Perfusion information extracted from resting state functional magnetic resonance imaging

Abstract

It is widely known that blood oxygenation level dependent (BOLD) contrast in functional magnetic resonance imaging (fMRI) is an indirect measure for neuronal activations through neurovascular coupling. The BOLD signal is also influenced by many non-neuronal physiological fluctuations. In previous resting state (RS) fMRI studies, we have identified a moving systemic low frequency oscillation (sLFO) in BOLD signal and were able to track its passage through the brain. We hypothesized that this seemingly intrinsic signal moves with the blood, and therefore, its dynamic patterns represent cerebral blood flow. In this study, we tested this hypothesis by performing Dynamic Susceptibility Contrast (DSC) MRI scans (i.e. bolus tracking) following the RS scans on eight healthy subjects. The dynamic patterns of sLFO derived from RS data were compared with the bolus flow visually and quantitatively. We found that the flow of sLFO derived from RS fMRI does to a large extent represent the blood flow measured with DSC. The small differences, we hypothesize, are largely due to the difference between the methods in their sensitivity to different vessel types. We conclude that the flow of sLFO in RS visualized by our time delay method represents the blood flow in the capillaries and veins in the brain.

Introduction

Functional magnetic resonance imaging (fMRI) has been widely utilized to study brain function. However, it does not image neuronal activations directly. It reflects the neurovascular changes correlated with neuronal activations. In short, local neuronal activations increase metabolic rate, leading to increased blood flow and volume at the site. These blood-related fluctuations change the local magnetic environment via changes in paramagnetic deoxy-hemoglobin concentration ([Hb]), resulting in alterations of the fMRI images. Therefore, the changes in [Hb] provide the contrast, called blood oxygenation level dependent (BOLD), in fMRI.^1,2 While neuronal activation occurs very quickly (with timescales in milliseconds), it produces a much slower blood reaction (2–6 s), the hemodynamic response, which is detectable by fMRI. Mathematically, in BOLD fMRI time courses, the hemodynamic response function serves as a low-pass filter for neuronal signals,³ and because of that the neural signals that are detected by BOLD fMRI are seen as low frequency oscillations (LFOs) (<0.15 Hz).

In addition to neuronal activations, many other non-neuronal physiological processes, such as respiration and cardiac pulsation, can also change the voxel [Hb] through changes in blood flow, volume, and oxygenation.¹ Thus, they affect the BOLD fMRI signal. Understanding these non-neuronal BOLD signal components is critical for two reasons. First, characterizing the non-neuronal components of BOLD can help to uncover the real neuronal signals by allowing for more effective removal of this unwanted variance. This is especially critical where the neuronal effect size is small and with unknown temporal characteristics, for example, resting state (RS) studies.⁴ In these studies, no external stimulation is used, leading to spontaneous (unpredictable) and relatively small neuronal activations. Secondly, most of these non-neuronal oscillations are not random noise. They reflect ongoing physiological processes, and carry their own information about the subject, which also might be interesting topics of study, yielding different insights.

In this study, we will focus on one of these non-neuronal signals in BOLD. In our recent studies, we have identified a global LFO in BOLD data, which oscillates in the range of 0.01–0.15 Hz. It seems to arrive at different voxels at different times, judging by the appearance of its temporally shifted versions of the same signal in different brain voxels. The dynamic pattern of this LFO arrival through the brain closely resembles the cerebral blood flow pattern with a period of 6–9 s, arriving latest in the large draining veins at the back of the brain.^5,6 This LFO observed in the brain can also be found at peripheral sites (e.g. in the fingertip by near infrared spectroscopy (NIRS)⁷) with other time delays, as changes in the oxy and deoxy-hemoglobin concentration. It even becomes one of the major contributing signals in RS data due to the absence of external stimulation.^8–10 These findings strongly imply that this LFO is a non-neuronal signal. Therefore, we refer to it as a systemic LFO (sLFO), to distinguish it from LFOs of neuronal origin. The origin and function of these non-neuronal sLFOs are not clear.¹¹ They are suggested to be associated with Mayer waves,¹² oscillations in CO₂ level,^13,14 oscillations in vasomotion,¹⁵ etc. Regardless of its source and function, based on its features, we have hypothesized that the sLFO is a signal intrinsic to blood, which travels with the blood to different parts of the brain with different time delays. Its dynamic pattern through the brain reflects the passage of cerebral blood (i.e. blood flow). Although, in previous studies we did not test this hypothesis directly. Here we provide a direct comparison of LFO flow patterns with those measured using dynamic susceptibility contrast (DSC).

DSC, also called bolus tracking, is a commonly used perfusion method^16,17 which captures the first passage of an intravenously injected paramagnetic contrast agent (gadolinium (Gd)-based chelate) through the brain with an echoplanar image acquisition. By kinetic analysis of these data, several important hemodynamic parameters, such as blood flow, volume, and mean transit time can be derived. Most importantly, this method allows the quantitation of the bolus as it flows throughout the brain (i.e. flow of the blood).

To test our hypothesis that the dynamic patterns of sLFO (derived from RS) moving through the brain reflect the cerebral blood flow, we compared the dynamic flow of the bolus derived by the DSC method with that of the sLFO derived from RS data of the same subject. In addition, we applied a novel analytical method to compare these two independent measurements quantitatively to understand the similarities and differences in these two measurements.

Materials and methods

Protocol

Eight healthy control subjects (7M, 1F, average age ± SD, 33 ± 12 years) participated in the study. The RS scan was conducted followed by DSC MRI. These data were collected as part of a longer protocol, the data from the other modalities will be reported elsewhere. In the RS scan, participants were asked to lie quietly in the scanner with their eyes open and view a gray screen with a fixation point in the center. The RS scan lasted 360 s. DSC lasted 180 s. All subjects provided informed, written consent. McLean Hospital (Partners) Institutional Review Board approved the research protocol, which was conducted in accordance with the ethical principles of the Belmont Report.

All MR data were acquired on a Siemens TIM Trio 3T scanner (Siemens Medical Systems, Malvern, PA) using a 32-channel phased array head matrix coil. After scans for localization and automated alignment, multiecho multiplanar rapidly acquired gradient-echo (ME-MPRAGE) structural images were acquired with the following parameters (Repetition Time: TR = 2530 ms, Echo Time: TE = 3.31,6.99,8.85,10.71 ms, Inversion Time: TI = 1100 ms, slices =128, matrix = 256 × 256, flip angle = 7°, resolution =1.0 mm × 1.0 mm × 1.33 mm, 2 × GRAPPA, total acquisition time 4:32). For RS, multiband EPI data were obtained (University of Minnesota sequence cmrr_mbep2d_bold R010^18,19) with the following parameters: TR/TE = 720/32 ms, matrix 86 × 86 on a 212 × 212 mm Field of View, multiband factor = 8, 64 2.5-mm slices with no gap parallel to the AC–PC line extending down from the top of the brain. This was followed by a Gd-DSC scan (TR/TE = 1510/21 ms, 120 time points, flip angle = 60°, 1.8 × 1.8 mm in plane resolution, 29 3.5 mm slices with a 0.5 gap, 3:09). Twenty seconds after the beginning of this scan, 20 ml of Prohance was injected into the antecubital vein at 4 ml/s, followed by a flush of 20 ml of saline using an Medrad Spectris Solaris power injector.

Preprocessing

For each participant, the standard fMRI preprocessing steps, including brain extraction, motion correction, slice-time correction, and smoothing (3 mm) were applied to the original BOLD signals (using FEAT v6.00 of FSL 5.0)²⁰ in both RS and DSC data. The acquisition time of the individual slices during each TR was read from the image DICOM headers to generate a “slice time” file for use in FSL. This allows correction even when slices are acquired simultaneously in a multiband sequence. The RS data (TR = 0.72 s) were filtered with a zero-phase digital filter function filtfilt in MATLAB (The Mathworks, Natick, MA), which uses a third order Butterworth filter applied in the forward and reverse direction to cancel phase delays. The lower and upper cutoff frequencies of the bandpass filter were set to be 0.01 and 0.15 Hz. The reason for this range is two fold: 1) based on previous research,^21–23 the spectral range of LFO in BOLD extends beyond 0.1 Hz and can be as high as 0.2 Hz; 2) the respiration in RS is normally about 0.2–0.3 Hz. To maximize the low frequency signal and not include the respiration signal, we chose a range of 0.01–0.15 Hz. The final step was to register these data (DSC and filtered RS) to the subject’s own anatomical scan and then to the standard brain. The resulting 4D data of both RS and DSC formed the basis of the calculation described below.

Time lag maps of RS and DSC

We have demonstrated how to calculate time lag maps from RS data using the seed voxels from peripheral NIRS recordings or from the BOLD signal itself.²⁴ In order for readers to replicate our finding, we further simplified the procedure (as shown in Figure 1). In short, we first generated a superior sagittal sinus (SSS) mask based on the standard brain. This mask was used on each subject’s filtered RS data to extract the time course from SSS (used it as a seed regressor), which represents a non-neuronal fluctuations in the BOLD signal. Then, we calculated the voxel-wise cross correlations with this seed regressor, resulting in the maximum cross-correlation coefficient and corresponding time lag for each voxel (using in house MATLAB program). We set the range of the cross-correlation to be ±6 s based on previous studies⁵ and literature.²⁵ To prevent spurious correlation, we set a strict lower threshold for the maximum correlation coefficient at 0.3 based on our previous research.²⁶ In other words, the voxel will be considered “valid” only if the maximum cross-correlation coefficient is greater than 0.3. The final time lag map is a 3D image showing the corresponding time lag of each valid voxel (as seen in Figure 1). The histogram of these valid voxels according to their time lag is calculated for each subject, as shown in Figure 1. We calculated the group histogram and group time lag map for all eight subjects.

Figure 1.

Flowchart of the method used to calculate the time lag map for each subject from the filtered RS data (0.01–0.15 Hz). A filtered seed regressor was selected from the superior sagittal sinus (SSS) (mask). This regressor was then cross-correlated with all other BOLD signals to select the voxels that have significant maximum correlation coefficients (Maxcc > 0.3). The corresponding time lags of these voxels form the time lag map. The histogram of these time lags is also shown. Lastly, a dynamic sLFO movie was created from the RS time lag map. Each frame of the movie shows the RS time lag map consisting of the voxels whose time lags are within a certain time range, namely the time window shown as blue bar in the histogram. As the window moves from −6 to 6 s, a series of non-overlapping frames are created. TS: time series.

We also calculated the time lag map and corresponding histogram from the DSC data for each subject. The time lag of each voxel is supposed to represent the relative arrival time of the bolus to that voxel. Since passage of a Gd bolus through the brain causes a dramatic transient decrease in MR signal intensity, the arrival time is properly defined as the time point where the DSC signal starts to decrease. However, it is hard to locate the arrival time accurately due to fluctuations in the baseline of the DSC signal. Therefore, time to peak (TTP) value, which is a widely used parameter in DSC, was used as a proxy to represent the arrival time. TTP is typically defined as the time from the start of the scan to peak DSC signal loss, which is easily identified, thus avoiding the ambiguities of deciding the first arrival.²⁷ To accurately calculate the TTP values, we used the program Perfx developed by Chris Rorden (http://www.mccauslandcenter.sc.edu/CRNL/tools/pwi). The program applies a Gamma variate fit to the DSC data, producing accurate TTP maps for each subject. The group time lag map as well as the corresponding group histogram was calculated as done in the RS data.

A movie reflecting the dynamic flow of the sLFO and the Gd bolus can be generated from the two time lag maps, respectively. An example (i.e. movie from RS) is shown at the bottom of Figure 1. In short, each frame of the movie consists of voxels whose time lags are within a certain time range (e.g. 0.18 s time window). Sliding the time window stepwise (by the same time range: 0.18 s) from −6 to 6 s generated consecutive non-overlapping frames. A movie was made by concatenating all these frames in sequence.

Comparison between DSC and RS

In order to validate the flow of sLFO (RS) using that of the bolus (DSC), we applied a novel analytical method, in which the frames from the RS movie were used as masks to extract a series of corresponding DSC time courses. The time lags between these DSC time courses were evaluated based on the time difference between the frames. This method allows us to compare two flow metrics quantitatively. The detailed steps for each subject are as follows (as shown in Figure 2).

1. A set of 3D masks was generated from each frame of the RS movie (as shown in Figure 1). For example (Figure 2), mask N is from the Nth frame of the RS movie and consists of all the voxels which time lags are between the time t_N and t_N +0.18 s (Δt = 0.18 s).

2. Each RS-based mask was applied onto the same subject’s DSC data (4D) to extract corresponding DSC time courses. We then calculated TTP from the averaged DSC time course (as shown in Figure 2). For example, mask N was applied to the DSC data to extract the averaged time course. The corresponding TTP of the time course was then obtained, as TTP (N).

3. Finally, to compare the arrival time of the bolus with that of LFOs, we plotted values of TTP derived from the average DSC time course from each mask versus the average RS time lag of each mask. For example (as shown in Figure 2), the RS time lag of mask N is t_N, the corresponding TTP is TTP(N). The time lag of mask N + 1 is t_N + 0.18 s, with a corresponding TTP of TTP (N + 1). Each mask was represented by a dot in the TTP/time lag graph, and a total of 67 dots are shown on the TTP/time lag graph with the x-axis reflecting the propagation of RS LFOs (in sec), and the y-axis reflecting corresponding propagations of DSC bolus (in sec). If the propagation of RS LFOs matches that of the DSC bolus perfectly, a line of unit slope (Δy = Δx) is expected. We calculated the individual TTP/time lag graph as well as the group result (N = 8).

Figure 2.

A schematic diagram showing the way to calculate the TTP/time lag graph. In detail, each frame of the sLFO movie (from Figure 1) was used as a mask to extract the corresponding DSC time series (ts) from the same subject. The TTP value was obtained from these averaged DSC time series. In the TTP/time lag graph, we plotted values of TTP vs. time lag for each frame.

Results

Figure 3(a) shows the histogram of RS time lags for each subject (color lines). The curves are consistent among the eight subjects, which is a slightly left modal distribution covering roughly −6 to 4 s. Please note here, the peak position of each histogram is not the same, because the relative time delay of the seed voxels (from SSS) to the rest of the voxels is subject-specific. For the purpose of comparison, we therefore shifted the curves to align the peaks horizontally (to that of subject 1). The spikes at both ends of the curves represent the number of voxels that have extremely long lags (<<−6 or >>6 s) with respect to the SSS seed. Mostly, they are from extracerebral regions and near ventricles.

Figure 3.

Histograms of the RS time lag maps (in (a)) and corresponding TTP/time lag graphs (in (b)) from 8 subjects. For the purpose of comparison, the histograms were shifted on the x-axis to align the peaks with that of subject 1. The curves in (b) were also shifted accordingly. The blue shaded area in (b) shows the region with the robust results obtained from large number of voxels.

The corresponding TTP/time lag graphs of the eight subjects are shown in Figure 3(b). We found positive correlations between sLFO time lag of RS and TTP of DSC in all the subjects. Moreover, the subjects’ data and correlations are highly consistent with each other, indicating the robustness of the phenomenon. For consistency with the histograms, we applied the same horizontal shifts (i.e. x-axis) as in the TTP/time lag graph. Moreover, since the relative (not the absolute) TTP values of DSC are important (i.e. y-axis), all the initial values in TTP/time lag graph are shifted vertically to 0 for the purpose of comparison (note that: after subjects’ TTP/time lag graphs are shifted horizontally to be aligned, the common RS time lag of TTP initial value is at −4.74 s instead of −6 s).

In order to develop a clear understanding of the trends observed in Figure 3(b), we plotted the group TTP/time lag graph (in black) with the standard deviation in Figure 4(a). We excluded the data outside before −4.7 and after 2 s (shown in Figure 3(b)) because these TTP values are obtained from a significantly smaller number of voxels (as shown in Figure 3(a), the number of voxels per histogram bin <15% of the peak), making them noisy and unreliable. From Figure 4(a), the scatterplot describing the relationship between the two measures can be separated into two ranges based on its local slope. First, from roughly −4.7 to −1.7 s (blue range in Figure 4(a)), the relationship is almost flat (indicated by the blue line). Second, from −1.7 to 2 s (red range in Figure 4(a)), the slope of the line is close to 1 (indicated by the red line). The spatial distributions of the voxels in these two time-lag ranges are shown in Figure 4(b) and (c), respectively. From these two figures, we can see that the spatial distributions corresponding to the different slopes are distinct (and non-overlapping). Those voxels associated with the “flat” slope (slope ≈ 0) are largely found around the central gyrus and gray matter as shown in Figure 4(b), while those voxels associated with slope ≈ 1 largely consist of some gray matter, white matter and large draining veins as shown in Figure 4(c). Lastly, the average standard deviations of the TTP values extracted by each RS mask were shown as shaded area in Figure 4(a). The range is relatively large due to the number of voxels (i.e. TTP values) involved from each mask (2000–10,000). However, the trend of the curves stays the same.

Figure 4.

(a) Group DSC TTP/RS time lag graph (black line) from the center region of Figure 3(b). The error bars are the standard deviations calculated from 8 subjects’ curves, as shown in Figure 3(b). The standard deviation of the TTP measurements for each RS mask was calculated for all the subjects and the averaged result was shown as shaded area in (a). Two time lag ranges are differentiated by the slope of the TTP/time lag curve and were colored according to the separation with blue and red. In the blue region, the slope is almost flat (blue line), while in the red region, the regression line is close to 45° degrees (slope = 1). In (b) and (c) the corresponding spatial distribution of the voxels are depicted for the blue (voxel type 1) and red (voxel type 2) time lag ranges, respectively. (NB: blue represents early RS time lags, red represents late time lags.)

The group time delay maps derived from RS and DSC data are shown in Figure 5(a) and (b), respectively. There, in order to maximize the pseudocolor dynamic range in FSLView, the time delays have been re-centered at 0 s. The following features can be observed. First, the RS time lag map is similar to maps from our previous works, in which, early time lags can be found in symmetric central regions, such as the motor cortex. Late time lags are mainly concentrated at white matter and draining veins. Compared to the RS time lag maps, DSC maps are more uniform in most of the gray matter. However, the RS and DSC maps are similar in that the voxels with late time lags are concentrated in the white matter and draining veins. The dynamic movies derived from these two time lag maps are shown in the Supplemental Material with the top one from RS data and bottom one from DSC data. The time range of each frame is 0.72 s, instead of 0.18 s, to control the file size. The durations of both movies are 8.64 s, which is consistent with our previous findings.^5,6

Figure 5.

Group averaged time lag maps calculated from RS data (a) and DSC data (b).

We show the group histogram of the RS and DSC time lag maps in Figure 6. The dark line represents the averaged histogram of DSC time lag maps, while shaded lines represent that of the RS time lag maps (same as that in Figure 3(a)). The peaks of both curves have been aligned. Compared to the curve of RS, the curve of DSC has a more pronounced positive skewness.

Figure 6.

Histograms of the group DSC TTP map (black line) and the group RS time lag map (gray line).

Discussion

This is the first study to explore the relationship between dynamic features of the sLFO in RS BOLD fMRI and cerebral blood flow measured by DSC. We have demonstrated that the dynamic flow of sLFOs in the brain match that of cerebral blood flow measured by DSC to a large extent, defined by the time range of the measurements. This flow matching has been visualized using novel TTP/time lag graphs as a tool, which can be observed in Figures 3(b) and 4(a), where the x-axis represents the propagation of sLFO in the RS data, while the y-axis represents the corresponding bolus propagation. If these two independent measurements (RS vs. DSC) are not related, random patterns (around group average of zero) would be expected in these figures. On the contrary, highly consistent positive correlations are found in all subjects as shown in Figure 3(b), which offers direct evidence that the movement of sLFO in the brain reflects brain blood flow. However, there are also notable discrepancies in these two flow metrics. As mentioned before, if they were exactly the same, a line of equality (i.e. slope = 1) would be expected in Figure 4(a). However, this is true only in the red-shaded range of Figure 4(a), when relative time lags are greater than −2 s. In the blue-shaded range (<−2 s), an almost “flat” line is shown indicating there is no clear dependence between time lags measured by DSC and RS sLFOs. Moreover, the discrepancy can also be seen when comparing the two time lag maps in Figure 5. Instead of being the same, the time lag map of DSC is much more uniform than that of RS, especially in large areas of gray matter. We will discuss the robustness of the RS time lag map in “Robustness of the time lag maps” section and the discrepancies between these two flow metrics in “Consistency and difference between DSC and RS” section.

Robustness of the time lag maps

The RS time lag maps are highly consistent among the eight subjects, and this inter-subject consistency is also reflected in the similarity of the individual subject histograms (Figure 3(a)). The consistency demonstrates the robustness of the procedure. More importantly, it indicates the wide presence of sLFO signals in RS data and highly repeatable dynamic patterns among subjects throughout the brain. These patterns are clearly visible in the movie we generated from the group averaged RS time lag map (see Supplemental Material).

As stated in “Time lag maps of RS and DSC” section, spurious correlation must be considered in the calculation. Two steps in the analysis procedure result in systematic inflation of the correlation coefficient; traditional statistical methods for assessing the significance of a correlation are not appropriate in this case, and must be revised.²⁸ The first is that the BOLD signals are filtered into the low frequency range 0.01–0.15 Hz to select only the LFO signals; lowpass filtering data broadens the distribution of chance correlations. The second step is the selection of the maximum correlation value over a range of time lags, which both increases the average chance correlation values and skews their distribution towards positive values. To control for spurious correlations in our study, we implemented a very strict lower threshold of acceptance (=0.3). The new threshold was obtained in our previous study,²⁶ in which we used simulated data to calculate the proper threshold for similar data and procedures. On average 36% of ‘valid’ voxels are in gray matter, 25% in white matter and 22% in CSF. The reason this does not add up to 100%, is largely due to the registration error. Conversely, 73%, 63%, and 69% voxels from gray matter, white matter, and CSF are ‘valid’ voxels.

To further understand the impact of spurious correlation, we performed the same procedure (as shown in Figure 1) to generate random (incoherent), non-subject specific, RS time lag maps by using swapped seed regressors. For example, we used subject 1’s seed regressor (from SSS) to cross correlate with subject 2’s RS data and generated “fake” time lag maps for subject 2. We did this swapped seed regressor procedure for all the subjects and found that: 1) most of the voxels did not pass the threshold (i.e. max cross correlation coefficient <0.3); 2) for the surviving voxels, the averaged group time lag map showed random patterns with very little spatial coherence compared to Figure 5(a). This provides strong evidence that the strict threshold has largely prevented spurious correlation and the RS time lag maps reflect the real meaningful correlation values between sLFOs. We included the “fake” averaged group RS time lag map as Figure 1 in the Supplemental Material.

The high consistency is also observed in the TTP/time lag graph, as shown in Figure 3(b). The positive correlation between the propagations of bolus and sLFO is observed in all the subjects. The consistency strongly indicates the robust relationship between these two flow measures (sLFO and bolus) in all subjects, supporting the contention that the sLFO propagation reflects the movement of blood.

It is critical to understand the impact of the frequency range selection on the result. Moreover, it is of great importance to understand the frequency content of this sLFO. We have calculated four TTP/time lag curves from one subject’s RS data filtered to four different frequency ranges: 1) 0.01–0.15 Hz; 2) 0.01–0.1 Hz; 3) 0.1–0.25 Hz; 4) 0.15–0.25 Hz (Figure 2 in the Supplemental Material). We also calculated averaged TTP/time lag graphs (Figure 3 in the Supplemental Material) for all the subjects using the bandpass filter range 0.01–0.1 Hz (commonly used in RS analysis). From these two figures, we can see the results are very similar only in the low frequency range (<0.15 Hz), with decreased slope from 0.01 to 0.1 Hz data, which implies that there is useful flow information between 0.1 and 0.15 Hz in sLFO and thus confirmed our initial selection (0.01–0.15 Hz).

Consistency and difference between DSC and RS

From Figure 4(a), we have demonstrated that the propagation of sLFO has a close link to the blood flow (i.e. DSC). However, it is not a perfect match; if it were, the measurements from the two methods would fall on the line of equality (i.e. slope = 1). We believe the main reason lies in the complicated interplay between a number of factors: 1) the different effects of the contrast agents in these two imaging technologies (the Gd agent for DSC and [Hb] for RS) on image intensity in different vessel types, namely arteries, veins, and capillaries; 2) the relative proportion of different blood vessel types in each voxel (artery, capillary and vein) and 3) different blood flow velocities in each of these vessel types. In DSC imaging scans, the Gd contrast agent that is injected into the subject is highly paramagnetic, and is visible by the scanner as soon as it enters the brain and in all vessel types. In RS fMRI, the contrast agent is [Hb]; arterial blood oxygenation is close to 100%. As a result [Hb] is negligible in arteries, and therefore, arteries and arterioles are almost invisible to BOLD fMRI. As shown in Figure 7(a), the vascular system within each voxel can be thought to consist of three parts, artery (including arterioles), capillary and vein (including venules). Depending on the location, the BOLD signal from a voxel can be dominated by signal originating from any one (or two) of these three vessel types.

Figure 7.

Schematic diagram of blood vessel components in a typical voxel (a). The voxels of type 1 are dominated by signals from the arteries and the capillary (b) and the voxels of type 2 are dominated by signals from the capillary and veins (c). Upper cases of A, C, and V represent the dominant arterial, capillary, and vein signal component in the voxel, while lower cases represent the same components with minimal influence on the signal. The length of the red arrow represents the blood flow velocity, while the length of the black arrow represents the time lags. The table (d) summarizes the interactions among all these parameters. DSC detects the blood in large arteries, capillary, and small veins in voxels of type 1, where the blood moves very fast (long red arrow), resulting in small or no time lags in DSC signals (very short black arrow). However, RS only detects the blood in the capillary and veins in voxels of type 1, thus the blood flow is much slower (short red arrow), resulting in large time lags in RS signals (large black arrow). The same logic applies to voxels of type 2.

Lastly, the blood flow velocity in artery (∼80 cm/s) is much faster than those in capillary (∼0.1 cm/s) and vein (∼10 cm/s).²⁹ The earliest stage of cerebral blood flow (the blue-shaded time-lag range in Figure 4(a)), involves voxels in which arteries, arterioles, and some capillaries dominate the signal, as depicted in Figure 7(b). fMRI, which is insensitive to deoxy-hemoglobin-lacking arteries measures only the slow blood flow in these small capillaries where deoxy-hemoglobin appears, while DSC can detect the much faster arterial flow from the same voxels with the help of exogenous injected contrast. Since the arterial blood flow is so fast that it can reach these voxels almost instantaneously (especially when TR = 1.5 s in DSC acquisition), there is almost no time lag between these DSC signals, resulting in the flat line of TTP in the blue-shaded time-lag range of Figure 4(a). Figure 4(b) shows the spatial distribution of these voxels in blue (group result), in which we can see that they are concentrated around central gyrus, orbitofrontal cortex, large areas of gray matter, etc. These are the areas that have higher arterial density and are near the beginning of the cerebrovascular flow path.

A subsequent stage of cerebral blood flow involves voxels in which capillaries and veins dominate the signal, as shown in Figure 7(c). In these voxels, the blood flow (in the capillaries and veins) is equally observed by both methods, resulting in a line with slope close to 1 (as depicted in the red-shaded time-lag range shown in Figure 4(a)). The same argument explains why the line is a little “flatter” than 45°; DSC data from these voxels reflect the integrated flow from all the compartments (artery, capillary and vein), leading to faster average flow velocity detection than that of only capillary and veins (seen in fMRI RS signal). Lastly, as depicted in Figure 4(c), the red-colored voxels reside in the areas of high vein density, including white matter and draining veins. Many large surface veins are not fully visible here (including the SSS) because they have been removed by the standard brain mask.

These two scenarios are depicted in Figure 7(d), in which red arrows indicate the blood flow velocity seen by each method, while black arrows represent the corresponding time lags calculated from each method. For each method, the faster the blood flow (longer red arrow), the smaller time lags were found (shorter black arrow). For each voxel type, the ratio of the time lags (i.e. ratio of the two black arrows) from two methods determines the slope of the line in Figure 4(a).

This explains both the similarities and differences in the two time lag maps in Figure 5. We can see similar delay patterns (yellow) in the white matter and the veins of both maps, indicating both methods measure blood flow similarly in these areas. However, in the gray matter, the DSC method produces more uniform TTP values (blue) in comparison to the RS time lag method. As we argued, this indicates that much faster blood flow (from arteries and arterioles) is observed by the DSC method, resulting in undifferentiated time lags compared to that of the RS method, which is sensitive to signal from capillary and veins from the same voxels. We do not fully understand why the early detectable sLFO appear in the central motor area (as light blue in Figure 5(a)), which is clearly visible in the first few frames of the movie (RS). This pattern has been observed repeatedly from all eight subjects as well as from our previous research. We hypothesize that these areas might be the first brain regions to reach the critical threshold of [Hb], in order to be seen by fMRI. More studies are needed to clarify this issue.

Lastly, the sharp rising edge observed in the histogram of DSC time lag map (black) in Figure 7 is also explained by the differential sensitivity of each blood flow measure to different vascular compartments. DSC TTP measures are sensitive to arterial and arteriolar signals (i.e. high blood velocity), it is capable of detecting numerous type 1 voxels in which TTP is quite short. This allowed detection of a large number of voxels with relative delays of less than 1 s.

To thoroughly assess these two flows, we reversed the procedure shown in Figure 2 to start with the DSC movie and calculate time lags from the RS time courses. The resulting time lag/TTP graph was shown as Figure 4 in the Supplemental Material. Again, we observed consistently the linear positive correlation between these two dynamic patterns. However, as in Figure 4, it is not a perfect match. We believe it is still largely due to the bias of each method. However, there might be hidden physiological or methodological reasons, which require further studies.

Neuronal vs. physiological LFO

We have demonstrated here and in many of our previous publications that the sLFO in RS data is very robust. They contribute a significant amount of LFO signal to fMRI over widespread brain regions,³⁰ which can influence RSN analyses, potentially resulting in non-neuronal correlations between brain regions. However, we would also like to point out that many procedures adopted in this study, including seed selection (SSS) and cross correlation, etc. are all fine-tuned to efficiently differentiate the sLFO from the BOLD LFO, as sLFO is the focus of this study. We are in no way denying the presence of neuronal signals, which have been demonstrated many times in multimodal^31,32 and invasive studies.³³ We believe both signals contribute to RS BOLD data. Recently, several interesting studies demonstrated the neuronal component in the global BOLD signal.^33,34 Many effective methods have been developed to identify and remove the confounding physiological noises and to expose the real neuronal signal. One major contribution of this manuscript is to demonstrate that some sLFOs are not static. Their dynamic and spatially evolving nature should be considered when developing effective denoising methods for low frequency BOLD signals, such as RS data.³⁵

The main observation of this study is that the evolving spatial patterns of sLFO in RS BOLD fMRI match cerebral blood flow. The origins and functions of sLFO are still not clear. Some evidence indicates that the sLFO likely have an extracerebral origin.⁷ However, as we pointed out in “Consistency and difference between DSC and RS” section, regardless of their origin and function, sLFO will not be visible in fMRI in the voxels heavily dominated by the arterial blood. Thus, we are not able to track the inflow of the blood using sLFO from RS fMRI. On the other hand, the sLFO method is very sensitive to the blood flow in the capillary and vein-dominated voxels containing detectable levels of [Hb]. Compared to DSC, it does not require any bolus injection and can be derived easily from any RS scan. We have tested a similar method in patients with intracranial stenosis, including Moyamoya disease, which show the great potential of this technique to assess clinically important flow parameters.³⁶ More studies are being conducted to validate the method in populations with cerebrovascular abnormalities.

Limitations and future directions

We utilized a very simple approach to calculate the time lag of DSC signals, in which TTP was used to represent the arrival time of the bolus. The benefits of this method are: 1) it does not require any assumptions of the bolus shape, vessel types, etc., 2) the signal to noise ratio around the peak of DSC signal is high, and 3) when fitted with a gamma variate curve, the temporal resolution of lags can be very high, and is not limited by TR. However, there are several factors that might affect the accuracy of the method, first, the dispersion of the bolus. The bolus is injected over time and occupies substantial space in the blood along the length of the vessel,³⁷ which would affect the accuracy of interpretation of TTP as analogous to bolus arrival time. Moreover, there are other effects that also affect the accuracy to a lesser degree, including MR scan sequence effects (from gradient-echo vs. spin-echo sequence),³⁸ leakage of contrast due to blood–brain barrier compromise,³⁹ etc. However, the superior signal to noise ratio makes TTP the most robust parameter for DSC, which offers an accurate measurement in regard to the relative temporal delays.²⁷

We gave an explanation about the discrepancy of the two flow measurements. However, definitive evidence is not available in the data we collected in this study. In future studies, we will introduce gas challenges, such as hypercarbic and hyperoxic challenges to alter blood flow, volume, and oxygenation, leading to visible BOLD signals in the voxels corresponding to the inflow of the blood. We hypothesize that the results of sLFO flow in the gas challenge should closely match the bolus flow of DSC. Moreover, other fMRI methods, such as iVASO^40,41 can be incorporated into the study to directly assess the signals from the arterial compartment. We are also working on quantifying our sLFO flow result to make it comparable to other methods, such as Arterial Spin Labeling. Lastly, the movie and time lag maps (Figure 5) are likely to reflect the properties of the underlying vascular territories. We will increase the spatial resolution of these maps by including more subjects in the future studies and compare these maps with those from cerebral arterial/venous territories.

Conclusion

Previously, we have demonstrated that we can extract sLFO from RS fMRI data and map its dynamic patterns as it moves through the brain. We have hypothesized that it is a natural, endogenous signal and moves with the blood; therefore, the dynamic patterns represent cerebral blood flow. In this study, we tested this hypothesis by conducting both a DSC scan (bolus tracking) and a RS fMRI scan in healthy subjects. By comparing the flow patterns of the bolus with that of sLFO, we found that the flow of sLFO does to a large extent represent the blood flow in the brain. However, due to fact that the contrast agent in BOLD fMRI is deoxy-hemoglobin, which is not readily detectable in arteries, the sLFO method is sensitive mainly to blood flow in the capillaries and veins.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: the National Institutes of Health, Grants K25 DA031769 (YT), R21 DA032746 (BdeBF).