基于Deep CG的肺部电阻抗成像方法  

Reconstruction of Thorax Image Based on Deep CG Method for Electrical Impedance Tomography

作　　者：王子辰 付荣 张新宇 王迪 陈晓艳[1] Wang Zichen;Fu Rong;Zhang Xinyu;Wang Di;Chen Xiaoyan(College of Electronic Information and Automatic,Tianjin University of Science and Technology,Tianjin 300222,China)

机构地区：[1]天津科技大学电子信息与自动化学院,天津300222

出　　处：《中国生物医学工程学报》2023年第2期148-157,共10页Chinese Journal of Biomedical Engineering

摘　　要：针对电阻抗图像重建空间分辨率不足问题,基于深度学习理论提出一种共轭梯度快速预重建与深度堆栈式自编码器后处理的电阻抗成像方法(Deep CG)。该方法的核心思想是:融合数值重建算法与深度学习算法,使胸腔内肺部的结构和电导率分布更加精准。首先采用共轭梯度算法进行图像预重建,获得边界电压与胸腔内部电导率分布的预映射关系;再采用深度堆栈式自编码器,将编码和解码层级连接,充分利用不同空间特征信息,实现特征提取和图像重建;最后根据公开的80名临床患者的CT结构图像构建了数据集,采用混合式监督训练方法调参,不仅避免了深度网络中信息流和梯度流弥散问题,而且优化了算法模型。采用图像相对误差、相关系数进行量化指标评价,并与常用的数值图像重建算法和全连接神经网络模型进行对比。结果显示,Deep CG算法的比常用图像重建算法图像和全连接神经网络模型相对误差从0.50和0.24降低到0.11,相关系数由0.80和0.90提高到0.96。该方法获得了空间分辨率高,尤其边界更清晰的电阻抗图像,有望进一步推动EIT技术在临床的应用研究。To improve the spatial resolution of electrical impedance tomography,a novel method based on conjugate gradient(CG)with rapid pre-reconstruction and deep stack autoencoder post-processing was proposed(Deep CG).The core idea is that the merging of numerical reconstruction algorithm with deep learning-based method makes the structure and conductivity distribution of the thorax more accurate.Firstly,the mathematical reconstruction algorithm CG was adopted to pre-reconstruct the coarse image,and the mapping between boundary voltage and conductivity distribution in the chest was achieved.Next,to take full advantages of the different spatial features,the stack autoencoder was employed to connect the encoding and the decoding modules hierarchically,which realized the feature extraction(FE)and the image reconstruction(IR).Finally,to train the model,the dataset was constructed from the number of 400 clinical CT slices,a mixed supervised method was employed to adjust the model parameters,which not only avoided the dispersion of the information flow and gradient flow,but also optimized the parameters of Deep CG method.The relative error(RE)and correlation coefficient(CC)were adopted to evaluate the image quantitively.The images were compared to the traditional numerical algorithm and a full connected neural network.The results showed that the RE was decreased to 0.11 from 0.5 and 0.24,and the CC was improved to 0.96 from 0.8 and 0.9.The proposed method was able to reconstruct EIT images with higher spatial resolution and clear boundary,which is expected to put forward EIT techniques to the further applications and researches.

关 键 词：电阻抗成像 堆栈式自编码器 深度学习 图像重建 共轭梯度 

分 类 号：R318[医药卫生—生物医学工程]