Belardinelli P, Jalava A, Gross J, Kujala J, Salmelin R
O.V. Lounasmaa Laboratory, Brain Research Unit, Aalto University, Espoo, Finland,
Cogn Process. 2013 Nov;14(4):357-69. doi: 10.1007/s10339-013-0568-y. Epub 2013 Jun 1.
Over the past decade, various techniques have been proposed for localization of cerebral sources of oscillatory activity on the basis of magnetoencephalography (MEG) or electroencephalography recordings. Beamformers in the frequency domain, in particular, have proved useful in this endeavor. However, the localization accuracy and efficacy of such spatial filters can be markedly limited by bias from correlation between cerebral sources and short duration of source activity, both essential issues in the localization of brain data. Here, we evaluate a method for frequency-domain localization of oscillatory neural activity based on the relevance vector machine (RVM). RVM is a Bayesian algorithm for learning sparse models from possibly overcomplete data sets. The performance of our frequency-domain RVM method (fdRVM) was compared with that of dynamic imaging of coherent sources (DICS), a frequency-domain spatial filter that employs a minimum variance adaptive beamformer (MVAB) approach. The methods were tested both on simulated and real data. Two types of simulated MEG data sets were generated, one with continuous source activity and the other with transiently active sources. The real data sets were from slow finger movements and resting state. Results from simulations show comparable performance for DICS and fdRVM at high signal-to-noise ratios and low correlation. At low SNR or in conditions of high correlation between sources, fdRVM performs markedly better. fdRVM was successful on real data as well, indicating salient focal activations in the sensorimotor area. The resulting high spatial resolution of fdRVM and its sensitivity to low-SNR transient signals could be particularly beneficial when mapping event-related changes of oscillatory activity.
在过去十年中，已经提出了各种基于脑磁图（MEG）或脑电图记录来定位振荡活动脑源的技术。特别是，频域中的波束形成器已被证明在这项工作中很有用。然而，这种空间滤波器的定位精度和功效可能会受到脑源之间相关性偏差以及源活动持续时间短的显著限制，这两个都是脑数据定位中的关键问题。在这里，我们评估一种基于相关向量机（RVM）的振荡神经活动频域定位方法。RVM是一种用于从可能超完备数据集中学习稀疏模型的贝叶斯算法。我们将频域RVM方法（fdRVM）的性能与相干源动态成像（DICS）进行了比较，DICS是一种采用最小方差自适应波束形成器（MVAB）方法的频域空间滤波器。这些方法在模拟数据和真实数据上都进行了测试。生成了两种类型的模拟MEG数据集，一种具有连续的源活动，另一种具有瞬态活动源。真实数据集来自缓慢的手指运动和静息状态。模拟结果表明，在高信噪比和低相关性条件下，DICS和fdRVM具有可比的性能。在低信噪比或源之间高相关性的条件下，fdRVM的性能明显更好。fdRVM在真实数据上也取得了成功，表明在感觉运动区域有明显的局灶性激活。当绘制振荡活动的事件相关变化时，fdRVM所具有的高空间分辨率及其对低信噪比瞬态信号的敏感性可能特别有益。
