A convolutional neural network approach to preserve image quality for sparse array data

Abstract

High-frame-rate ultrasound (HiFRUS) has seen recent interest in resolving highly dynamic spatiotemporal events but transferring the large amount of data generated from the probe remains a hurdle for real-time imaging. One method to lessen the data traffic while preserving the field of view is to reduce the channel count, but this can lead to image quality degradation and the appearance of grating lobe artifacts. In this work, we present a convolutional neural network (CNN) based framework that uses a sparse array (half the channel count) and infers the remaining channels to mimic a fully-populated array. On unfocused transmissions, our results show that on a beamformed image of a multiple point target phantom, grating lobe artifacts are reduced from over 8dB (sparse array) to less than 1 dB (CNN interpolated) when compared to an image beamformed using the full array. Additionally, reconstructions from CNN generated data demonstrated improvement (10 dB) in carotid echolucent flow regions in vivo. Our work demonstrates that, using a deep learning approach to channel-domain radiofrequency data interpolation, the required physical channel count on an array and the corresponding data transfer bandwidth can both be reduced without significant image quality trade-off.