Complex Residual Attention U-Net for Fast Ultrasound Imaging from a Single Plane-Wave Equivalent to Diverging Wave Imaging

Abstract

Plane wave imaging persists as a focal point of research due to its high frame rate and low complexity. However, in spite of these advantages, its performance can be compromised by several factors such as noise, speckle, and artifacts that affect the image quality and resolution. In this paper, we propose an attention-based complex convolutional residual U-Net to reconstruct improved in-phase/quadrature complex data from a single insonification acquisition that matches diverging wave imaging. Our approach introduces an attention mechanism to the complex domain in conjunction with complex convolution to incorporate phase information and improve the image quality matching images obtained using coherent compounding imaging. To validate the effectiveness of this method, we trained our network on a simulated phased array dataset and evaluated it using in vitro and in vivo data. The experimental results show that our approach improved the ultrasound image quality by focusing the network's attention on critical aspects of the complex data to identify and separate different regions of interest from background noise.

Keywords: complex convolutional neural networks; deep learning; image reconstruction; in-phase/quadrature signal; ultrasound imaging.

Figure 1

Block diagram of the proposed complex max-pooling layer. Solid and dashed lines indicate real and imaginary parts, respectively.

Figure 2

Block diagram of the proposed complex up-sampling layer with a factor of 2. The solid and dashed lines indicate real and imaginary parts, respectively.

Figure 3

Block diagram of the proposed complex residual attention U-Net. $\tilde{x}$ is the low-quality (LQ) complex input and $\widehat{x}$ is the high-quality (HQ) reconstruction.

Figure 4

A sample from the phantom and training sets. The red area indicates a background region, the green area indicates an anechoic region, and the blue area indicates an hyperechoic region. (a) A computer-generated phantom used to simulate the dataset; (b) a B-mode image of low-quality IQ data acquired using single insonification, (c) a B-mode image of high-quality IQ data from standard compounding 20 steered acquisitions.

Figure 5

Three B-mode samples (a–c) from test datasets comparing different techniques. All results are from a single-plane wave insonification except for the standard compounding, which is obtained by coherently compounding 20 steered insonifications. Res-Att-UNet and C-Res-Att-UNet show visual improvements by separating different regions of noise.

Figure 6

B-mode results of Res-Att-UNet and C-Res-Att-UNet with 1PW using the PICMUS dataset [25] compared to DAS (1PW) and standard compounding (75 PWS). (a) In vitro CIRS phantom, (b,c) simulated, and (d) in vivo carotid longitudinal section.

Figure 7

Lateral profiles of (a) the in vitro Figure 6 at a 26 mm depth and (b) the simulated point target Figure 6 at a 33 mm depth. Depth is indicated with a dashed yellow line.