A Fully Convolutional Neural Network for Rapid Displacement Estimation in ARFI Imaging

Abstract

Ultrasound elasticity imaging in soft tissue with acoustic radiation force requires extracting displacement information, typically on the order of several microns, from raw data. In this work, we implement a fully convolutional neural network for ultrasound displacement estimation. We present a novel method for generating ultrasound training data, in which virtual displacement volumes are created with a combination of randomly-seeded ellipsoids. Network performance was tested on the virtual displacement volumes as well as an experimental phantom dataset and human in vivo prostate data. In simulated and phantom data, the proposed neural network accurately reconstructed the ARFI displacements, performing similarly to a conventional phase-shift displacement estimation algorithm. Application of the trained network to in vivo prostate data enabled the visualization of the prostatic urethra and peripheral zone.