Inverted Pulse Estimation in Pulse Inversion Harmonic Imaging using Deep Learning

Conventional methods such as filtering and multi-pulse acquisition techniques play an important role in ultrasonic nonlinear imaging, specifically harmonic imaging. However, the former suffers from spectral leakage, while the latter is hindered by the low frame rate and increased susceptibility to motion artifacts caused by the need to transmit two or more pulses. In this work, deep learning concepts are exploited to achieve a single-shot harmonic imaging application. This is accomplished by implementing an asymmetric convolutional autoencoder, capable of learning the nonlinear relationship between the echo resulting from the first transmitted pulse (network input) and the echo resulting from the second transmitted (inverted) pulse (network output) used in the conventional pulse inversion technique. The proposed approach yielded harmonic images with comparable contrast to noise and contrast ratios to the conventional pulse inversion technique, yet at approximately double the frame rate. These findings pave the way for harmonic imaging to be used with quality that is on par with the quality achieved by current harmonic imaging methods but with a higher framerate and fewer motion artifacts.