A deep learning approach to increase the frame rate of acoustic radiation force impulse displacement imaging

Multiparametric imaging based on acoustic radiation force impulse (ARFI) displacement and shear wave (SW) imaging improves diagnostic accuracy. However, there is a potential for misalignment error and extended exposure to the radiation force if ARFI and SW images are collected separately. Therefore, the objective of this study is to generate “on-axis” displacement images from the “off-axis” SW displacement data using a convolutional neural network (CNN), thereby enabling high-frame-rate ARFI imaging. The input to the CNN is the “off-axis” displacement images, and the learning task of the model is to predict the “on-axis” displacement image. The LS-DYNA3D-simulated training data consist of five heterogeneous materials, with a background elasticity of 1.5–5 kPa and viscosity of 0.3–2.6 Pa s, and a 3-mm radius spherical inclusion with elasticity of 6–20 kPa and viscosity of 0.3–4 Pa s (F-number = 2.5 and focal depth = 20 mm). There is an on-average decrease in the contrast-to-noise ratio (CNR) of 8.1% and 32.52% compared to conventional ARFI imaging when increasing the framerate by 12.5 and 25 times, respectively. While the CNR improved by 2.85–9.5%, increasing the framerate by 5 –8.33 times due to averaging from multiple pushes. Future study will investigate ultrasonic tracking and validate it experimentally.