Vector flow imaging using a deep neural network

Abstract

Vector flow imaging (VFI) estimates blood flow velocities in both azimuth and axial dimensions. VFI has promising applications in the characterization of complex flow patterns, including cardiac flow and abdominal flow imaging. Conventional VFI relies on the use of multiple angles in transmit or receive, or speckle tracking. They are computationally intensive and estimate quality may be sacrificed to improve computational speed. In this work, we report a vector flow estimation technique using a deep neural network. The network extracts feature from high-pass filtered beamsummed RF data of two consecutive Doppler packets. For each packet, the RF data have azimuth and axial dimensions. It then performs estimation of vector flow in the feature space, and maps the estimate back to the spatial domain. The total computation time is 0.11 s for a pair of Doppler frames of 1024 × 128 samples. The performance of the method is characterized using Field II simulation studies, flow phantom studies, and an in vivo liver study with a Verasonics Vantage 256 scanner. The simulation and flow phantom studies show good agreement between the estimates and the ground truth. The in vivo studies demonstrate that the method is capable of characterizing complex flow patterns in human liver vessels.

ENNS: Variable Selection, Regression, Classification and Deep Neural Network for High-Dimensional Data

“…Deep Neural Network (DNN) is a neural network with a deep structure of hidden layers, which has better performance than the shallow neural network (neural network with only one hidden layer) in many aspects in a broad field of applications such as pattern recognition, speech recognition and computer vision, see for example [48,43,36]. The deep structure has a greater approximation power than a shallow neural network.…”