Time Sequence Learning for Electrical Impedance Tomography Using Bayesian Spatiotemporal Priors
Time Sequence Learning for Electrical Impedance Tomography Using Bayesian Spatiotemporal Priors
复制标题
DOI：
10.1109/tim.2020.2972172
复制
发表时间：
2020-09-01
影响因子：
5.6
通讯作者：
Ji, Jiabin
中科院分区：
文献类型：
--
作者：
Liu, Shengheng;Cao, Ruisong;Ji, Jiabin
作为一种新兴的连续监测有界域的技术，电阻抗断层成像(EIT)在各种应用中得到了越来越广泛的应用。尽管取得了前所未有的进展，但现阶段的EIT逆解算器还不能保证足够的保真度以及对内阻抗动力学的有效研究。在此背景下，本文提出了一种时空结构感知的稀疏贝叶斯学习(SA-SBL)框架，用于求解时间连续的EIT逆问题。具体地说，在EIT时间序列的重建过程中，利用分层贝叶斯模型和结构感知先验，在无监督的情况下探索和利用了帧内空间聚类和帧间时间连续性。建立了多测量向量模型来捕捉时空相关性，并描述了潜在的多维重建问题。通过将近似消息传递应用于期望更新，有效地解决了由此产生的大尺度反演问题。与原SA-SBL算法相比，加速比达到O(N-2/M)。仿真结果表明，与已有方法相比，该算法具有更好的重建性能，量化指标的评价分数至少提高了17%。由于改进了图像质量和恢复效率，因此该算法具有更广泛的适用性。
As an emerging technology for continuous monitoring of a bounded domain, electrical impedance tomography (EIT) gains increasing popularity in various applications. Despite unprecedented progress, the EIT inverse solvers at the present stage are incompetent to guarantee sufficient fidelity as well as efficient investigation of the internal impedance dynamics. In this context, this article introduces a spatiotemporal structure-aware sparse Bayesian learning (SA-SBL) framework for solving the time-continuous EIT inverse problems. Specifically, in the process of reconstructing the EIT time sequence, both intraframe spatial clustering and interframe temporal continuity are explored and exploited in an unsupervised manner by using the hierarchical Bayesian model and structure-aware priors. A multiple measurement vector model is established to capture the spatiotemporal correlations and describe the underlying multidimensional reconstruction problem. The resultant largescale inversion is efficiently solved by applying the approximate message passing to the expectation updating. A speedup ratio of O(N-2/M) is achieved compared with original SA-SBL. Simulation results indicate that the proposed algorithm exhibits superior reconstruction performance to the existing methods, where the scores evaluated by the quantitative metrics are improved by at least 17%. The presented algorithm is envisioned to offer broader applicability since it yields improved image quality and recovery efficiency.