A neural network based classifier for ultrasonic raw data of the myocardium

The following work describes a new noninvasive method to characterize myocardial tissue using backscattered ultrasonic RF data of the human heart. The assumption is that it is possible to relate ultrasonic RF data to structural properties of myocardium due to chronic heart failure. The ultrasonic RF data was collected using a transoesophageal multiplane probe. The region of interest was at a mean depth of 7.5-10 cm. The analog RF-data was digitized with 8 bits at a sampling rate of 16 MHz using a modified ultrasound system. The RF-data was filtered and segmented in areas representing myocardium, pericardium and blood. Different features were extracted from the signal originating in the myocardium. These features were applied to a Multilayer Perceptron which distinguishes between five different states of disease. A total number of 200 data sets of ultrasonic RF-data were used in the investigation. They were taken in vivo from 33 different patients. Additionally, with each patient a biopsy was performed to get samples of myocardial tissue