Bayesian Electromagnetic Spatio-Temporal Imaging of Extended Sources Based on Matrix Factorization
Bayesian Electromagnetic Spatio-Temporal Imaging of Extended Sources Based on Matrix Factorization
复制标题
DOI：
10.1109/tbme.2018.2890291
复制
发表时间：
2019-09
影响因子：
4.6
通讯作者：
Ke Liu;Z. Yu;Wei Wu;Z. Gu;Jun Zhang;Ling Cen;S. Nagarajan;Yuanqing Li
中科院分区：
文献类型：
--
作者：
Ke Liu;Z. Yu;Wei Wu;Z. Gu;Jun Zhang;Ling Cen;S. Nagarajan;Yuanqing Li
从脑电图和脑磁图（E/MEG）中准确估计神经源的位置和范围是具有挑战性的，特别是对于深度和高度相关的神经活动。在这项研究中，我们提出了一种新的完全数据驱动的源成像方法，源成像基于时空基函数（SI-STBF），这是建立在贝叶斯框架，以解决这个问题。SI-STBF是基于源矩阵的因式分解作为稀疏编码矩阵和时间基函数（TBF）矩阵的乘积，其包括几个TBF。TBF的先验以经验贝叶斯的方式设置。类似地，对于空间约束，SI-STBF假设编码矩阵的先验协方差为若干空间协方差分量的加权和。通过变分贝叶斯推理，同时从E/MEG中学习TBF和编码矩阵。为了能够在高分辨率源空间上进行推理，我们使用凸分析推导出了一个可扩展的算法。使用模拟和实验E/MEG记录评估SI-STBF的性能。与$L_2$-范数约束方法相比，SI-STBF在重建扩展源时具有上级优势，且空间扩散小，定位误差小。由于对源矩阵进行了时空分解，对于高相关性和深源，SI-STBF比仅空间约束方法具有更高的估计精度。
Accurate estimation of the locations and extents of neural sources from electroencephalography and magnetoencephalography (E/MEG) is challenging, especially for deep and highly correlated neural activities. In this study, we proposed a new fully data-driven source imaging method, source imaging based on spatio-temporal basis function (SI-STBF), which is built upon a Bayesian framework, to address this issue. The SI-STBF is based on the factorization of a source matrix as a product of a sparse coding matrix and a temporal basis function (TBF) matrix, which includes a few TBFs. The prior of the TBF is set in the empirical Bayesian manner. Similarly, for the spatial constraint, the SI-STBF assumes the prior covariance of the coding matrix as a weighted sum of several spatial covariance components. Both the TBFs and the coding matrix are learned from E/MEG simultaneously through variational Bayesian inference. To enable inference on high-resolution source space, we derived a scalable algorithm using convex analysis. The performance of the SI-STBF was assessed using both simulated and experimental E/MEG recordings. Compared with $L_2$-norm constrained methods, the SI-STBF is superior in reconstructing extended sources with less spatial diffusion and less localization error. By virtue of the spatio-temporal factorization of source matrix, the SI-STBF also produces more accurate estimations than spatial-only constraint method for high correlated and deep sources.