AAR-Net：用于声学异质介质光声图像重建的深度神经网络

PDF下载HTML

Title:

AAR-Net: a deep neural network for photoacoustic image reconstruction in heterogeneous acoustic media

作者:

孙美晨¹, 孙正^1,2, 候英飒¹

1. 华北电力大学 电子与通信工程系, 河北 保定 071003;
2. 华北电力大学 河北省电力物联网技术重点实验室, 河北 保定 071003

Author(s):

SUN Meichen¹, SUN Zheng^1,2, HOU Yingsa¹

1. Department of Electronic and Communication Engineering, North China Electric Power University, Baoding 071003, China;
2. Hebei Key Laboratory of Power Internet of Things Technology, North China Electric Power University, Baoding 071003, China

关键词:

图像重建; 图像增强; 光声光谱成像; 声学特性; 反射; 深度学习; 深度神经网络; 梯度方法

Keywords:

image reconstruction; image enhancement; photoacoustic spectroscopy; acoustic properties; reflection; deep learning; deep neural networks; gradient methods

分类号:

TP391.41；R445

DOI:

10.11992/tis.202212024

文献标志码:

2023-11-14

摘要:

在光声成像中，由于组织的吸收和扩散等引起的超声波衰减、由声速变化引起的相位偏差以及与声衰减相关的信号波形展宽都会降低图像的空间分辨率，针对该问题，提出一种基于深度学习的声学特性非均匀组织图像重建方法。通过将深度梯度下降（deep gradient descent，DGD）网络与U-Net相结合构建声伪影去除网络（acoustic artifacts removal network，AAR-Net）。DGD模块利用梯度信息减少非均匀声学特性对重建图像质量的影响，实现信号域到图像域的转换。U-Net模块实现对DGD模块输出的低质量图像的优化，实现图像域到图像域的转换。仿真、仿体和在体试验结果表明，与传统的非学习图像重建方法和最新的基于图像后处理的深度学习方法相比，采用该方法重建的图像结构相似度和峰值信噪比分别可提高约20%和10%。AAR-Net无需任何有关成像对象声学特性的先验知识，即可重建高质量图像。

Abstract:

Photoacoustic imaging suffers from degraded image quality owing to distorted and attenuated ultrasound waves propagating in an acoustic attenuating medium, phase deviation caused by changes in sound speed, and signal broadening related to acoustic attenuation. To address this issue, a deep learning method is proposed to reconstruct photoacoustic images of acoustically heterogeneous medium. A deep neural network is constructed, named acoustic artifacts removal network (AAR-Net) by combining deep gradient descent (DGD) network with U-Net. The DGD module aims to achieve the conversion from signal domain to image domain, which uses the gradient information to reduce the impact of heterogeneous acoustic properties on reconstructed image quality. U-Net module aims to optimize low-quality images output by the DGD module and realize the image-to-image conversion. The simulation, phantom and in vivo studies show that the proposed method outperforms traditional non-learning methods and the state-of-the-art post-processing based deep learning method. The image similarity and peak signal-to-noise ratio obtained by this method are improved by about 20% and 10%, respectively. AAR-Net enables reconstruction of high-quality images without any prior knowledge of acoustic properties of imaging objects.