基于生成对抗网络的注入电流式热声成像逆问题研究

The Study on the Inverse Problem of Applied Current Thermo-Acoustic Imaging Based on Generative Adversarial Network

注入电流式热声成像结合了电阻抗成像高对比度和超声成像高分辨率的优势,在生物医学成像领域具有广阔的应用前景。该成像方法激励效率高、检测信噪比强,但在较低频率的电磁激励下,重建目标体电导率的高分辨率图像仍然具有很大的挑战。该文提出一种基于生成对抗网络的电导率重建新方法,包含三个步骤:首先用维纳滤波反卷积,将超声探头输出的电信号还原为真实声信号;然后利用滤波反投影获得初始声源图像;最后将初始声源图像和电导率图像进行匹配,作为生成对抗网络的训练样本,构建用于电导率重建的深度学习模型。经理论分析与仿真研究发现,新方法通过引入深度神经网络,能够挖掘出蕴含在数据中的逆问题求解模型,进而重建高分辨率的电导率图像,且具有很强的抗干扰特性。新方法的提出为解决注入电流式热声成像的电导率重建问题提供了新思路。

Applied current thermo-acoustic imaging(ACTAI)has a prospect applications in medical imaging,with the advantages of high contrast of ecclectrical impedance tomography and high spatial resolution in ultrasound imaging.Although the imaging modality has high excitation efficiency and signal-to-noise ratio,there is still a great challenge on high resolution reconstruction of conductivity under low frequency electromagnetic excitation.In this paper,a new method for reconstructing the conductivity based on generative adversarial network is proposed.The proposed algorithm consists of the following three main steps:firstly,by using Wiener filtering deconvolution method,the original acoustic fields on the detected boundary are reconstructed by the acoustic signals originated from the ultrasonic transducers.And then the initial acoustic source image is obtained by the filtering back projection method.Finally,the initial acoustic source image are used as training samples and labels of a deep learning network,which is designed for the conductivity reconstruction.Theoretical analysis show that the method proposed in this paper can solve the inverse problem of the conductivity reconstruction by the machine learning models and obtain the accurate and stable images.This provides a new idea for solving the problem of conductivity reconstruction in the applied current thermo-acoustic imaging.