Low Frequency Systemic Hemodynamic "Noise" in Resting State BOLD fMRI: Characteristics, Causes, Implications, Mitigation Strategies, and Applications

Abstract

Advances in functional magnetic resonance imaging (fMRI) acquisition have improved signal to noise to the point where the physiology of the subject is the dominant noise source in resting state fMRI data (rsfMRI). Among these systemic, non-neuronal physiological signals, respiration and to some degree cardiac fluctuations can be removed through modeling, or in the case of newer, faster acquisitions such as simultaneous multislice acquisition, simple spectral filtering. However, significant low frequency physiological oscillation (∼0.01-0.15 Hz) remains in the signal. This is problematic, as it is the precise frequency band occupied by the neuronally modulated hemodynamic responses used to study brain connectivity, precluding its removal by spectral filtering. The source of this signal, and its method of production and propagation in the body, have not been conclusively determined. Here, we summarize the defining characteristics of the systemic low frequency noise signal, and review some current theories about the signal source and the evidence supporting them. The strength and distribution of the systemic LFO signal make characterizing and removing it essential for accurate quantification, especially for resting state connectivity, when no stimulation can be compared with the signal. Widespread correlated non-neuronal signals obscure and distort the more localized patterns of neuronal correlations between interacting brain regions; they may even cause apparent connectivity between regions with no neuronal interaction. Here, we discuss a simple method we have developed to parse the global, moving, blood-borne signal from the stationary, neuronal connectivity signals, substantially reducing the negative correlations that result from global signal regression. Finally, we will discuss some of the uses to which the moving systemic low frequency oscillation can be put if we consider it a "signal" carrying information, rather than simply "noise" complicating the interpretation of resting state connectivity. Properly utilizing this signal may offer insights into subtle hemodynamic alterations that can be used as early indicators of circulatory dysfunction in a number of neuropsychiatric conditions, such as prodromal stroke, moyamoya, and Alzheimer's disease.

Keywords: cerebrovascular reactivity; denoising; low frequency oscillation; noise modeling; physiological noise; physiological noise modeling; vascular mapping.

FIGURE 1

Power spectrum (left) and time domain data (right) presented in different spectral bands, from a voxel in a resting state data (TR = 0.4 s) of one participant. Three distinct spectral ranges corresponding to different physiological processes were marked. The spectral area captured by various TR values is also depicted on the power spectrum. The right hand panel shows a BOLD timecourse from a resting state fMRI scan (TR = 0.4 s) without filtration (A, in red) and its band-passed versions in panel (B) 0.01–0.2 Hz; (C) 0.3–0.4 Hz; and (D) 0.8–1.0 Hz (in blue) (Figure adapted from Tong and Frederick, 2014a).

FIGURE 2

Independent components (1–7) from a group analysis of 10 subjects’ resting state data that have high, significant positive correlations with simultaneously recorded peripheral NIRS data (Figure adapted from Tong et al., 2015).

FIGURE 3

Synthetic data consisting of progressively delayed sum of sinusoids was placed inside two identical blocks (a). The red arrows indicate the direction of the moving wave (increasing time delay). The examples of moving waves at the circles (1–3) are shown in panel (b). Six independent components resulting from ICA are shown in panel (c) with the corresponding color bars (Figure adapted from Tong et al., 2015).

FIGURE 4

Results from group ICA on 11 subjects’ real BOLD data were shown in panel (A). Results from group ICA on 11 subjects’ synthetic data were shown in panel (B). The value in each result showed the spatial correlation coefficient calculated between that component and the corresponding RSN from the template (Beckmann et al., 2005). The two components in the red block are the same (Figure adapted from Tong et al., 2015).

FIGURE 5

A schematic representation of the RIPTiDe regressor refinement procedure (Figure reproduced from Erdogan et al., 2016).

FIGURE 6

The effect of static and dynamic global signal regression on group level connectivity strengths from a posterior cingulate cortex (PCC) seed to ROIs in panel (A) major default mode network (DMN) ROIs (task negative regions), (B) Task positive network (TPN) ROIs, (C) and reference regions thought not to be involved in either network. L, left hemisphere; R, right hemisphere; VIS, visual cortex ROI. Both static and dynamic global signal regression remove spurious connectivity within the DMN (panel A), while preserving the expected anticorrelations with regions of the task positive network (panel B). Spurious positive correlations with unrelated reference regions were eliminated with both types of regression; however this came at the cost of large, significant spurious anticorrelations using static GSR, but not with dGSR (panel C) (Figure reproduced from Erdogan et al., 2016).

FIGURE 7

Blood arrival time delay values (in seconds) obtained from rapidtide analysis of (A) resting state fMRI data and from (B) dynamic susceptibility contrast imaging during the same imaging session in healthy controls (N = 8) (Figure adapted from Tong et al., 2017).

FIGURE 8

Averaged correlation parameters (lag time of maximum correlation and maximum correlation value) for 487 subjects from the 500 subjects release of the Human Connectome Project data. Each subject had four scans (LR and RL phase encode in two sessions, REST1 and REST2) (Figure adapted from Frederick et al., 2017).