“Beamforming is the most common class of algorithms for adaptive distributed-source imaging that is used in EEG and MEG” (Mike X,Analyzing Neural Time Series Data p. 313) 波束形成是应用于EEG和MEG的自适应分布式源成像最常见的一类算法。

“The main advantage of adaptive spatial filters such as beamforming is, as the name suggests, that the weights are adapted to the data. This means that the weights can change over time, frequency, condition, and subject, according to changes in patterns of covariance in the data. This provides increased sensitivity for detecting subtle features of the results and provides a possible advantage over electrode-level recordings. For example, an effect with low amplitude might not be detected by electrode-level statistics, but if that low-amplitude activity is spatially coherent in a manner consistent with a certain brain source, an adaptive spatial filter may isolate this activity pattern and thus make it more visible in the data. Simulation studies suggest that beamforming provides results with higher accuracy compared to other source-imaging methods and that beamforming has the least amount of overestimation of the spatial extent of the activation” (Mike X,Analyzing Neural Time Series Data p. 313) 自适应空间滤波器如波束形成的主要优点是，顾名思义，权重是适应数据的。这意味着权重可以根据数据中协方差模式的变化而随时间、频率、条件和主体而变化。这为检测结果的细微特征提供了更高的灵敏度，并提供了相对于电极级记录的可能优势。例如，低幅值的效应可能无法通过电极水平的统计检测到，但是如果低幅值的活动在空间上是以与某个脑源一致的方式相干的，那么自适应的空间滤波器可能会隔离这种活动模式，从而使其在数据中更加可见。仿真研究表明，与其他源成像方法相比，波束形成提供了更高的精度，并且波束形成对激活的空间范围的高估量最少

“The main disadvantages of adaptive distributed source methods are the number of parameters that must be set and the consequences those parameters can have on the results. How much you have to consider these choices depends on what the goal of the analysis is, and on whether you write your own beamformers or use a toolbox or software package that uses predefined default values.” (Mike X,Analyzing Neural Time Series Data p. 313) 自适应分布式源方法的主要缺点是必须设置的参数数量以及这些参数对结果的影响。你需要考虑这些选择的程度取决于分析的目标是什么，取决于你是否编写自己的波束形成器或使用使用预定义默认值的工具箱或软件包。

“Options include the algorithm to use, whether and which frequencies to analyze, whether to use time-domain covariances or frequency-domain cross-spectral densities, the length of the time window for computing the covariances and whether that window should change as a function of frequency, whether and how to regularize the covariance matrix, whether to compute weights based on all conditions or for each condition, what to use as a normalization baseline, how many voxels to estimate and where they are located, what type of forward model to use (and, for EEG, what values to use for skull and scalp conductances), whether to fix dipole orientations with respect to the cortex or use three cardinal orientations, and, if using three orientations, how to estimate a voxel’ s activity from projections onto the three orientations, and so on. The number of parameters and choices should not dissuade you from applying adaptive spatial filters, but it is important to consider the parameters of the analyses and their possible influences on the results.” (Mike X,Analyzing Neural Time Series Data p. 314) 包括使用的算法，是否和哪些频率进行分析，是否使用时域协方差或频域互谱密度，计算协方差的时间窗口的长度以及该窗口是否应该随着频率的变化而变化，是否和如何正则化协方差矩阵，是否根据所有条件计算权重或者对于每个条件，使用什么作为标准化基线，估计多少体素和它们位于哪里，使用(对于EEG ,颅骨和头皮电导有什么价值?)的前向模型是什么类型，是固定偶极子相对于皮层的方向还是使用三个基数方向，如果使用三个方向，如何估计一个体素从投影到三个方向的活动，等等。参数的数量和选择不应该阻止你应用自适应空间滤波器，但重要的是要考虑分析的参数及其对结果的可能影响。