Vector Flow Velocity Estimation from Beamsummed Data Using Deep Neural Networks

Abstract

Vector flow imaging (VFI) is a novel velocity measurement technique that provides flow velocity information in both azimuth and axial dimensions. Compared to conventional color Doppler imaging, VFI provides velocity estimation that is independent of flow directions. Previous VFI techniques utilize either multiple transmit or receive beams or angles, or speckle tracking. This creates a trade-off between computational intensity and estimate quality or equipment cost. In this work, we present a vector flow velocity estimation technique based on deep neural networks using only beamsummed radio-frequency (RF) data. The deep neural network extracts features from the RF data, and performs flow velocity estimation on the features, and maps the estimates back to the spatial domain. The structure and training of the neural network model is presented. The performance of the technique is demonstrated and evaluated using simulations and flow phantom experiments.