Partitioning of Physiological Noise Signals in the Brain with Concurrent Near-Infrared Spectroscopy and fMRI

Abstract

The blood–oxygen level dependent (BOLD) signals measured by functional magnetic resonance imaging (fMRI) are contaminated with noise from various physiological processes, such as spontaneous low-frequency oscillations (LFOs), respiration, and cardiac pulsation. These processes are coupled to the BOLD signal by different mechanisms, and represent variations with very different frequency content; however, because of the low sampling rate of fMRI, these signals are generally not separable by frequency, as the cardiac and respiratory waveforms alias into the LFO band. In this study, we investigated the spatial and temporal characteristics of the individual noise processes by conducting concurrent near-infrared spectroscopy (NIRS) and fMRI studies on six subjects during a resting state acquisition. Three time series corresponding to LFO, respiration, and cardiac pulsation were extracted by frequency from the NIRS signal (which has sufficient temporal resolution to critically sample the cardiac waveform) and used as regressors in a BOLD fMRI analysis. Our results suggest that LFO and cardiac signals modulate the BOLD signal independently through the circulatory system. The spatiotemporal evolution of the LFO signal in the BOLD data correlates with the global cerebral blood flow. Near-infrared spectroscopy can be used to partition these contributing factors and independently determine their contribution to the BOLD signal.

Keywords

BOLD signal fMRI low-frequency oscillation near-infrared spectroscopy physiological noise resting state

Introduction

Functional magnetic resonance imaging (fMRI) has been widely used in studies of neuronal activations in the brain (Bandettini et al, 1992; Ogawa et al, 1992). However, the blood–oxygen level dependent (BOLD) signal measured by fMRI is not only influenced by neuronal activation through neurovascular coupling, but also by other physiological processes, such as respiration and heartbeat, and autonomic blood pressure variations. This physiological noise can obscure neuronal activity by aliasing into the same temporal frequencies and by occupying the same anatomical brain regions as the target activation. Therefore, it is important to understand how these physiological processes influence the fMRI time series. There are three predominant physiological processes that directly affect the BOLD signal: (1) spontaneous low-frequency oscillations (LFOs), defined as the signal between 0.01 and 0.1 Hz, with unknown origin; (2) respiration (around 0.2 to 0.3 Hz); and (3) cardiac pulsation (around 1 Hz).

The sampling rate of fMRI typically ranges between 0.3 and 0.6 Hz, so the effects of respiration and cardiac pulsations in the fMRI data are aliased. Various methods for characterizing and removing these signals from the fMRI data have been studied extensively (Birn et al, 2008; Glover et al, 2000). Some studies have used external physiological recording (chest belt, end-tidal CO₂) with high sampling rate to identify the time courses of the processes likely to affect the BOLD signals (Birn et al, 2006; Cole et al, 2010; Glover et al, 2000). These measurements do not examine blood parameters directly, and are at best an indirect measure of the corresponding cerebral blood changes that affect fMRI signals, due to the complicated hemodynamic responses that affect both the shapes and delays in the BOLD waveform. Some studies have concentrated on using shorter repetition time (TR) (500 milliseconds) to avoid the aliasing problem altogether (Frank et al, 2001), but this requires a significant reduction in the spatial resolution or extent of brain coverage of the fMRI data.

Low-frequency oscillation is a dominant signal in BOLD fMRI, especially in the case of resting state connectivity studies. There appear to be multiple sources of variation in this frequency range, not all of which are understood. There have been many studies aimed at understanding the biological mechanism of the LFO in BOLD fMRI (Fox and Raichle, 2007; Katura et al, 2006; Obrig et al, 2000; Yan et al, 2009). It has been shown that scanner instability might contribute to it (Smith et al, 1999; Zarahn et al, 1997). Other studies provide evidence suggesting that the respiratory and cardiac pulsation effects contribute to the LFO through aliasing (Bhattacharyya and Lowe, 2004; Birn et al, 2006, 2008; Chang et al, 2009; Chang and Glover, 2009; Shmueli et al, 2007; van Buuren et al, 2009). Another contributor to the LFO may be neuronal signal from ‘resting state functional connectivity’ networks (Damoiseaux et al, 2006; Fox and Raichle, 2007) that are presumed to arise from spontaneous neuronal events (at frequencies below 0.1 Hz) that are synchronized in time between distant brain regions, leading to fluctuations in metabolic demands in the resting brain (Balduzzi et al, 2008; Fransson, 2005; Fukunaga et al, 2008; Suckling et al, 2008; Wink et al, 2008; Zuo et al, 2010). Since these fluctuations occur in the same frequency band as the LFO, the relationship to physiological noise is unclear.

Functional near-infrared spectroscopy (NIRS) is a complementary brain imaging technique to fMRI. Like fMRI, it is sensitive to blood-related changes, rather than directly to neuronal activity; unlike fMRI, it has very high temporal resolution. Concurrent NIRS and fMRI studies show high temporal correlations between the measured signals: Δ[HbO] (or Δ[Hb]) in NIRS, and BOLD signal in fMRI, in response to tasks (Sassaroli et al, 2006; Strangman et al, 2002; Toronov et al, 2003). In our previous concurrent studies, we have found that the spatiotemporal evolution of correlations between the signals in the LFO band in the brain suggest that this signal is closely associated with cerebral blood flow (CBF) (Tong and Frederick, 2010). However, we did not rule out the possibility that the LFO signal arose from aliased respiratory and/or heartbeat signals, neither did we study the temporal and spatial effects of respiration and heartbeat.

The aim of the current study is to separate three physiological signals, namely LFO, respiration, and cardiac pulsation, and examine their individual spatial and temporal effects on the BOLD signal during a resting state fMRI study in isolation. The high temporal resolution of NIRS (12.5 Hz) makes it possible to extract the individual physiological signals based on their spectral signatures without aliasing (Greve et al, 2009).

Materials and methods

Protocol

Details of the protocol can be found in our previous work (Tong and Frederick, 2010). In short, concurrent NIRS and fMRI resting state studies were conducted on six healthy volunteers during a resting state acquisition (4 M, 2F, average age 28 years).

All MR data were acquired on a Siemens TIM Trio 3T scanner (Siemens Medical Systems, Malvern, PA, USA) using a 12-channel phased array head matrix coil (260 time points, TR/TE=1500/30 milliseconds, flip angle 75°, matrix=64 × 64 on a 220 × 220 mm² field of view, 29 3.5 mm slices with 0.5 mm gap parallel to the Anterior-Posterior Commissure line extending down from the top of the brain). Physiological waveforms (pulse oximetry and respiratory depth) were recorded using the scanner's built-in wireless fingertip pulse oximeter and respiratory belt. After the completion of the structural and the functional acquisition, 3D velocity encoded phase contrast images were acquired to map the blood vessels in the head.

An MRI compatible NIRS optical probe, with three collection and two illumination fibers and source-detector distances of 1 or 3 cm, was placed over the right prefrontal area of each participant (roughly between Fp1 and F7 in the 10 to 20 system) and held in place by an elastic band around the head. The position of the probe was chosen due to its easy accessibility (no hair) and relatively short distance from the scalp to the cortex (11 to 13 mm on average) in this area (Okamoto et al, 2004). The sampling rate of NIRS data acquisition was 12.5 Hz. The fMRI data were collected for 6 minutes 30 seconds; NIRS data were recorded continuously during this time as well as for several minutes before and after the fMRI acquisition.

Near-Infrared Spectroscopy Regressor Preparation

Each pair of raw NIRS time courses (690 and 830 nm data) were converted into two time courses representing temporal changes of the concentrations of oxy-hemoglobin (Δ[HbO]) and deoxy-hemoglobin (Δ[Hb]) according to the differential path length factor method (Matcher, 1994) using a Matlab program (The Mathworks, Natick, MA). Figure 1 shows the time courses of Δ[HbO] and Δ[Hb] (measured from one of the NIRS 3 cm source-detector pairs) in Figures 1A and 1B and their spectra in Figure 1C from subject 3. From Figures 1A and 1B, the temporal trace of Δ[HbO] has a much higher signal-to-noise ratio than that of the Δ[Hb] (four- to sixfold bigger in the heartbeat signal). Figure 1C shows a comparison of the spectra of Δ[HbO] and Δ[Hb]. The signals from Δ[HbO] showed much stronger signals in the spectral bands of interest, including the LFO band (<0.1 Hz), respiration band (around 0.2 Hz), and heartbeat band (around 1 Hz). This pattern was observed in all six subjects.

Figure 1

Temporal traces (and their power spectra) of the near-infrared spectroscopy (NIRS) data collected from one source-detector pair (3 cm) on subject 3. (A) Temporal traces of Δ[HbO] (gray) and Δ[Hb] (black) calculated from the NIRS recording. (B) Enlarged section of (A) indicated by the black block. (C) Spectra from Δ[HbO] (gray) and Δ[Hb] (black).

While it is true that BOLD is sensitive to the Δ[Hb], it is also sensitive to the changes in cerebral blood volume and CBF (Buxton et al, 1998). The focus of this study is the effects of blood-related physiological variations caused by heartbeat, respiration, and LFO. These fluctuations primarily affect the cerebral blood volume and CBF, which are better detected using the Δ[HbO] waveform; moreover, the effects of global blood–oxygen saturation variations over time are also captured in this signal. The sensitivity of BOLD to Δ[Hb] is most important in detecting local hemodynamic changes due to neural activation, which in this study is not of interest. The Δ[HbO] obtained by NIRS has been demonstrated to capture the BOLD variations we are looking for with greater statistical significance than the Δ[Hb] time course (Tong and Frederick, 2010; Vernieri et al, 2004). Therefore, to correlate accurately between the NIRS signal and BOLD, the temporal trace of Δ[HbO], instead of Δ[Hb], from NIRS was used for further analysis.

The temporal trace of Δ[HbO] was first spectrally separated into four bands using a zero delay Fourier domain bandpass filter in Matlab. The four bands were defined as very-LFO (VLF), VLF<0.01 Hz; LFO, 0.01<LFO<0.1 Hz; respiration (Resp), 0.2<Resp <0.6 Hz; cardiac pulsation (Card), >0.8 Hz. The temporal signal from each band was then resampled at every 1.5 second according to the TR of the fMRI acquisition. Figure 2 shows the process in detail for subject 3. The left panel of Figure 2 shows the power spectra of original data (A), VLF (B), LFO (C), Resp (D), and Card (E). The middle panel displays the time traces of the original data (A), VLF (B), LFO (C), Resp (D), and Card (E) corresponding to left panel. The right panel of Figure 2 displays the enlarged section of the corresponding time traces (gray areas) in the middle panel with the resampled points (separated by 1.5 seconds) shown as closed circles. The resulting four time traces were used as regressors in a Generalized Linear Model (GLM) analysis of the fMRI data. Since the blood-related signal was detected with NIRS at the forehead, we expected that this same signal would also travel through out the brain, and arrive in each voxel with variable latency. To capture all the voxels that correlated with the NIRS signal at any latency, we applied a range of time shifts to the NIRS signal before filtering and resampling. A positive time shift value corresponds to events that happened before the time the NIRS data were recorded. We considered 55 time shift values covering the range of −4.32 to +4.32 seconds with respect to the unshifted signal with steps as small as 0.16 seconds. The number of time shifts were determined as the result of a preliminary analysis to find the range over which significant correlations were observed.

Figure 2

(A) Power spectra and temporal traces of Δ[HbO] collected on subject 3. (B) Very low-frequency oscillation (VLF) component of Δ[HbO] is represented in spectral domain (left), in time domain (middle). Enlarged section of the temporal trace (indicated by the gray area) with resampled data marked with circles (right). Similar plots are shown for low-frequency oscillation (LFO) (C), respiration (D), cardiac pulsation, (E) components of Δ[HbO].

Combined Analysis

For each subject, both NIRS and fMRI data were analyzed using FEAT, part of the FSL analysis package (FMRIB Expert Analysis Tool, v5.98; http://www.fmrib.ox.ac.uk/fsl, Oxford University, UK) (Smith et al, 2004). Details can be found in our previous work (Tong et al, 2011).

This image analysis was repeated 55 times at sequential time shifts. In each analysis, the time-shifted NIRS Δ[HbO] time course was split into four spectral bands, which were then downsampled to the fMRI TR to generate four regressors. The resulting thresholded z-statistic maps of each frequency band (VLF, LFO, Resp, and Card) were concatenated over all time lags in sequence to determine the evolution of the pattern of the frequency-specific (Δ[HbO]-correlated) regions over time. The maximum z value along the time course of each voxel in the concatenated z-statistic maps of each spectral band was used to form a new 3D thresholded z-statistic map (referred to here as a ‘max z-statistic map’).

Results

Near-Infrared Spectroscopy Signal

From the spectra of the NIRS Δ[HbO] and the data from the respiration belt acquired from all subjects, we observed strong cardiac and LFO signals of similar magnitude. However, respiration signals were only clearly observed in three subjects, and with much smaller magnitudes, indicating that the Δ[HbO] signal at the respiration frequency is much smaller than that caused by the other two factors, at least at the location where the NIRS probe was placed. The heart rate of subject 6 was lower than that of the other subjects, thus the bandpass filter range was modified for this subject to produce similar spectral signal components (VLF: <0.01; LF: 0.01 to 0.1; Resp: 0.2 to 0.5; Card: >0.6).

Dynamic Patterns

Figure 3 shows the thresholded z-statistic maps obtained by using Δ[HbO] regressors with various time shifts (indicated by the number below each z-statistic map) for subject 3. The z-statistic maps in Figure 3A correspond to the Δ[HbO] regressors in the spectral range of LFO, while those in Figure 3B correspond to the Δ[HbO] regressors in the range of heartbeat. Figure 3A shows the patterns of activation that highly correlated with NIRS regressors delayed in time from −4.32 to +4.32 seconds in 1.44 seconds increments, which we discussed in detail in a previous paper (Tong and Frederick, 2010). Here, we refined and extended those results by removing the respiratory and heart rate components (which will alias) from the LFO NIRS signal before resampling. In Figure 3A, when viewed in sequence (optimally as a movie), the dynamic pattern outlined by the time-delayed NIRS z-statistic maps resembles the blood flow through the brain, predominantly in the gray matter and the veins, although not clearly in the arteries. The time over which the pattern moves through the whole brain (from appearance to disappearance) is about 8 seconds, which is consistent with cerebral circulation (Crandell et al, 1973).

Figure 3

The thresholded z-statistic maps obtained by using resampled Δ[HbO] regressors of certain time shifts (indicated by the number below each z-statistic map) for subject 3. The z-statistic maps in panel A corresponded to the Δ[HbO] regressors in the spectral range of low-frequency oscillation (LFO), while that in panel B corresponded to the Δ[HbO] regressors in the cardiac range. The maps were arranged by their corresponding time shifts to assess the dynamic patterns.

Resampling the cardiac and respiratory signals to the fMRI TR will cause aliasing in the resulting signals; however, by resampling the actual waveform at the appropriate sample rate and delay, this procedure generates signals that are aliased in the same way as those resulting from the fMRI acquisition process. This is a distinct advantage of using the critically sampled NIRS regressor rather than trying to estimate these waveforms from the undersampled MR data.

Using the aliased heartbeat signal of Δ[HbO] and its temporal shifts as regressors generates z-statistic maps that suggest another dynamic pattern of blood flow repeating itself about every second. One of the cycles is presented in Figure 3B, which shows the z-statistic maps obtained by using the NIRS Δ[HbO] regressor (at cardiac frequency) time shifted by 0.16 seconds across −0.48 to +0.48 seconds, as indicated beneath each map. Since the acquisition rate of NIRS was 12.5 Hz (0.08 s/sample), the time shift can be as small as 0.08 seconds. In this case, 0.16 seconds was sufficient. From the maps, we observed that the first z-statistic map (−0.48 seconds shift) has a similar pattern to that of the last one (+0.48 seconds), indicating that the dynamic pattern shown in Figure 3B repeats itself temporally at a frequency of about 1 Hz. Extending the time shifts from −4.32 to 4.32 seconds further supported this idea; the pattern repeated itself about eight times (data not shown).

Blood Map

Maximum z-statistic maps were generated to include all the voxels that had highly significant correlations (z>3). Figure 4 compares the max z-statistic maps of subject 3 obtained by using Δ[HbO] regressors of different spectral ranges (LFO in Figure 4B; respiratory in Figure 4C; cardiac in Figure 4D) with the phase contrast MRI map of the blood vessels from the same subject (Figure 4A). In the LFO max z-statistic map (Figure 4B), activated voxels (in LFO) appear mostly in gray matter and veins, when compared with the phase contrast map (Figure 4A). However, the extensive arterial network at the circle of Willis, which was obvious in the phase contrast map could not be seen in the LFO map.

Figure 4

Comparison between the blood vessel map and maximum z-statistic results obtained from subject 3. (A) The blood vessel map obtained from phase contrast magnetic resonance imaging (MRI). (B–D) The z-statistic maps of the brain as results of using near-infrared spectroscopy (NIRS) Δ[HbO] of different spectral bands, low-frequency oscillation (LFO) in (B), respiration in (C), and heartbeat in (D) as regressors. Images are z-statistic map overlaid on the high-resolution scan transformed into standard (MNI152) space. The left hemisphere of the brain corresponds to the right side of the image.

The respiratory max z-statistic map (Figure 4C) is difficult to interpret. The map showed some activation near the eye sockets as well as on the surface of the brain, especially toward the top. While the map resembles surface vasculature, and vessels can be seen in the area of the insula, the amplitude of the respiratory NIRS signal is quite low, and we suspect contamination with a motion artifact makes it difficult to make firm conclusions.

Among the three max z-statistic maps, Figure 4D, which was generated using a NIRS signal from the cardiac spectral band as a regressor, displayed the clearest resemblance to the phase contrast blood vessel map in Figure 4A, including arteries and veins. Visible arteries include the basilar artery, posterior cerebral artery, middle cerebral artery, internal carotid artery, anterior cerebral artery, and circle of Willis. The visible veins include the superior sagittal sinus, transverse sinus, straight sinus, and internal cerebral vein. A robust heartbeat signal was detected in some of the ventricles, such as the fourth ventricle and lateral ventricles, in addition to blood vessels.

Figure 5 shows the averaged results of six subjects displayed in the same order as the single subject results in Figure 4, except Figure 5C, which is the averaged result of three subjects whose spectra showed signal in the respiratory frequency band. The results were obtained by registering each subject's structural map, blood map, and z-statistic maps to the MNI152 standard head (Evans, 1993) using FLIRT in FSL (Smith et al, 2004) and then taking the average max z-statistic across subjects for each voxel. Average max z-statistic maps over all subjects thresholded at z>3 showed spatial features similar to those shown in the single subject analysis shown in Figure 4. Also, in the average LFO map (Figure 5B), brain regions prominent in several established resting state networks were also observed (Damoiseaux et al, 2006). To highlight this observation, the sagittal, coronal, and axial images of the LFO map (Figure 5B) are shown in Figure 6. In Figure 6A, peristriate area, and lateral and superior occipital gyrus (Brodmann area (BA) 19) are clearly shown to have high maximum LFO z-statistics, as well as dorsal anterior cingulate (BA 32), superior parietal cortex (BA 7). In Figure 6B, middle frontal and orbital cortices (BA 6/9/10), superior parietal cortex (BA 7/40), middle temporal gyrus (BA 21), and posterior cingulate (BA 23/31) were also shown.

Figure 5

Averaged blood vessel map and averaged maximum z-statistic results in various frequency bands for comparison. Data were obtained from six subjects, except in (C), where the averaging was performed on the three subjects whose respiration components were visible. (A) The averaged blood vessel map obtained from phase contrast magnetic resonance imaging (MRI). (B–D) The averaged maximum z-statistic maps of the brain as results of using near-infrared spectroscopy (NIRS) Δ[HbO] of different spectral bands, low-frequency oscillation (LFO) in (B), respiration in (C), and cardiac pulsation in (D) as regressors. Images are averaged z-statistic maps overlaid on the averaged high resolution scan transformed into standard (MNI152) space. The left hemisphere of the brain corresponds to the right side of the image.

Figure 6

Averaged maximum z-statistic maps of six subjects using near-infrared spectroscopy (NIRS) low-frequency oscillation (LFO) Δ[HbO] regressors show activations in regions contained in familiar resting state patterns in (A, B). Images are averaged z-statistic maps overlaid on the averaged high-resolution anatomic scan transformed into standard (MNI152) space. The left hemisphere of the brain corresponds to the right side of the image.

Discussion

Benefits of Using Near-Infrared Spectroscopy

In this study, the signal (Δ[HbO]) detected by NIRS was used as a probe to correlate with the BOLD fMRI data. Advantages of this method are twofold. First, both fMRI and NIRS are sensitive to the same blood-related changes, which offers a justification to compare these two signals, and even use them interchangeably. This is not true for other modalities used concurrently with fMRI studies, such as electroencephalogram (Laufs et al, 2003). Attempts have previously been made to use the signals obtained from a respiratory chest belt and end-tidal CO₂ (Birn et al, 2008; Chang and Glover, 2009) as confound regressors to remove the respiration noise from the BOLD. However, the changes in these signals do not directly reflect the corresponding blood signal changes in BOLD. To make them work, some crucial factors, such as the shape and the temporal delay of the hemodynamic function, have to be modeled. An appropriately filtered and systematically delayed NIRS-measured signal is a more direct measure of the hemodynamic changes to which BOLD is sensitive. Second, NIRS has much higher temporal resolution than fMRI (12.5 versus 0.67 Hz in this study). It can sample physiological respiration and heartbeat signals without aliasing. Thus, it offers a way to spectrally separate these signals and study their effects individually. Moreover, the high temporal resolution of NIRS also allows us to generate temporally shifted NIRS regressors (with steps as small as 0.08 seconds) to accurately match the time latencies of these physiological signals at different voxels.

Low-Frequency Oscillations

In this study, the dynamic spatial pattern of the LFO was reproduced (Figure 3A) (Tong and Frederick, 2010) by using only the filtered low-frequency component of the NIRS signal as the source of time-delayed regressors; the direct influences of other physiological noise bands (respiratory and cardiac) have been completely removed from the data. We were, therefore, able to consider the LFO signal in isolation, and some of its proposed origins.

Although it has been suggested that LFO is instrumental noise (Smith et al, 1999; Zarahn et al, 1997), NIRS is a completely independent method for estimating hemodynamic signals, which does not share any instrumental noise mechanisms with fMRI. The correlated LFO detected by both instruments, therefore, cannot be the result of instrumental noise. Additionally, although it has also been suggested that LFO is the aliased signal of respiratory or cardiac signals (Bhattacharyya and Lowe, 2004; Birn et al, 2006, 2008; Chang and Glover, 2009; Shmueli et al, 2007; van Buuren et al, 2009), in this study, the components of respiration and heartbeat were spectrally filtered out from the critically sampled NIRS signal to obtain the NIRS LFO regressors (no aliased respiratory or cardiac signals were presented in the LFO). Thus, the findings from the LFO regressor analyses cannot be explained by aliasing, which is consistent with the work of Kiviniemi et al (2005).

As to whether LFO in BOLD arises from hemodynamic fluctuation related to spontaneous neuronal activities, which is a basic assumption of resting state connectivity analyses, the results were more equivocal. On one hand, the areas of high correlation between the NIRS and BOLD signals in the LFO band overlap regions prominent in a number of resting state networks described by Damoiseaux et al (2006), as shown in Figure 6. On the other hand, the dynamic evolution of these patterns (as shown in Figure 3A) indicates that at least some part of the LFO detected by both modalities is closely correlated with CBF, and not with the regional neuronal activities. A possible explanation for this apparent inconsistency is that both physiological signal (mainly from the skin and skull) and neuronal signal (mainly from the cortex) were contained within the NIRS data, thus enabling the detection of both resting state networks and blood circulation. The image at the right of Figure 6A shows that the brain region under the NIRS probe was highly correlated with NIRS signal, indicating that blood oscillation in this region (which could be neuronal or circulatory) was detected by NIRS. The LFO signal in any brain regions likely contains contributions from both sources. However, our work suggests that if NIRS probes can be designed and placed carefully, detecting only physiological and not neurogenic signals, then the portion of the signal arising from global blood parameters fluctuations can be studied in isolation.

In conclusion, our results suggest that BOLD signal contains a LFO component that has the following features: (1) it is not directly related to respiration and heartbeat by aliasing; (2) it is widely distributed in the brain, especially in the gray matter as well as areas rich in veins; and (3) it is closely related to cerebral blood circulation and moves through the brain with the blood, thus it is not necessarily, and certainly not entirely, related to the neuronal activation. This finding is consistent with the studies performed by Katura et al (2006), who reported that the contribution of heart rate and arterial blood pressure to the variations of LFO is <50%. Yan et al (2009) postulated that brain physiology, instead of scanner instability or cardiac/respiratory pulsations, might be the main source of the LFO.

Heartbeat

In this study, the portion of the NIRS signal in the cardiac frequency range was isolated using the highpass filter. Then, it was resampled at every 1.5 seconds to form the regressor, which together with its time-shifted versions, was used to locate those brain regions in which BOLD signal was influenced by the heartbeat. From Figure 2E, it can be seen that the resampling process produced substantial aliasing of the heartbeat signal. However, by using this aliased ‘heartbeat’ signal and its time-shifted versions as regressors, the dynamic spatial pattern of correlated brain signal (Figure 3B) that oscillated around 1 Hz was detected. Interestingly, this demonstrated that the highly aliased heartbeat signal existed in the BOLD signal at voxels located mainly in and around the arteries and big veins. More importantly, the temporal correlations among the BOLD signals of these voxels faithfully reflected that of the real underlying cause (matching the heartbeat at around 1 Hz), despite a resampling rate of 0.67 Hz (TR) and a time shift increment of 0.16 seconds. In this study, NIRS has been shown to be an effective tool to unveil not only voxels that had good correlations with heartbeat, but also their underlying temporal relationships connected by the progression of the heartbeat pressure wave.

The dynamic spatial pattern correlated with the time-delayed resampled heartbeat waveforms (Figures 3B and 4D) did not resemble the orderly flow from arteries, through capillaries, to veins that was observed in the LFO data shown in Figure 3A. This might be due to the fact that the primary effect of cardiac pulsations on the BOLD signal is due to the pressure wave traveling along the blood vessels, rather than from the bulk motion of the blood. Thus, it propagates through the circulatory system much faster than the blood does. The activation patterns nevertheless appeared at locations of main blood vessels, including arteries and veins. This is further demonstrated in Figures 4D and 5D, which clearly show the cerebral vasculature (compared with Figures 4A and 5A), especially in the areas where the dense arteries and big veins are located, such as the circle of Willis and superior sagittal sinus. This result is not surprising since the heartbeat signal is known to be the strongest in the arteries (veins are detectable primarily in proportion to their sizes). The brain regions with BOLD signal influenced by heartbeat were also found to be located in the areas adjacent to cerebrospinal fluid (particularly in brainstem), which are susceptible to the effects of pulsatile motion. Compared with the LFO maps shown in Figures 4B and 5B, the areas that were affected by heartbeat are different. Aliased cardiac signals are mainly apparent in voxels close to the blood vessels (arteries and big veins), while LFO-correlated BOLD signals are widely apparent in gray matter, veins, the bilateral frontal cortex, and midline of the brain, which could potentially be the result of circulatory and neuronal contributions to the NIRS-measured LFO signal as discussed previously.

Respiration

In Figure 4C, the size of the statistically significant area and the averaged z value were smaller compared with the other two maps (Figures 4B and 4D). As we concluded from the single subject data, this may be partially due to the fact that the respiration signal obtained by NIRS had a much lower amplitude than signals in other bands resulting in comparatively low signal-to-noise ratio, reducing the detection power in this spectral band. In fact, only a subset (3/6) of the participants had respiratory signal detectable by NIRS under these experimental conditions. More studies are needed to draw any conclusions. However, the respiratory signal analysis offered a few clues worth noting: (1) brain areas containing large vessels or groups of vessels did not show high correlations with respiration signals and (2) the voxels in which BOLD signal is correlated by NIRS signal in the respiratory frequency band are located mostly at the surface of the brain in Figures 4C and 5C. Further study is required to better characterize the response in this band.

Very-Low Frequency Oscillations

The z-statistic maps produced by using VLF components of the NIRS signal as regressors were not considered meaningful in this study. This is because that the VLF signal obtained by NIRS, though mainly a physiological signal, is likely to be strongly contaminated by slow instrumental drift (from thermal effects on the lasers and detectors). In BOLD fMRI, Yan et al (2009) found that the spatial distribution of VLF (0.00 to 0.01 Hz) in human brain followed a tissue-specific pattern, namely greater drift in the gray matter than in the white matter. This indicated the brain physiology was the main source of the VLF. More studies with carefully measured instrumental noise controls are needed to find the origin of the VLF.

Like instrumental noise in NIRS, the scanner noise was also present in the BOLD data (Zarahn et al, 1997). However, these two instrumental noise signals arise in totally different instruments, and thus have no correlations with each other. To remove the effects of VLF, we used the high pass filter offered in FSL to remove the any signal below 0.01 Hz from the time-shifted NIRS regressors as well as the BOLD signals.

Repetition Time

In this study, the TR was 1.5 seconds, which was shorter than the typical TR used for fMRI (2 to 3 seconds). However, we believe that the results we found in this study, namely different spatial and temporal patterns of voxels affected by corresponding physiological signals, could be generalized to other studies with more typical TR values. Since it has been shown that the temporal correlation between the NIRS and BOLD signals was detectable, and varied at the appropriate physiological frequency (i.e., heartbeat at 1 Hz), even when the underlying physiological signal was strongly aliased in the BOLD data, we would, therefore, not expect a longer TR to affect the results greatly. Under conditions where the TR is longer and the resampling rate is lowered, the resulting z-statistic map would not be as detailed as one produced with a shorter TR and higher resampling rate, but temporal and spatial patterns from both maps would nevertheless be similar.

In conclusion, the primary goal of this work was to understand both the temporal and spatial relationships between BOLD fMRI signal and various physiological processes, including LFO, respiration, and cardiac pulsation. We demonstrated that NIRS is a useful tool for detecting these physiological processes with much higher temporal frequency compared with BOLD fMRI. Filtered and time-delayed regressors obtained from the NIRS data were successfully applied in analyses of the BOLD fMRI data and corresponding temporal and spatial patterns for the different processes were identified, despite the fact that the BOLD sample rate caused aliasing of these physiological effects. Our results suggest that (1) there is a strong LFO signal in BOLD fMRI, that is neither the aliased signal of respiration and cardiac pulsation, nor does it have neuronal origin. This LFO was closely related to global CBF and may partially obscure resting state networks; (2) the BOLD signals affected by heartbeat are detectable are mainly at the locations with high concentrations of large vessels; and (3) the influence of respiration on BOLD signal remains unclear due to the low signal-to-noise ratio in NIRS data.

Footnotes

The authors declare no conflict of interest.

References

Balduzzi

Riedner

Tononi

(2008) A BOLD window into brain waves. Proc Natl Acad Sci USA 105:15641–2

Bandettini

Wong

Hinks

Tikofsky

Hyde

(1992) Time course EPI of human brain function during task activation. Magn Reson Med 25:390–7

Bhattacharyya

Lowe

(2004) Cardiac-induced physiologic noise in tissue is a direct observation of cardiac-induced fluctuations. Magn Reson Imaging 22:9–13

Birn

Diamond

Smith

Bandettini

(2006) Separating respiratory-variation-related fluctuations from neuronal-activity-related fluctuations in fMRI. Neuroimage 31:1536–48

Birn

Smith

Jones

Bandettini

(2008) The respiration response function: the temporal dynamics of fMRI signal fluctuations related to changes in respiration. Neuroimage 40:644–54

Buxton

Wong

Frank

(1998) Dynamics of blood flow and oxygenation changes during brain activation: the balloon model. Magn Reson Med 39:855–64

Chang

Cunningham

Glover

(2009) Influence of heart rate on the BOLD signal: the cardiac response function. Neuroimage 44:857–69

Chang

Glover

(2009) Relationship between respiration, end-tidal CO2, and BOLD signals in resting-state fMRI. Neuroimage 47:1381–93

Cole

Smith

Beckmann

(2010) Advances and pitfalls in the analysis and interpretation of resting-state FMRI data. Front Syst Neurosci 4:8

10.

Crandell

Moinuddin

Friedman

Robertson

(1973) Cerebral transit time of 99 m technetium sodium pertechnetate before and after cerebral arteriography. J Neurosurg 38:545–7

11.

Damoiseaux

Rombouts

Barkhof

Scheltens

Stam

Smith

Beckmann

(2006) Consistent resting-state networks across healthy subjects. Proc Natl Acad Sci USA 103:13848–53

12.

Evans

Collins

Mills

Brown

Kelly

Peters

(1993) 3D Statistical neuroanatomical models from 305 MRI volumes. IEEE-Nuclear Science Symposium and Medical Imaging Conference. pp, 1813–7

13.

Fox

Raichle

(2007) Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging. Nat Rev Neurosci 8:700–11

14.

Frank

Buxton

Wong

(2001) Estimation of respiration-induced noise fluctuations from undersampled multislice fMRI data. Magn Reson Med 45: 635–44

15.

Fransson

(2005) Spontaneous low-frequency BOLD signal fluctuations: an fMRI investigation of the resting-state default mode of brain function hypothesis. Hum Brain Mapp 26:15–29

16.

Fukunaga

Horovitz

de Zwart

van Gelderen

Balkin

Braun

Duyn

(2008) Metabolic origin of BOLD signal fluctuations in the absence of stimuli. J Cereb Blood Flow Metab 28:1377–87

17.

Glover

Ress

(2000) Image-based method for retrospective correction of physiological motion effects in fMRI: RETROICOR. Magn Reson Med 44:162–7

18.

Greve

Goldenholz

Kaskhedikar

Polimeni

Moran

Schwartz

Fischl

Wald

Rosen

Triantafyllou

Boas

(2009) BOLD physiological noise reduction using spatio-spectral-temporal correlations with NIRS. ISMRM, Hawaii

19.

Katura

Tanaka

Obata

Sato

Maki

(2006) Quantitative evaluation of interrelations between spontaneous low-frequency oscillations in cerebral hemodynamics and systemic cardiovascular dynamics. Neuroimage 31:1592–600

20.

Kiviniemi

Ruohonen

Tervonen

(2005) Separation of physiological very low frequency fluctuation from aliasing by switched sampling interval fMRI scans. Magn Reson Imaging 23:41–6

21.

Laufs

Kleinschmidt

Beyerle

Eger

Salek-Haddadi

Preibisch

Krakow

(2003) EEG-correlated fMRI of human alpha activity. Neuroimage 19:1463–76

22.

Matcher

(1994) Use of the water absorption spectrum to quantify tissue chromophore concentration changes in near-infrared spectroscopy. Phys Med Biol 39:177–96

23.

Obrig

Neufang

Wenzel

Kohl

Steinbrink

Einhaupl

Villringer

(2000) Spontaneous low frequency oscillations of cerebral hemodynamics and metabolism in human adults. Neuroimage 12:623–39

24.

Ogawa

Tank

Menon

Ellermann

Kim

Merkle

Ugurbil

(1992) Intrinsic signal changes accompanying sensory stimulation: functional brain mapping with magnetic resonance imaging. Proc Natl Acad Sci USA 89:5951–5

25.

Okamoto

Dan

Sakamoto

Takeo

Shimizu

Kohno

Oda

Isobe

Suzuki

Kohyama

Dan

(2004) Three-dimensional probabilistic anatomical cranio-cerebral correlation via the international 10-20 system oriented for transcranial functional brain mapping. Neuroimage 21:99–111

26.

Sassaroli

BFB

Tong

Renshaw

Fantini

(2006) Spatially weighted BOLD signal for comparison of functional magnetic resonance imaging and near-infrared imaging of the brain. Neuroimage 33:505–14

27.

Shmueli

van Gelderen

de Zwart

Horovitz

Fukunaga

Jansma

Duyn

(2007) Low-frequency fluctuations in the cardiac rate as a source of variance in the resting-state fMRI BOLD signal. Neuroimage 38: 306–20

28.

Smith

Lewis

Ruttimann

Sinnwell

Yang

Duyn

Frank

(1999) Investigation of low frequency drift in fMRI signal. Neuroimage 9:526–33

29.

Smith

Jenkinson

Woolrich

Beckmann

Behrens

Johansen-Berg

Bannister

De Luca

Drobnjak

Flitney

Niazy

Saunders

Vickers

Zhang

De Stefano

Brady

Matthews

(2004) Advances in functional and structural MR image analysis and implementation as FSL. Neuroimage 23(Suppl 1):S208–19

30.

Strangman

Culver

Thompson

Boas

(2002) A quantitative comparison of simultaneous BOLD fMRI and NIRS recordings during functional brain activation. Neuroimage 17:719–31

31.

Suckling

Wink

Bernard

Barnes

Bullmore

(2008) Endogenous multifractal brain dynamics are modulated by age, cholinergic blockade and cognitive performance. J Neurosci Methods 174:292–300

32.

Tong

Bergethon

Frederick

(2011) An improved method for mapping cerebrovascular reserve using concurrent fMRI and near-infrared spectroscopy with regressor interpolation at progressive time delays (RIPTiDe). Neuroimage 56:2047–57

33.

Tong

Frederick

(2010) Time lag dependent multimodal processing of concurrent fMRI and near-infrared spectroscopy (NIRS) data suggests a global circulatory origin for low-frequency oscillation signals in human brain. Neuroimage 53:553–64

34.

Toronov

Walker

Gupta

Choi

Gratton

Hueber

Webb

(2003) The roles of changes in deoxyhemoglobin concentration and regional cerebral blood volume in the fMRI BOLD signal. Neuroimage 19:1521–31

35.

van Buuren

Gladwin

Zandbelt

van den Heuvel

Ramsey

Kahn

Vink

(2009) Cardiorespiratory effects on default-mode network activity as measured with fMRI. Hum Brain Mapp 30:3031–42

36.

Vernieri

Tibuzzi

Pasqualetti

Rosato

Passarelli

Rossini

Silvestrini

(2004) Transcranial Doppler and near-infrared spectroscopy can evaluate the hemodynamic effect of carotid artery occlusion. Stroke 35: 64–70

37.

Wink

Bullmore

Barnes

Bernard

Suckling

(2008) Monofractal and multifractal dynamics of low frequency endogenous brain oscillations in functional MRI. Hum Brain Mapp 29:791–801

38.

Yan

Zhuo

Xie

Aguirre

Wang

(2009) Physiological origin of low-frequency drift in blood oxygen level dependent (BOLD) functional magnetic resonance imaging (fMRI). Magn Reson Med 61:819–27

39.

Zarahn

Aguirre

D'Esposito

(1997) Empirical analyses of BOLD fMRI statistics. I. Spatially unsmoothed data collected under null-hypothesis conditions. Neuroimage 5:179–97

40.

Zuo

Di Martino

Kelly

Shehzad

Gee

Klein

Castellanos

Biswal

Milham

(2010) The oscillating brain: complex and reliable. Neuroimage 49:1432–45