POSTDOC: The Development of a 3D Parallel Finite Element Model for EEG and MEG Source Localization

大脑中信息处理的时间动态可以用脑电图（EEG）或脑磁图（MEG）来研究，但由于数学模型粗糙，这些技术的空间分辨率往往很差。利用有限元模拟头部和颗粒的真实几何形状和电导率，可以提高精度和分辨率。基于有限元方法的二维脑电图源定位代码已经开发出来，但将其扩展到三维逼真形状的头部模型将需要大幅提高计算速度。3D有限元模型中的源定位需要在脑电波的每个时间点上进行数百次迭代，从而解决一个大型方程组（超过100,000个），这可能持续几秒钟。因此，为了使这种计算可行，3D脑电图代码将被优化和并行化，以便在匹兹堡超级计算中心的CRAY超级计算机之一或匹兹堡大学医学中心的80个惠普工作站集群上运行。在第二年，将开发和并行化一个三维MEG源定位代码。这是一个直接和相对简单的任务，一旦3D脑电图代码并行化，因为脑电图和脑磁图都需要求解相同的方程组。该奖学金将通过参加正式课程、讲习班、研讨会和与计算机科学、数学、神经生理学和工程等多学科顾问的咨询，帮助研究助理拓宽在代码优化、代码并行化和神经科学方面的知识。该项目开发的代码将通过该领域提供，以帮助其他研究人员使用准确有效的源定位模型。

The temporal dynamics of information processing in the brain can be studied using electroencephalography (EEG) or magnetoencephalpgrahpy (MEG), but the spatial resolution of these techniques is often poor due to crude mathematical models. Accuracy and resolution can be improved by using finite elements to model the realistic geometry and electrical conductivity of the head and grain. A 2D EEG source localization code based on a finite element method has been developed, but extending this to the 3D realistic-shaped head model will require a huge increase in computational speed. Source localization in a 3D finite element model requires the solution of a large system of equations (more than 100,000) for hundreds of iterations for each time point in a brainwave that may last for seconds. Therefore, to make this computations practicable the 3D EEG code will be optimized and parallelized to run on one of the CRAY supercomputers at the Pittsburgh Supercomputing Center or on a cluster of 80 Hewlett-Packard workstations at the University of Pittsburgh Medical Center. During the second year, a 3D MEG source localization code will be developed and parallelized. This is a straightforward and relatively simple task once the 3D EEG code is parallelized, because the solution of the same system of equations is needed for both EEG and MEG. The fellowship will help broaden the research Associate's knowledge in code optimization, code parallelization and the neurosciences through participation in formal courses, workshops, seminars and consultation with a group of multi-disciplinary advisops in computer science, mathematics, neurophysiology and engineering. The code developed from this project will be made available through the domain to help other researchers utilize accurate and efficient source localization models.