Brain Imaging Center, McLean Hospital, 115 Mill Street, Belmont, MA 02478, USA.
Neuroimage. 2012 Apr 15;60(3):1913-23. doi: 10.1016/j.neuroimage.2012.01.140. Epub 2012 Feb 9.
Confounding noise in BOLD fMRI data arises primarily from fluctuations in blood flow and oxygenation due to cardiac and respiratory effects, spontaneous low frequency oscillations (LFO) in arterial pressure, and non-task related neural activity. Cardiac noise is particularly problematic, as the low sampling frequency of BOLD fMRI ensures that these effects are aliased in recorded data. Various methods have been proposed to estimate the noise signal through measurement and transformation of the cardiac and respiratory waveforms (e.g. RETROICOR and respiration volume per time (RVT)) and model-free estimation of noise variance through examination of spatial and temporal patterns. We have previously demonstrated that by applying a voxel-specific time delay to concurrently acquired near infrared spectroscopy (NIRS) data, we can generate regressors that reflect systemic blood flow and oxygenation fluctuations effects. Here, we apply this method to the task of removing physiological noise from BOLD data. We compare the efficacy of noise removal using various sets of noise regressors generated from NIRS data, and also compare the noise removal to RETROICOR+RVT. We compare the results of resting state analyses using the original and noise filtered data, and we evaluate the bias for the different noise filtration methods by computing null distributions from the resting data and comparing them with the expected theoretical distributions. Using the best set of processing choices, six NIRS-generated regressors with voxel-specific time delays explain a median of 10.5% of the variance throughout the brain, with the highest reductions being seen in gray matter. By comparison, the nine RETROICOR+RVT regressors together explain a median of 6.8% of the variance in the BOLD data. Detection of resting state networks was enhanced with NIRS denoising, and there were no appreciable differences in the bias of the different techniques. Physiological noise regressors generated using Regressor Interpolation at Progressive Time Delays (RIPTiDe) offer an effective method for efficiently removing hemodynamic noise from BOLD data.
BOLD fMRI 数据中的混杂噪声主要源于血流和氧合的波动，这些波动是由心脏和呼吸效应、动脉血压的自发低频振荡 (LFO) 以及与任务无关的神经活动引起的。心脏噪声尤其成问题，因为 BOLD fMRI 的低采样频率确保了这些效应在记录的数据中被混淆。已经提出了各种方法来通过测量和转换心脏和呼吸波形来估计噪声信号（例如 RETROICOR 和呼吸容积/时间（RVT）），并通过检查空间和时间模式来进行无模型噪声方差估计。我们之前已经证明，通过对同时采集的近红外光谱 (NIRS) 数据应用特定于体素的时间延迟，可以生成反映全身血流和氧合波动效应的回归器。在这里，我们将该方法应用于从 BOLD 数据中去除生理噪声的任务。我们比较了使用从 NIRS 数据生成的各种噪声回归器集去除噪声的效果，还比较了 RETROICOR+RVT 的效果。我们比较了使用原始和过滤后的噪声数据进行静息状态分析的结果，并通过从静息数据计算零分布并将其与预期的理论分布进行比较，评估了不同噪声过滤方法的偏差。使用最佳的处理选择集，六个具有特定于体素的时间延迟的 NIRS 生成的回归器解释了大脑中 10.5%的方差中位数，其中灰质的降低最大。相比之下，9 个 RETROICOR+RVT 回归器一起解释了 BOLD 数据中 6.8%的方差中位数。使用 NIRS 去噪可以增强静息状态网络的检测，并且不同技术的偏差没有明显差异。使用渐进时间延迟回归器插值 (RIPTiDe) 生成的生理噪声回归器为从 BOLD 数据中有效去除血液动力学噪声提供了一种有效方法。
