Dynamics of the cerebral blood flow response to brief neural activity in human visual cortex

Abstract

The blood oxygen-level dependent (BOLD) functional magnetic resonance imaging (fMRI) signal depends on an interplay of cerebral blood flow (CBF), oxygen metabolism, and cerebral blood volume. Despite wide usage of BOLD fMRI, it is not clear how these physiological components create the BOLD signal. Here, baseline CBF and its dynamics evoked by a brief stimulus (2 s) in human visual cortex were measured at 3T. We found a stereotypical CBF response: immediate increase, rising to a peak a few second after the stimulus, followed by a significant undershoot. The BOLD hemodynamic response function (HRF) was also measured in the same session. Strong correlations between HRF and CBF peak responses indicate that the flow responses evoked by neural activation in nearby gray matter drive the early HRF. Remarkably, peak CBF and HRF were also strongly modulated by baseline perfusion. The CBF undershoot was reliable and significantly correlated with the HRF undershoot. However, late-time dynamics of the HRF and CBF suggest that oxygen metabolism can also contribute to the HRF undershoot. Combined measurement of the CBF and HRF for brief neural activation is a useful tool to understand the temporal dynamics of neurovascular and neurometabolic coupling.

Introduction

Functional magnetic resonance imaging (fMRI) based on the blood oxygen level dependent (BOLD) contrast is widely used for the quantification of brain function and network connectivity.^1,2 The BOLD signal is sensitive to changes in oxygen extraction fraction (OEF), directly related to underlying physiological responses³; the OEF is modulated by cerebral blood flow (CBF) and oxygen metabolism (CMRO₂) in local parenchyma. Moreover, venules and veins contain a dominant share of deoxygenated blood, so that changes in their cerebral blood volume (CBV) could affect deoxyhemoglobin concentration and thereby the BOLD signal.^4–6 The interplay of these three physiological parameters (CBF, CBV and CMRO₂) is mainly expected to drive the BOLD signal.

Direct measurement of CBF is therefore of strong interest. Arterial spin labeling (ASL) can measure arterial blood delivered to a tissue,^7,8 by pairwise acquisition of label and control images.⁸ Quantitative CBF measurement with ASL has been used to localize brain activation in visual cortex,^9,10 motor cortex,^11,12 and prefrontal cortex.^9,13

ASL-based CBF measurement is useful for investigating the mechanisms of neurovascular coupling, and thereby the interpretation of the BOLD signal.¹⁴ While simultaneous ASL and BOLD measurements have shown correlations between BOLD and CBF signals,^15–17 there have been a few studies that addressed the temporal dynamics of BOLD and corresponding physiological metrics such as CBF, CMRO₂ (and CBV) responses.^4,6,18–20 However, these studies were performed with long-duration stimuli (>10 s), which is not ideal to quantify the dynamics of neurovascular coupling. Lu et al.⁶ estimated the post-stimulus increase of CMRO₂ using a multimodal fMRI scheme – BOLD, ASL and vascular space occupancy – in human visual cortex at 1.5 T. After a 30-s-duration stimulus, they observed that CBF and CBV returned to baseline, while the BOLD signal dropped below baseline (undershoot) for 20–30 s. Again using a long-duration stimulus, Obata et al.⁴ performed simultaneous measurement of BOLD and ASL using a dual-echo spiral sequence²¹ in human motor cortex at 1.5 T. Primary motor cortex showed undershoots in both BOLD and CBF, while a BOLD undershoot was found only in supplementary motor areas. However, inconsistent post-stimulus flow dynamics were reported for similar experimental paradigms.^19,22–25 Moreover, inconsistent multiple peaks during stimulation were commonly found in these studies.

The BOLD hemodynamic response function (HRF) generally refers to the stereotypical BOLD response evoked by a brief (∼2-s) stimulus. Although the shape and underlying dynamics of the HRF vary somewhat, it often consists of an initial delay or dip, a hyperoxic peak, followed by an undershoot with possible oscillatory ringing; total duration of HRF dynamics is 20–30 s.²⁶ Recently, we showed that HRF temporal dynamics were remarkably stable, and the majority exhibited a significant post-stimulus undershoot.²⁷ The stable dynamics of the HRF across cortex justify the frequent assumption of a standard form for the HRF to permit linear analysis of fMRI data. Thus, the HRF evoked by 2-s duration stimuli is of strong interest to the fMRI research community.

A short stimulus paradigm is useful to understand the physiology of BOLD signals for at least two reasons. First, a long stimulus does not necessarily evoke a temporally steady neural response; phenomena such as variable attention and adaptation can modulate the response, making it difficult to distinguish the details of the physiological responses. The use of a brief stimulus separates the fast time-scales of the neural activity (order of milliseconds) from the much slower evoked vascular and metabolic responses (order of seconds); the aggregate activation evokes a temporally stereotypical flow response.^28–31 Second, the physiology is simpler for a short stimulus because we expect that the BOLD signal is driven only by CBF and CMRO₂. Animal experiments have shown prompt arterial volume change and no venous volume changes with a short stimulus, while weak and prolonged venous volume changes were observed with a long stimulus.^32–34 Thus, with a short stimulus, venous CBV changes can be neglected. Moreover, our modeling predicted that a short stimulus would evoke similar CBF temporal dynamics across subjects in early visual cortex.³⁰ In particular, the modeling predicted that CBF responses should feature a substantial undershoot. Correlations between the measurements and model predictions bolstered the likelihood of a stereotypical flow undershoot.

Previously, there have been a few efforts to explore the temporal dynamics of the CBF and BOLD HRF using event-related stimuli.^35–39 These studies generally used flow-sensitive alternating inversion recovery (FAIR) MR sequences combined with stimulus-dithering methods to acquire time-resolved flow and compare these results to the BOLD HRF. However, these studies used coarse spatial resolution (3-mm pixels and slice thicknesses of 7–10 mm) with fairly low contrast-to-noise ratio (CNR), and were therefore unable to resolve the late-time dynamics of the flow response.

Huppert et al.⁴⁰ used methods similar to those we present here to compare the HRF measured by BOLD to that obtained by ASL (as well as optical NIRS) with 0.5-s temporal sampling. These experiments focused on the early temporal dynamics of the HRF, and also utilized coarse spatial resolution (3.75-mm pixels, 6-mm slices). Our experiments provide a useful extension of their results to higher spatial resolution.

Here, we quantify the temporal dynamics of CBF evoked by brief neural activation across human visual cortex with fine spatial resolution at 3T. A brief visual stimulus combined with a fast-paced task produced brief periods of brain activity. A stimulus dithering scheme was used to obtain fast (1.25-s) temporal sampling for ASL. The same stimulus paradigm was used for BOLD HRF measurements during the same session. We found reliable and stereotypical flow dynamics across subjects and ROIs; these dynamics were very similar to that of the HRF. Significant flow undershoots were observed, thus supporting the assumption of an underdamped flow response for brief neural activation offered by our model.^28,30,31 This method provides an effective means to understand neurovascular and neurometabolic coupling by distinguishing and quantifying the underlying flow responses from the BOLD HRF.

Methods

Participants

Seven volunteers (ages 20–59 years, three females) participated in experiments after providing informed consent according to a protocol approved by the local ethics committee for human study, Baylor College of Medicine Institutional Review Board. Our human-subjects experimentation protocol conforms to Baylor College of Medicine's “Ethical and Regulatory Mandate for Protecting Human Subjects,” which places emphasis on the principles in the Belmont Report.

Stimulus

We specifically chose a low-level and abstract visual stimulus to limit variability due to higher-level cognitive responses across subjects. Subjects viewed a display (Cambridge Research Systems, BOLDscreen32) positioned at the back of the scanner bore. Visual stimulation consisted of three circular regions (5° radius) of flickering colored dots (half bright, half dark) on a mean-gray screen. A region of dots appeared at a random selection of one of the three screen locations for 667 ms, followed by a second differently colored (red, yellow, green) region of dots at a different location for the next 667 ms, then a third, Figure 1(a). Subjects were instructed to follow three consecutive dot-region presentations with their eyes and pressed corresponding buttons that matched the color and position of the dot display. Between impulses, there was 35-s inter-stimulus-interval (ISI) for flow evolution (27.5 s for HRF measurements). Subjects were instructed to maintain fixation on a central colored dot and perform a slow-paced, non-challenging color-detection task to encourage a stable, cognitively impoverished mental state. After each ISI, the central fixation dot changed color to cue the subject 0.5 s before the next 2-s duration stimulation period. We measured the subject's performance by analyzing the latency and accuracy of their responses. Subjects practiced the task outside of the scanner to maintain performance level of <600 ms latency and >80% accuracy before a scan session.

Figure 1.

(a) Three circular regions of flickering dots are presented with random spatial orders during 2 s; the example shows yellow-red-green circular regions presentations (left) followed by an ISI that includes a slow-paced fixation-point color-detection task (center), and then second set of circular-region presentations, this time green-yellow-red (right). During the stimulus period, subjects were required to follow these circular regions with their eyes and push buttons that correspond to their colors and positions. (b) Example of prescription for acquisition and labeling regions for ASL MR sequence: 25 oblique quasi-axial slices with 2-mm thickness, 256 mm FOV for functional measurements (blue box), and 100 mm inversion slab (yellow box). (c) T1-weighted image with color overlays delineating ROIs, blue: V1, red: V2, green: V3, cyan: V3AB, and magenta: hV4. (d) Same ROIs on gray-white matter inflated surface.

MRI protocol

FMRI data were obtained on a 3T Siemens Trio scanner (Erlangen, Germany) with a product 32-channel head coil. For flow measurements, experiments were performed on 25 quasi-axial slices with 2-mm thickness (no gap) covering early visual cortex using a pulsed-ASL (PICORE/Q2TIPS tagging); 256-mm field of view (FOV), 128 × 128 matrix, 90° flip angle, 1698-Hz bandwidth and GRAPPA factor 3. The arterial spins were labeled with a 100-mm inversion slab and a 12.5-mm gap inferior to the image slices, Figure 1(b).

We chose TI1 = 700 ms and TI2 = 1300 ms with saturation stop time, TIs = 750 ms, similar to those used in previous ASL studies of the HRF.^18,40 The repetition time, TR = 2.5 s, was the minimum given the desired 2-mm isotropic voxels, so that alternating tag and control measurements were obtained every 5 s. Each run started with baseline flow measurement (no stimulus) for 35 s followed by 16 35-s duration stimulation events. To obtain 1.25-s temporal sampling, 4:1 stimulus-onset time dithering was used. Stimulus onset occurred 1.25 s later on each subsequent presentation, so that the duration of every fourth event was 30 s, whereupon the cycle was repeated. Total acquisition time for each run was 10 min including the initial baseline period.

We first experimented with simultaneous measurements of BOLD and its corresponding CBF response by using a 33-ms TE to obtain BOLD HRF responses from the unlabeled image time series while obtaining CBF response from the label-tag differences. However, we observed significant BOLD/ASL contamination effects (Figure S1); the longer TEs tended to reduce the observed perfusion and blur their dynamics. These simultaneous measurements also strongly underestimated the magnitude of the BOLD HRF, possibly due to magnetization-transfer.^41–43 Therefore, we performed ASL runs at the minimum (13-ms) TE to measure only the flow response.

Separate runs measured the BOLD HRF with the same FOV and voxel size using a blipped-CAIPI^44,45 echo-planar imaging (EPI) with 3× simultaneous multi-slice acceleration and 2× GRAPPA acceleration (1250-ms TR, 33-ms TE, flip-angle 67°, 1562-Hz bandwidth). Each experimental session interleaved four runs of ASL (35 s × 17 events), and two runs of BOLD (27.5 s × 12 events) using the same slice prescription.

A set of T1-weighted structural images (3D FLASH with minimum TE and TR) was obtained with the same slice prescription as functional scans at the beginning and end of each session: 1 × 1×2 mm³ voxels, 256-mm FOV, 32 slices, 15° flip angle, 3.5-min total acquisition time. These images were used to align the functional data to the segmented high-resolution structural reference volume collected from a separate session using a high-resolution (0.7-mm isotropic voxels) MP-RAGE sequence (900-ms TI, flip-angle 9°, 2600-ms TR).^46,47

At the end of each session, we obtained one 2-min resting-state proton-density-weighted EPI run with the same slice prescription to create an intensity calibration volume for ASL quantification: 4-s TR, minimum (13-ms) TE, flip-angle 90°, 1698-Hz bandwidth, 2 × GRAPPA acceleration and 7/8 partial Fourier.

Data preprocessing

ASL data were split into tag and control time series, then intensity-based motion correction performed.⁴⁸ Sequential slice acquisition timing was corrected followed by slow baseline-intensity drift compensation using a high-pass filter.

The stimulus-onset dithering scheme described above was used to obtain high temporal resolution. Analysis requires the use of an interpolation scheme throughout the time series, so it is not possible to utilize standard approaches such as pairwise or surround subtraction. Instead, both tag and control data were 4:1 upsampled using Hermite-polynomial cubic interpolation to estimate tag and control values at each stimulus time point. This is similar to surround subtraction because each interpolated time point preserves temporal details from surrounding time points; however, our approach differs because both tag and control measurements are separately interpolated by cubic interpolation to obtain the desired temporal resolution. After upsampling, dither offsets were removed by circular shifts of each HRF time series (Figure S2). Note that this procedure greatly expands the temporal bandwidth of the ASL measurement, and this expansion is partly dependent on the form of interpolation used. We experimented with cubic spline (smoothest, least bandwidth expansion), hermite-polynomial (intermediate bandwidth), and sinc (widest bandwidth) interpolation. All three produced similar results, but dynamics became noisier as bandwidth increased. We chose the Hermite-polynomial approach as a compromise between temporal resolution and signal-to-noise ratio (SNR). To avoid the effects of under-sampling at the end of each event created by the dithering, we chose a longer event duration for the CBF measurements (35 s) than for the HRF measurements (28 s). We preserved only the first 27.5 s of the CBF time series in each event for subsequent analysis.

We performed the same post-processing for the BOLD HRF data excluding the control-tag split and dither correction (no onset dithering was used during stimulus presentation). All functional data, both CBF and BOLD, were then spatially transformed to a high-resolution reference volume anatomy using the same robust intensity-based alignment algorithm⁴⁸ applied to the T1-weighted inplane image volume mentioned above.

ROIs

To quantify regional changes in the CBF, we used visual-area ROIs estimated with population receptive-field methods from a separate session for each subject,^49,50 Figure 1(c) and (d). Data were analyzed in the portions of each prescription that overlapped with seven visual areas: V1, V2V, V2D, V3V, V3D, hV4, and V3AB. However, in one subject, S7, V3AB was not well delineated, and in another, S6, the functional prescription did not cover V2V, V3V, hV4, and V3AB. We then selected a subset of voxels in the middle of the gray matter for each ROI using a normalized distance mapping scheme^30,51 to minimize partial volume effects from outside of the gray matter. We treated the signal observed in each ROI as an independent sample of the BOLD and CBF dynamics. Altogether, this yielded 44 individual flow and BOLD dynamics measurements from seven subjects.

Data analysis

To obtain quantitative CBF from our ASL data, we used the standard kinetic model;¹⁸

$$\text{CBF}=\frac{\lambda \Delta \text{M}}{2\alpha ·{\text{M}}_{0\text{A}}\text{TI}1·\text{exp}(-\text{TI}{2}_i/\text{T}1\text{A})}$$

where ΔM is the difference between control and tag signals, α is labeling efficiency, M_0A is equilibrium arterial blood magnetization (from the calibration scan), TI2_i is TI2 for slice i, and T1A is T1 of arterial blood (1664 ms).

Baseline flow was measured at the beginning of each run for 35 s without stimulus and task. Four 35-s baseline flow measurements in each subject were averaged within each ROI.

We obtained 64 CBF events for each ROI in each subject. We averaged these together within each ROI. We set the mean of the first and last points equal to the measured baseline flow in each CBF time series; then we calculated percentage flow changes. The flow time series was characterized by its peak amplitude (and percentage change), time-to-peak (TTP), full-width-half-maximum (FWHM), undershoot amplitude (and percent change), and time-to-undershoot (TTU). BOLD HRF data were analyzed in a similar way, except a zero baseline assumed as appropriate for these relative measures of functional contrast.

We used a bootstrapping scheme^52–54 to test the statistical significance of the flow measurements. We resampled the 64 repeated flow measurements obtained in each session and ROI with replication, followed by averaging and processing to extract the various parameters, including peak, undershoot, and several timing parameters. This procedure was repeated 1000 times to estimate the parameter distributions. Confidence intervals and p values can then be evaluated from these distributions. This procedure was particularly useful to test for the existence of a post-stimulus flow undershoot, which has long been controversial. We considered an undershoot as significant when >95% of the bootstrapped undershoot amplitudes were less than the baseline flow (p < 0.05). In addition, we defined CNR as a ratio of the mean peak amplitude to the length of its confidence intervals, which was used as a metric for reliability of the flow dynamics.

Results

We obtained strong and significant measurements of baseline flow across early visual cortex on both hemispheres, Figure 2(a). Our central gray-mater selection method successfully excluded most strong comfounding signals from feeding pial arteries, Figure S3. We observed spatial variability in flow, with mean values ranging from 30 to 85 ml/100-g/min across subjects. There was a spatial gradient of baseline perfusion, with weaker perfusion in dorsal areas becoming stronger ventrally. Baseline flow responses showed interesting and significant differences among visual areas (Figure 2(b)). The pattern of perfusion across the visual areas matched the general dorsal-to-ventral gradient, with significantly higher perfusion (p≪ 0.001) in areas hV4, V3V, and V2V than in V2D, V3D, and V3AB in S2, S3, and S4. Similarly, in S1 and S5, we found significantly higher perfusion (p ≪ 0.001) in hV4, V3V, and V2V than in V3D, and V3AB. When averaged across all subjects (Figure 2(c)), the differences remain significant (p < 0.003). Moreover, a weak peak becomes evident in V2V, while perfusion in V1 is significantly larger than any of the dorsal areas (p < 0.002) but not significantly smaller than the ventral regions.

Figure 2.

(a) An example of baseline flow on the gray-white surface in early visual cortex for one subject; each ROI is delineated with a color-coded outline. (b) Mean and standard deviation of baseline flow measurements in each subject and ROI, and (c) mean and standard deviation of baseline flow measurements across all subjects.

Our 2-mm-isotropic-voxel ASL measurement with brief stimuli produced reliable responses. CBF measurements of peak flow were significant (p < 0.003) in all ROIs and subjects. Mean and standard deviation of the CNR for peak CBF across subjects was 6.4 ± 3.6, and CNR > 3 was found for all subjects and ROIs, Figure 3(a). CNR was ∼3 × higher for the BOLD HRF (19 ± 5.7) than CBF, Figure 3(b); however, their spatial pattern was similar, Figure 3(c).

Figure 3.

Histograms of the CNR for (a) CBF and (b) HRF across subjects and ROIs. (c) An example of the CNR spatial patterns for CBF (left) and HRF (right) on the gray-white surface in early visual cortex for one subject; each ROI is delineated with a color-coded outline.

All CBF responses (Figure 4(a), top, green lines) showed similar dynamics, starting with a prompt increase during the stimulus, reaching a peak a few seconds after the stimulus, followed by an undershoot with possible ringing; mean (solid black line) and standard deviation (dashed black lines) across subjects and ROIs. Similar dynamics were observed for the BOLD HRF (Figure 4(a), bottom). For two individual subjects (Figure 4(b)), both CBF (green lines) and HRF (orange) responses across visual areas also display similar responses with stereotypical dynamics. TTP was similar between CBF and HRF responses. A flow undershoot was frequently evident in both subjects, but its amplitude and timing were more diverse than the peak amplitude and timing within each subject. Spatial pattern of the flow peak amplitude (Figure 4(c), left) was similar to that of the HRF (right). We found strong CBF and HRF responses in most ROIs, but some weak and negative responses also presented in ROIs apart from primary visual cortex (e.g. V3AB, and hV4).

Figure 4.

(a) Time series of 44 individual flow (upper panel, green lines) and HRF (bottom panel, orange lines) responses. Mean (solid black line) and standard deviation (dashed black) are also shown. (b) Examples of CBF and HRF responses for two example subjects. Each line shows time series of CBF (green) and HRF (orange) for each ROI. (c) An example of the spatial patterns of peak amplitude for CBF (left) and HRF (right) on the gray-white surface in early visual cortex for one subject; each ROI is delineated with a color-coded outline.

Figure 5 shows histograms of amplitude and timing parameters of the 44 CBFs and HRFs across subjects and ROIs. Flow-response parameters were: peak 23.9 ± 11%, undershoot −9.3 ± 5.6%, FWHM 6.3 ± 0.9 s, TTP 6.2 ± 0.9 s, and TTU 14.3 ± 1.5 s. Note that flow amplitude is percentage change based on baseline flow amplitude. HRF-response parameters were: peak 1.39 ± 0.4%, undershoot −0.46 ± 0.18%, FWHM 4.8 ± 0.7 s, TTP 6.4 ± 1 s, TTU 13.3 ± 2.5 s. The similar values for each CBF parameter across ROIs and subjects suggest stereotypical flow-response dynamics. The HRF FWHM was significantly narrower than that of the CBF (p ≪ 0.001). There was no significant difference for both TTP and TTU between CBF and HRF. However, the distributions of the CBF and HRF TTPs were narrower than their corresponding TTU distributions.

Figure 5.

Histograms of (a) CBF and (b) HRF parameters across subjects and ROIs.

We examined the significance of the undershoot using bootstrapping. When averaged across all subjects/ROIs, the flow and HRF undershoots were very significant (p ≪ 0.001); 34/44 of the flow undershoots were significant (p < 0.05, Figure 6(a)), while HRF undershoots were significant in all 44 ROIs, Figure 6(b). Significant flow undershoots were evident across most ROIs in most subjects, except for S5. There was a qualitatively similar pattern observed between CBF and HRF undershoots. For example, strong reliable undershoots were evident in S2, while weaker but reliable undershoots were seen in S7 for both CBF and HRF.

Figure 6.

Undershoot amplitude of (a) CBF and (b) HRF for subjects and ROIs; 34/44 flow undershoots were significant, while 44/44 HRF undershoots were significant. The undershoot in each ROI and subject was considered as significant when >95% of the bootstrapped undershoot amplitudes (n = 1000) were less than the baseline flow (p < 0.05).

We evaluated correlations between CBF and HRF parameters, Figure 7. We found strong and significant correlations of their peak amplitudes, Figure 7(a): R² = 0.46 (p < 0.001), with moderate and significant correlation for their undershoot amplitudes (Figure 7(b)): R² = 0.28, p = 0.002. Remarkably, these correlations disappeared when the peak and undershoot flow responses were normalized by their baseline values (Figure S4). For timing parameters, we found strong and highly significant correlation for TTP (R² = 0.45, p < 0.001; Figure 7(c)), and moderate but significant correlation for TTU (R² = 0.33, p < 0.001; Figure 7(d)). Also, there was moderate but significant correlation for FHWM between HRF and CBF (R² = 0.35, p < 0.001; Figure 7(e)). Strong and significant correlations between TTU and TTP were also found for CBF (R² = 0.66, p < 0.001; Figure 7(f)), and HRF (R² = 0.42, p < 0.001; Figure 7(g)).

Figure 7.

Correlations between HRF parameters and CBF parameters (a–g); correlations between baseline flow and (h) flow peak amplitude and (i) HRF peak amplitude across subjects and ROIs (44 data points).

Strong correlations were also observed between peak amplitudes and baseline flow. The correlation between peak and baseline flow (Figure 7(h)) was remarkably strong (R² = 0.88, p < 0.001); the strength of neurovascular coupling reflects the baseline perfusion evident in each visual area. A similarly significant but moderate correlation (R² = 0.37, p < 0.001), was also evident between the BOLD HRF and baseline flow (Figure 7(i)), consistent with the flow-to-baseline correlation just mentioned, and the BOLD-to-flow correlation shown in Figure 7(a). Thus, the strength of the BOLD response also reflected the baseline perfusion in each visual area.

Discussion

We successfully quantified the CBF dynamics evoked by brief stimulation in multiple regions of human visual cortex. A short stimulus produced relatively stereotypical CBF dynamics that included a flow undershoot, consistent with our theoretical modeling results.^30,31 We utilized a stimulus-onset-time dithering scheme to provide an effective 1.25-s sampling. We combined this with 2-mm isotropic voxel sampling and careful segmentation methods to focus our ASL measurements specifically on cortical gray matter. Our repetitive ASL measurements produced reliable flow time series, which can offer calibrated flow quantification in cortical gray matter similar to that of BOLD measurements.

Our 2-mm-isotropic-voxel ASL measurements obtained flow dynamics with satisfactory CNR, despite using much higher spatial resolution than typical ASL studies of the human brain.^8,55,56 There are at least two possible reasons for this high-quality performance. First, in both flow and BOLD signals, thermal noise is not dominant, and physiological noise (which is mostly a form of nuisance rather than uncorrelated noise) is the primary contributor to the total image noise for 2-mm isotropic voxel sampling at 3T. Because physiological noise is also proportional to the signal strength, which in turn decreases total noise with decreasing voxel size, higher resolutions can improve CNR in this physiological noise-dominant regime.^57,58 Second, the use of 2-mm isotropic voxel size to sample only the center of the gray matter can improve CNR by decreasing partial volume effects. Weak signals in the white matter worsen the CNR by decreasing contrast, while relatively strong but noisy signals from the pial vasculature can contaminate and confound signals from the gray matter.⁵¹ The combination of these two effects appears to have improved CNR for both measurements.

One of the main goals of this study was to understand the detailed dynamics of brain functional activation. There have been a number of efforts to examine brain functional activation physiology using multi-modal MRI methods.^{4,6,19,22–25} However, relatively long stimuli were used in these studies, and complex CBF dynamics were observed, e.g. multiple peaks during the stimulation period, and slow decay of the CBF back to the baseline post-stimulus. In recent optical imaging experiments,³² such results were interpreted to be consistent with prompt non-linear volume changes in surface arterioles and slow and steady non-linear venous volume changes during and after long stimuli. Moreover, long stimuli do not necessarily evoke a temporally steady neural response; phenomena such as variable attention and adaptation can further confound the results. Thus, it is far more difficult to distinguish the details of the physiological response evoked by long-duration stimuli.

However, for a brief stimulus, arterial impulse volume changes were observed in previous animal studies with no significant venous volume changes.^32–34,59 Here, we employed a short stimulus to generate the BOLD HRF and its corresponding CBF component. Because neuronal activity generally closely follows the temporal window of the stimulus, the use of short stimuli separates the fast time-scales of the neural activity and the much slower evoked vascular and metabolic responses. A short stimulus is also beneficial because we expect only arterial volume changes; we can exclude BOLD changes introduced by non-linear venous volume changes. Additionally, a short stimulus is known to produce a stereotypical BOLD HRF response consisting of hyperoxic peak, followed by undershoot and possible ringing.^29–31 These stereotypical BOLD dynamics are the basis of linear approaches to fMRI data analysis.

Our measurements substantially expand upon previous efforts to resolve the temporal dynamics of the flow HRF component with short-duration stimuli.^35–39 Those measurements used very coarse spatial resolution (particularly in the slice direction) and were carried out at lower field strength (1.5 or 1.9T). Most also used FAIR sequences to measure flow. So, direct comparison of their results to our own is somewhat difficult. Nevertheless, all of the studies showed rough correspondence between BOLD and CBF responses but noted small timing differences. One of the studies³⁹ obtained similar time-to-peak for the flow response as compared to our own (6.2 s), but the others observed faster flow responses (3.7–4.2 s), which may reflect earlier contributions from the superficial pial vasculature because of their coarse resolution. These studies did not attempt to assess the details of the late-time behavior, but Yee et al.³⁹ did show a significant flow undershoot 15 s after stimulus onset, and Su et al.³⁷ showed an undershoot in their data but did not characterize its significance or discuss its mechanism.

A more recent study used MRI methods very similar to our own at 3T,⁴⁰ but used a somatosensory stimulus. Their results showed a significant flow undershoot, which they noted but did not discuss at length. They also observed faster time-to-peak for their flow responses than we observed, but once again their coarser spatial resolution (3.75-mm pixels and 6-mm slices) could reflect dominant contributions from pial arteries. Faster CBF TTP could also correspond to the different region of cortex studied (somatosensory versus visual).

A larger FWHM was observed for the CBF than for the HRF, which was somewhat unexpected. Because the BOLD signal is dominated by downstream deoxygenated blood,^30,60,61 one might expect earlier flow responses with a narrower FWHM. There are at least two possible explanations for this observation. First, it could be partial-volume effects from feeding pial arteries. Although most voxels with confounding signals in the gray matter were excluded with our central gray-matter selection methods, there were a few selected gray-matter voxels that were very close to artifactually large signals near the superior sagittal sinus or adjacent to the tentorium (Figure S3), which could broaden the CBF FWHM. Another confound remains BOLD contamination from these sinus regions, which can be strongly driven by our stimulus. Alternatively, the larger CBF FWHM could reflect the actual dynamics of microvascular perfusion. One mechanism can be heterogeneous transport times for arterial blood to reach different layers of the gray matter. Animal measurements during activation indicate that arterial vasodilation occurs relatively faster in deeper layers than superficial,^62,63 which is consistent with human fMRI measurements of HRF temporal latency.^51,64 Our 2-mm-isotropic sampling was likely to capture signals from both superficial and deep gray matter, which can also broaden the CBF response. Moreover, we found a significant FWHM correlation between CBF and HRF, Figure 7(e), suggesting that the larger CBF FWHM results from tight neurovascular coupling from capillaries. However, the group results were also strongly influenced by two subjects (S4 and S5) with particularly wide FWHM in their CBF response, Figure S5, which exaggerated the observed differences with the HRF. Further studies with higher spatial resolution will be necessary to discern the layer-dependent temporal characteristics of the CBF.

Post-stimulus flow undershoot in the brain has been controversial. Flow undershoot was reported in 1990s using laser-Doppler flowmetry^65,66; the authors were unsure of its mechanism and its reliability. Later, with ASL methods, an undershoot was also found in some studies,^{6,19,22,25,40} while no undershoot was observed in others.^18,24 In one study, a simple stimulus was found to evoke an undershoot only in primary motor cortex, but not in supplementary motor areas.⁴ In another study, an undershoot was observed with long forepaw stimulation (1 min), but no significant undershoot was evoked by a short stimulus (4 s).⁶⁷ These inconsistent observations are possibly the result of coarse spatial resolutions. More recently, Mullinger et al.⁶⁸ used ASL methods to find a flow undershoot following a long-duration stimulus period that showed a reliable correlation to inhibitory neural activity. In general, the use of long-duration stimulus paradigms could introduce complexity and non-linearity as we described above.

The flow undershoot needs to be more fully characterized. We demonstrated reliable flow undershoots in cortical gray matter using a brief stimulus. We found significant correlations of both amplitude and timing of the undershoot between CBF and HRF, suggesting that the CBF undershoot contributes to the HRF undershoot. However, there were difference in the late-time temporal dynamics of CBF and HRF. Specifically, we observed that the CBF undershoot occurred later but more reliably than the HRF undershoot, Figure 5. Altogether, our data suggest that the late-time behavior of the HRF after the hyperoxic peak is driven by multiple factors. Previous experimental studies showed slow decay of late-time CMRO₂ for brief stimulus, possibly another source of the HRF undershoot.^69–71 In fact, our previous modeling efforts attempted to separate the dynamics of CBF and CMRO₂ from measured BOLD HRFs³⁰ and predicted some variation of CBF responses as well as prolonged decay of CMRO₂ responses. Generally, the late-time behavior of the HRF is less stable from trial-to-trial than the early-time behavior,^27,28,30,31 suggesting variable contributions from both CBF and CMRO₂ and possibly other factors.

Our measurement of the CBF response showed stereotypical dynamics across subjects and ROIs, consisting of an initial peak followed by an undershoot and possible ringing. These dynamics are consistent with the assumption of an underdamped sinusoidal flow response used in our previous modeling efforts, motivated by the proximal-integration hypothesis.^31,59 A vasodilation signal proximal to neural activity in capillaries propagates upstream and generates brief vasodilation in pial arteries. This impulsive upstream arterial volume change generates an underdamped flow response to mediate oxygen delivery to downstream parenchyma. The oscillatory flow character is a consequence of inertial effects in upstream pial arteries because of their relatively large diameters and fast flow velocity.²⁸ Our demonstration of reliable flow undershoot (Figures 5 and 6) supports the hypothesis of underdamped oscillation.

Correlations between CBF and HRF parameters confirmed our earlier modeling predictions.³⁰ CBF and HRF peak amplitudes were strongly and significantly correlated (Figure 7(a)), consistent with a strong role of blood flow on the peak of the HRF. Interestingly, this correlation disappeared when peak flow was normalized (Figure 7(b)). Instead, the fractional changes in peak flow were relatively constant (Figure 4(a), 24.9 ± 12.5%) across subjects and ROIs, again consistent with our previous modeling results (21.6 ± 10.5%). Strong temporal correlations between CBF and HRF (Figure 7(c) to (e)) further support a biomechanical contribution of blood flow to BOLD contrast.

Baseline flow varied across ROIs and showed significant differences between dorsal and ventral visual areas in posterior occipital lobe, Figure 2(a). These differences were significant between all ventral and dorsal regions studied, with a weak peak observed in V2V. On a finer spatial scale, there are large variations across each visual area, possibly the consequence of variable perfusion between the superficial gyri and the banks of the sulci.^72,73 The apparent perfusion could also reflect variations in cortical thickness; the thinner cortex can show lower perfusion because of relatively greater partial volume effects. Further experiments at higher spatial resolution will be necessary to fully understand this interesting observation.

Baseline flow was strongly correlated with the peak flow evoked by a brief stimulus. Thus, flow responses were very strongly modulated by baseline perfusion. The neurovascular flow response is apparently dependent on these regional variations in perfusion, rather than local metabolic needs. Because the peak of the BOLD HRF is also moderately correlated with the baseline perfusion, it is likely that ventral early visual areas will yield greater BOLD signals than dorsal areas for similar levels of activation. However, it is possible that the more weakly perfused dorsal areas may instead receive longer periods of flow increases in compensation for their weaker peak values. This is consistent with our observation that dorsal early visual cortex exhibited larger FWHM and longer TTP for the BOLD HRF.²⁷ In any case, the variability in perfusion appears to modulate the strength of the BOLD response in a fashion independent of the stimulus or task, a nuisance modulation that needs to be considered when interpreting BOLD responses across cortex.

Several previous results appear to be inconsistent with our positive correlation between baseline CBF and BOLD response. For example, in some studies, caffeine produced a baseline CBF reduction with no significant change^74,75 or an increase^76,77 in BOLD activation. A lower BOLD response was reported when baseline CBF was increased by acetazolamide.^12,78 However, baseline flow changes induced pharmacologically may be fundamentally different than the natural variations in resting-state baseline. It is possible that the pharmacological manipulations induced vasoconstriction or vasodilation that also altered arterial volume changes evoked by neural activation, fundamentally changing the hemodynamics of the BOLD response. However, in this study, the variations in baseline flow may be the consequence of spatially variable vascular densities⁷³; this hypothesis is consistent with the greater flow observed on the gyri as compared to the sulci. Thus, our experiments observed that normal variations in hemodynamic coupling appear to follow spatial patterns of baseline perfusion, which are likely to be different from the dynamics evoked by pharmacological perturbations from normal perfusion.

Apparently, the fractional CBF response evoked by a brief, repetitive activity is largely constant, with the majority of the variability controlled by baseline perfusion. This supports the concept that the blood-flow response is fairly simple biomechanical process. That supposition is consistent with the observed regularity of cortical microvascular structure that does not respect columnar boundaries.⁷⁹ Such a stereotypical mechanical nature of the flow response would offer a powerful and attractive option to infer oxygen metabolism by the combined measurement of CBF and BOLD responses, together with the regional baseline perfusion, which avoids any need for carbogen inhalation or similarly uncomfortable challenges.

The use of short stimuli simplifies the physiological mechanisms that underlie the HRF, opening up the possibility that CMRO₂ changes can be inferred from the BOLD response. Here, we demonstrated the quantification of CBF dynamics with sufficient temporal resolution, a key step for estimating CMRO₂ responses corresponding to the HRF. Similar experiments that combine simultaneous BOLD and ASL measurements with a detailed computational model could then estimate CMRO₂ dynamics. Previously, we developed a computational model to estimate CBF and CMRO₂ responses associated with the HRF,³⁰ but it utilized a particular parametric description of the CMRO₂ dynamics. Further development of a computational model to estimate CMRO₂ will be necessary to enable detailed quantitation of neurovascular and neurometabolic coupling in the human brain.

We used pulsed ASL to capture the flow dynamics and separately used BOLD-EPI to obtain the HRF. Initially, we tried to use the control portion of the ASL measurement to obtain BOLD signal. However, a long TE produced a BOLD-contamination effect. Although it is not ideal to use separate measurements for CBF and HRF, our approach with repetitive measurements successfully captured flow dynamics as well as the HRF. Future work will implement dual-echo sequences (short TE for ASL and long TE for HRF) to permit simultaneous measurements.^4,19,21 However, our results (Figure S1) indicate that the BOLD response obtained from such methods will be substantially reduced in magnitude, but it should be possible to correct for this with a scale factor. We dithered the stimulus onset time to provide temporal resolution. A dithering approach can remediate a critical drawback of ASL, long acquisition time, to obtain flow dynamics evoked by a short stimulus. However, this approach reduces the number of sampling intervals, thereby decreasing SNR; a simultaneous measurement should partly alleviate this drawback. It would also open up the possibility of model-based estimates of CMRO₂ based on combined measurements of ASL and BOLD.

Conclusion

CBF dynamics are qualitatively similar to HRF dynamics, indicating that the neurovascular coupling provides a dominant drive for the early peak in BOLD contrast. The consistent fractional change in peak flow indicates that neural activation in local parenchyma induces transient flow responses that couple to the pial surface, supporting the mechanism of proximal integration. Subsequent flow undershoots are reliable and significantly correlated with the HRF undershoots, suggesting substantial contribution of the flow undershoot on the HRF undershoot. The flow undershoot confirms earlier modeling predictions of an underdamped flow response. Our measurement and modeling methods open up the possibility to infer neural metabolism to understand the physiology of neurovascular and neurometabolic coupling.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by NIH K25 HL131997, and NIH R01 NS095933 and NIH R01 EB028297.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Authors' contributions

Conceptualization, JK, AT, DJW, and DR; Methodology, JK, AT D.JW, and DR; Investigation, JK, AT, XZ, and DR; Writing–original draft, JK, AT, and, DR; Writing–Review and Editing, JK., AT, DJW, XZ, and DR; Funding acquisition, JK, and DR; Resources, JK, DJW and DR; Supervision, JK and DR.

Supplemental material

Supplemental material for this paper can be found at the journal website: http://journals.sagepub.com/home/jcb