 and 20 cases when using the attenuation and the velocity dispersion, respectively. The average magnitude of dispersion of liver is estimated to be 1.67 +/- 0.1 m/s/MHz and about 2.3 +/- 0.18 m/s/MHz in the normal and abnormal cases, respectively. The overall performance of the system for liver differentiation is 91% for normal cases, and 86% for abnormal cases. The data files are also differentiated using the nearest neighbor statistical classifier. The results show that of the 30 tested cases, 23 files are differentiated correctly using the attenuation coefficient.