Visualization of Untrained Target Shape and Distance Using a Multi-element Ultrasound Transducer and Neural Network

Abstract A scheme of detecting rectangular targets placed on a plane in water is realized using a multi-element ultrasound transducer. The transducer is designed aiming at future implementation of 3-D endovascular ultrasound endoscope for clinical diagnosis. A three-layer artificial neural network is employed for training and detecting targets placed on three vertical planes located at different positions in front of the transducer. Ultrasound echoes received by the transducer are fed to the neural network which employs the back propagation algorithm. Rough visualization of rectangular targets consisting of square units (pixels) is achieved by integrating detection results even for untrained shapes at untrained distance after training on limited shapes and sizes of primitive patterns formed by square units.