Resting‐State fMRI: Emerging Concepts for Future Clinical Application

Resting‐state functional magnetic resonance imaging (rsfMRI) has been developed as a method of investigating spontaneous neural activity. Based on its low‐frequency signal synchronization, rsfMRI has made it possible to identify multiple macroscopic structures termed resting‐state networks (RSNs) on a single scan of less than 10 minutes. It is easy to implement even in clinical practice, in which assigning tasks to patients can be challenging. These advantages have accelerated the adoption and growth of rsfMRI. Recently, studies on the global rsfMRI signal have attracted increasing attention. Because it primarily arises from physiological events, less attention has hitherto been paid to the global signal than to the local network (i.e., RSN) component. However, the global signal is not a mere nuisance or a subsidiary component. On the contrary, it is quantitatively the dominant component that accounts for most of the variance in the rsfMRI signal throughout the brain and provides rich information on local hemodynamics that can serve as an individual‐level diagnostic biomarker. Moreover, spatiotemporal analyses of the global signal have revealed that it is closely and fundamentally associated with the organization of RSNs, thus challenging the basic assumptions made in conventional rsfMRI analyses and views on RSNs. This review introduces new concepts emerging from rsfMRI spatiotemporal analyses focusing on the global signal and discusses how they may contribute to future clinical medicine.

Functional magnetic resonance imaging (fMRI) is a noninvasive method of measuring brain activity that is widely used in both neuroscience and clinical medicine.Although several types of contrast mechanisms have been proposed for fMRI, such as arterial spin labeling (ASL) 1,2 and vascular space occupancy, 3,4 gradient-echo blood oxygenation leveldependent (BOLD) contrast is by far the most frequently used. 5,6BOLD was first described by Seiji Ogawa as contrast based on paramagnetic deoxyhemoglobin in venous blood. 7,8gawa et al not only described the BOLD contrast mechanism but also showed that BOLD could provide noninvasive real-time mapping of altered metabolic demand or blood flow in rat brains at 7 T, 9 paving the way for its application to functional neuroimaging.][14][15] Like many other functional neuroimaging techniques, such as those using positron emission tomography (PET), single-photon emission computed tomography (SPECT), and ASL, BOLD fMRI relies on the basic fact that neuronal activation and cerebral blood flow (CBF) are coupled. 16Neural activation leads to a local CBF increase of up to 80%, whereas the increase in the cerebral metabolic rate of oxygen is only 25%. 17 This gap increases the oxygenation level of local venous blood, thereby increasing the BOLD signal within the activation area.Since it is based on hemodynamic changes, the BOLD response is a slow signal variation in the order of seconds, with a peak time of approximately 3-5 sec, 18,19 making it possible to track changes with a repetition time of several seconds (Fig. 1).In task-based fMRI, the analysis is often based on voxel-wise general linear model regression, 23 which can be expressed as a function of time as follows: where the response variable y t ð Þ comprises the measured fMRI time series, the explanatory variables x i t ð Þ, i ¼ 1::N i are the predicted time series for each task i, β i denotes the N i coefficients or weights to be estimated, and ε t ð Þ is the noise.The prediction is generally performed by assuming that the BOLD signal is the output of a linear time-invariant system 19 -that is, that the form of the response is independent of time-and that the responses to successive stimuli are superposed in a linear fashion, which allows us to express the predicted time series x i t ð Þ as the convolution of a stimulus function u i t ð Þ with a canonical hemodynamic response function h τ ð Þ as follows: Thus, we can take advantage of neurovascular coupling to infer the site and time course of neural activation by analyzing the MRI signal changes caused by vascular responses, as long as the assumptions about the hemodynamic response function hold (Fig. 2).

Resting-State fMRI
Thirty years after the inception of fMRI, the annual number of relevant studies continues to increase.A PubMed search using the criteria (brain OR cerebral) AND (functional magnetic resonance imaging OR functional MRI OR fMRI) yields 41,906, 106,180, and 184,135 articles related to fMRI published during 1992-2002, 2003-2012, and 2013-2022, respectively.One reason for this continuous growth is the emergence of resting-state fMRI (rsfMRI).RsfMRI is a FIGURE 1: Schematic of the physiological basis of a blood oxygen level-dependent (BOLD) response.The BOLD effect reflects uncoupling between changes in the cerebral blood flow (CBF) and cerebral metabolic rate of oxygen (CMRO 2 ).While neural activation leads to a local CBF increase of up to 80%, the increase in CMRO 2 is only 25%. 17 This gap increases the oxygenation level of local venous blood, thereby increasing the BOLD signal within the activation area due to a reduction in paramagnetic deoxyhemoglobin.Since it is based on hemodynamic changes, the BOLD response is a slow signal variation in the order of seconds, with a peak time of approximately 3-5 sec, 18,19 making it possible to track changes with a repetition time of several seconds.The initial dip (not always seen) is considered to reflect a rise in deoxyhemoglobin due to an increase in metabolism before the hemodynamic response starts. 20The signal increase is followed by a post-stimulus undershoot period, whose origin is still debated. 21The bottom left part shows maps of baseline CBF and CBF increments (mL/minute/100 g) measured using 3D arterial spin labeling while a healthy subject performs a hand-grasping task (superior view).CBF increases in the bilateral primary motor cortices.The bottom right part shows the fMRI signal time course using the same task (block design with alternating control and task epochs of 24 sec).Adapted from Amemiya et al. 22 method of investigating spontaneous neural activity, often based on its macroscopic organization characterized by the activity's coherence.Although rsfMRI is based on the same contrast mechanism, rsfMRI analysis strategies often differ from those of task-based fMRI.This is because, unlike in task-based fMRI, the time course of neural activity is unknown in rsfMRI.Therefore, rsfMRI analyses often rely on the temporal correlations of the signal time series between two regions, which are supposed to reflect the synchronization of spontaneous neural activities between them; hence, it is called "(functional) network connectivity."Although Biswal et al's 26 seminal work on rsfMRI focusing on the motor cortex dates back to the early days of fMRI, the significance of this approach did not become obvious until Greicius et al 27 showed that the so-called default mode network (DMN) could be replicated using rsfMRI correlation analysis and that the correlation is decreased in patients with Alzheimer's disease. 28Using PET, Raichle et al 29 originally defined the DMN as areas showing more intense activity during the resting state than when performing certain goal-oriented tasks, which led to the hypothesis that these regions constitute a network supporting a default mode of brain function.Following the discovery of the DMN, many other networks have been found similarly using rsfMRI. 30These networks, termed resting-state networks (RSNs), 31 are closely related to the anatomical connectivity among neural subsystems revealed by a wide variety of visual, sensorimotor, and cognitive task paradigms 32,33 (Fig. 3a).In short, rsfMRI has made it possible to identify multiple functional networks on a single scan of less than 10 minutes without assigning tasks to patients.It is easy to implement even in clinical practice, in which special devices for stimulating or recording responses are not always available, and patients are not always able to perform tasks.These advantages have accelerated the adoption and growth of rsfMRI.
Other than connectivity or network analyses based on temporal correlations, several rsfMRI metrics are used to indicate the magnitude of local neural activity, as CBF measurements are used for this purpose in other functional imaging techniques.For example, the amplitude of low-frequency fluctuations (ALFF) 26,34 is a voxel-wise measurement of the averaged amplitude of the low-frequency fluctuations of the rsfMRI time series, which is supposed to reflect the magnitude of spontaneous neural activity in each region.Not limited to ALFF measurements, the target frequency of rsfMRI analysis is generally a low-frequency component around 0.01-0.1 Hz (Fig. 3b).This is because frequencies outside this range are contaminated with noise from the scanner drift (0-0.015Hz) 35 and systematic physiological events, such as respiration (0.1 Hz) and heart rate (1 Hz). 36,37In other words, rsfMRI basically relies on the assumption that the  24,25 Neural activity leads to an increase in K + in the perivascular space surrounding microvessels to generate local hyperpolarization of the endothelial membrane.Hyperpolarization spreads to adjacent smooth muscle cells and induces smooth muscle relaxation and arteriolar dilation.Rapid arteriole dilation is also induced by nitric oxide (NO) binding.The increase in arteriole diameter leads to an increase in blood flow, thereby increasing the venous signal through a reduction in paramagnetic deoxyhemoglobin.This complex process is modeled using the hemodynamic response function (HRF) h t ð Þ in a task-based blood oxygen level-dependent (BOLD) fMRI analysis using a general linear model.The stimulus function (t) is convolved with HRF to produce a predicted signal time course x t ð Þ, assuming that the BOLD signal is the output of a linear time-invariant system. 19arget signal time series reflects neural activity based on 1) its spatial distribution, which corresponds to large-scale neural networks and 2) its temporal distribution (i.e., frequency range) outside the typical range of noise.Although other techniques exist, such as using physiological recordings to track respiration-and heart rate-related changes 37 or determining whether signal fluctuations are related to changes in R 2 * or in initial signal intensity (S0) based on echo-time dependence analysis, 38,39 no single approach is thought to completely remove nonneural contributions from rsfMRI data.As a matter of course, no measurement is free of noise, and its effect should matter only when its levels are sufficiently high to affect the measurement of the target signal.However, the effect of noise on fMRI is not trivial and the proportion of the signal variance attributed to neurobiological activity in rsfMRI is estimated to be small, in the order of 5%-20% of the total signal variance. 40,41Therefore, rsfMRI generally requires multiple preprocessing steps, such as spatial distortion corrections, realignment of volumes to compensate for subject motion, and regression of motion parameters to improve its detection performance. 42,43Global signal regression-that is, the regression of the global mean signal-is a preprocessing step that is widely used in rsfMRI studies.However, there has been considerable controversy over its use, rendering it particularly challenging in a consortium study. 44One of the main reasons is that, while it removes the global signal component that mostly arises from physiological events and motion, 45 it has been both theoretically and empirically shown that global signal regression mathematically introduces significant artifacts-namely, negative correlations-thereby significantly altering the interpretation of rsfMRI findings. 46lthough it has drawn less attention than the local network (i.e., RSN) component, the global signal in rsfMRI is not a mere nuisance or a subsidiary component.On the contrary, it is quantitatively the dominant component that accounts for most of the variance in the rsfMRI signal throughout the brain (Fig. 3c) and provides rich information on local hemodynamics that can serve as an individual-level diagnostic biomarker.Moreover, spatiotemporal analyses of the global signal have revealed that it is closely and fundamentally associated with the organization of RSNs, thus challenging the basic assumptions made in conventional rsfMRI analyses and views on RSNs.In the following sections, we introduce new concepts emerging from rsfMRI spatiotemporal analyses focusing on the global signal and discuss how they may contribute to future clinical medicine.

Hemodynamic Response Function and Perfusion
Here, we focus on the variability of the hemodynamic response function as the key to understanding the spatiotemporal characteristics of rsfMRI signals.BOLD fMRI is an indirect measurement of neural activity exploiting the hemodynamic responses, which is necessarily influenced by vascular dynamics.Regional variations in vascular architecture or anatomy that determine the diameter and density of local arterioles and capillaries [47][48][49] directly affect BOLD responses to neural activation.Specifically, BOLD responses are reduced in amplitude and delayed in areas supplied by small and/or collateral arteries often found in steno-occlusive cerebrovascular diseases. 22,50Such fMRI measurements can conversely provide local hemodynamic information that can serve as a noninvasive single patient-level diagnostic biomarker of impaired hemodynamics 22 (Fig. 4).In task-based fMRI, the measurement and interpretation of these metrics are relatively straightforward because we can control the stimulus time series.2][53] ).
5][56][57][58] In this case, although we have no information on the stimulus, we can take advantage of the fact that, as described above, the global mean signal is a ubiquitous and dominant component in the rsfMRI signal time series (Fig. 5a).Therefore, using the global mean signal as a reference, we can map the relative differences in signal amplitude and time delay in a voxel-wise manner (Fig. 5a,b).Interestingly, the signal delay map exhibits a time delay similar to that of perfusion imaging, as if the signal were traveling through the cerebrovascular system (Fig. 5b,c).0][61][62][63][64] The use of global mean signal was later introduced by clinical studies, which made it possible to compute the time delay without using external recording and even without defining a specific region of interest to extract the reference signal (Fig. 6a).These studies showed that the rsfMRI time delay under hypoperfusion or ischemia correlated with that of dynamic susceptibility contrast perfusion imaging. 54,56,57The delay was evident even in MR-defined ischemic penumbrae or stroke cores 54,57 (Fig. 6b), from which normal spontaneous neuronal activity is known to be absent. 65It has also been demonstrated that the rsfMRI signal delay has spatiotemporal characteristics similar to those of hemodynamic responses to acetazolamide 66 and carbon dioxide 67 challenges (Fig. 7).Due to the limited information available about the signal source, it is always more difficult to interpret rsfMRI than task-based fMRI findings.Nevertheless, all the findings described above consistently suggest that the main source of the global mean signal is a nonneural and systemic phenomenon and that its time lag is vascular in origin.

Global Signal and RSNs
RsfMRI connectivity or network analysis depends on the temporal correlation of the signal time series.The more similar two signal time series are, the stronger the correlation between them is.What, then, is the source of synchronization?Does each RSN have its own pacemaker that synchronizes the neural activity within it?Indeed, the neurophysiological mechanism underlying the synchronization within RSNs remains to be elucidated.On the other hand, accumulating evidence points to a shows that the time lag is amplified in the rsfMRI measurement in an area with earlier perfusion (blue voxels), while the slope is close to 1 in an area with delayed perfusion (red voxels).This is because, while the rsfMRI blood oxygen level-dependent signal is less likely to reflect an arterial contribution, the DSC perfusion signal does so.The rsfMRI time lag maps (b and e) show similar patterns across studies.Adapted from Tong et al. 58 54 substantial contribution of the global signal to the generation of structured synchronization within RSNs.Tong et al first reported that the time lag of the low-frequency fMRI signal could give rise to spatial patterns similar to those of RSNs. 68hey showed that a spatial independent component analysis (ICA) of a synthetic time series embedded with the measured time lag of rsfMRI data reproduced RSN-like components 68 (Fig. 8a).The contribution of the time lag to the generation of RSN synchronization was also confirmed in our a study that examined the spatiotemporal characteristics of the rsfMRI signal by decomposing the time series into global and local signals using temporal ICA 69 (Fig. 8b-d).
Using the time lag is advantageous for rsfMRI data analyses, in which it is not easy to know the time course of neural activity or to precisely extract the hemodynamic response functions to neural events in each region.Even in such a case, the time lag can provide information if we can assume a common condition for either the stimulus or the hemodynamic response function.However, it was previously unclear whether the time difference arises from differences in the stimulus function (i.e., delayed neural activation or delayed arrival of a vasodilator, such as carbon dioxide) or in the hemodynamic response function (delayed perfusion).Indeed, due to the lack of information, the source of the time lag has been the subject of debate.To explore its origin, we thus compared the regional variation in time delay maps obtained using rsfMRI and visual task-based fMRI using simultaneous wide-field visual stimulation.The results showed that the time lag patterns were similar regardless of whether the assumed source of the time series was physiological events (global mean signal), spontaneous neural activity (RSN signals), or simultaneous neural stimulation (checkerboard visual task fMRI signals) 70 (Fig. 9).These findings provide direct evidence that the delay is primarily caused by the regional variation in the hemodynamic response function, which causes differences in the time required to increase local blood flow, rather than by the propagation of neural activity.Another interesting phenomenon is that the global mean signal is tightly inversely coupled with the extracranial arterial signal, [70][71][72] with the latter arriving several seconds ahead of the former (Fig. 10).Given that the BOLD response to a stimulus is delayed by several seconds, stimuli seem to affect both the extracranial internal carotid artery and local intracranial pial/capillary arteries almost simultaneously.Taken together, these findings suggest that, as in taskbased fMRI, the delay in the rsfMRI signal likely arises from delayed hemodynamic responses to virtually simultaneous stimuli in the first place.Candidate phenomena include respirationand heartbeat-related changes in blood pressure and blood flow that can simultaneously affect the wide areas of arteries.Although the primary source of the global signal is likely physiological noise, we note that it may also have some neural components.4][75][76] In any case, once the delay is embedded in the signal, venous blood further flows into larger veins and the sinuses, which carry the BOLD signal, just as they carry the contrast agent.
Around the same time, Chen et al showed that the regional variability in hemodynamic responses to physiological events, such as respiration and heartbeat, leads to the generation  67 of "physiological networks" corresponding to those of RSNs. 77ll these findings consistently support the view that global physiological events triggering hemodynamic responses can create local networks similar to RSNs due to the similarity of the hemodynamic responses rather than to the synchronization of neural events within each RSN.Another study investigating rsfMRI spatiotemporal patterns using a complex-valued extension of principal component analysis also confirmed that the spatiotemporal patterns of the rsfMRI signal, including RSNs, can mainly be explained by a small number of global signal components. 78While these findings pose a challenge to the interpretation of rsfMRI studies based on the assumption of neural network synchronization, they further accentuate the importance of the global signal as a source of information related to anatomical and physiological phenomena that has considerable potential as a biomarker.

Global Signal Metrics' Reliability and Sensitivity to Altered Hemodynamics
As we have seen, the rsfMRI time delay is sensitive to local hemodynamic changes and is thus applicable to imaging perfusion or vascular reactivity.More precisely, the local hemodynamic status changes both the magnitude and time delay of fMRI signals, 22,50,79 thus also directly affecting their temporal correlations 54,80,81 (Fig. 11).Such measurements are potentially useful for managing patients with cerebral artery stenoocclusive diseases, evaluating the risk of future ischemic attacks, assessing eligibility for revascularization therapy, and monitoring postoperative hemodynamic changes. 82,83Clinical perfusion imaging mostly uses exogenous contrast materials or tracers, such as iodine or xenon for computed tomography and gadolinium for MRI, or radioactive tracers for PET and SPECT. 84ow do the characteristics of rsfMRI metrics differ from those of conventional measurements?Indeed, there are some differences between the time delay maps of rsfMRI and those of dynamic susceptibility contrast-enhanced perfusion imaging. 55,58Theoretically, fMRI metrics can be more sensitive to altered hemodynamics than CBF measurements.This is because the blood transit time is delayed even when an increase in blood volume results in the successful preservation of normal CBF. 85Moreover, since BOLD contrast originates in the veins, the rsfMRI time delay is amplified compared  69 with the corresponding arterial transit time 58 (Fig. 5).The problem is that the test-retest reliability of fMRI measurements can be affected by measurement errors.To explore the measurement characteristics of rsfMRI global signal metrics, we evaluated their perioperative test-retest reliability and sensitivity to hemodynamic changes in asymptomatic patients with large cerebral artery steno-occlusive diseases undergoing anterior circulation revascularization and compared the results   70 with those of [ 123 I]-iodoamphetamine SPECT CBF.For the rsfMRI metrics, we computed voxel-wise correlation coefficients and time delays using the cerebellar signal as a reference in five-minute scans.In most patients (11/12), a BOLD signal delay of several seconds preoperatively seen in the middle cerebral artery territory decreased postoperatively, with moderate recovery of the temporal correlation.While the reliability measured within the cerebellum was significantly lower for rsfMRI metrics than for SPECT CBF, the sensitivity to postoperative changes was higher for the rsfMRI time delay and equal for the rsfMRI temporal correlation. 86Although the postoperative rsfMRI measurements were affected by susceptibility artifacts caused by the metal plate and air bubbles, longitudinal changes within the anterior circulation area were visible at the single-patient level in both the rsfMRI correlation coefficient and the delay map (Fig. 12).These findings confirm that rsfMRI metrics can be used as biomarkers of subtle hemodynamic changes.The implication for fMRI connectivity studies focusing on neural synchronization is that the temporal correlation can significantly decrease due to altered hemodynamics, even in cases of normal CBF.

Clinical Applications of rsfMRI Global Signal Metrics
][88][89][90][91][92] As described in previous sections, the rsfMRI global signal primarily reflects non-neural contributions that are not lost even when neural activity is impaired.Therefore, it can be used to monitor altered perfusion even in acute stroke 54,57,[89][90][91] (Fig. 13).Since the BOLD signal reflects hemodynamic responses to vasodilatory stimuli, the signal delay indicates a reduction in the cerebrovascular reserve. 66In clinical practice, such information is usually obtained using a potent and potentially harmful vasodilatory agent-namely, acetazolamide-to monitor and identify patients with chronic hypoperfusion who can benefit from revascularization surgery.Although the use of BOLD responses for cerebrovascular reserve assessments is not new, such studies still require vasodilatory agents or tasks to alter the concentration of carbon dioxide in the blood. 66,67he rsfMRI global signal analysis provides similar information by taking advantage of the physiological fluctuations of the  70 BOLD signal.Other clinical applications include the detection of perfusion deficits in Alzheimer's disease that confirmed the rsfMRI signal delay in the bilateral precuneus, which is consistent with previous ASL studies showing hypoperfusion patterns in these areas. 93In a study evaluating the relationships between the rsfMRI time lag, functional connectivity, and   90 cognitive performance in neuropsychiatric systemic lupus erythematosus, the rsfMRI signal was more delayed in the patient group than in healthy control group and the delay in the cingulate cortex was significantly associated with neuropsychological test scores. 94Studies on patients with temporal lobe epilepsy have shown that the rsfMRI signal was delayed in the temporal lobe on the side of involvement. 95,96One study incorporated the rsfMRI time delay into a connectivity analysis, showing that the interhemispheric correlation metrics were significantly lower in a group of patients with unfavorable surgical outcomes. 96A study on the mechanism of hydrocephalus exploited the rsfMRI time delay to assess changes in venous drainage patterns, and found that venous drainage insufficiency contributed to ventricular enlargement independently of brain atrophy. 97Moreover, high-grade gliomas have been shown to have advanced rsfMRI signal time lags and higher ALFFs than low-grade gliomas, presumably due to increased tumor vascularization. 98It has also been shown that the rsfMRI signal time delay is altered in gliomas, although its relationship with tumor grades is more complicated. 99Finally, the global signal's temporal correlation is another possible biomarker, although its clinical application is still limited.As discussed above, it has theoretically and empirically been shown to reflect altered hemodynamics in patients with cerebrovascular diseases. 86herefore, it may be used to assess the status of such diseases, as the global signal time lag does for various disorders.

Conclusions
Thirty years after its inception, BOLD fMRI continues to develop and remains one of the most useful and powerful methods for investigating neurophysiology and pathology.Like many other functional neuroimaging techniques, BOLD fMRI takes advantage of neurovascular coupling. 16This sometimes complicates its use as a tool for evaluating neural activity.At the same time, this challenge highlights the existence of a phenomenon that is not taken into account in our assumptions but is too important to neglect.Low-frequency rsfMRI signals can indicate physiological fluctuations in venous blood deoxyhemoglobin concentrations.One of the main sources of these fluctuations is likely local hemodynamic responses to virtually simultaneous physiological events. 70herefore, the time lag of the signal can indicate local hemodynamic status, 54,[56][57][58]66,72,81,89 which is readily applicable to clinical diagnoses. Moreover, i has been shown that the time lag can lead to spatial patterns corresponding to RSNs, 68,70,77 indicating that RSNs may represent common global (neuro)physiological phenomena rather than local network-specific activity.While these findings pose challenges to the interpretation of rsfMRI studies, they further support the view that rsfMRI can offer detailed information related to global (neuro)physiological phenomena and local hemodynamics that have considerable potential as biomarkers.

FIGURE 2 :
FIGURE 2: Schematic of neurovascular coupling and fMRI analysis.The bottom right part illustrates two pathways regulating local cerebral blood flow (CBF) as the central mechanism of neurovascular coupling.24,25Neural activity leads to an increase in K + in the perivascular space surrounding microvessels to generate local hyperpolarization of the endothelial membrane.Hyperpolarization spreads to adjacent smooth muscle cells and induces smooth muscle relaxation and arteriolar dilation.Rapid arteriole dilation is also induced by nitric oxide (NO) binding.The increase in arteriole diameter leads to an increase in blood flow, thereby increasing the venous signal through a reduction in paramagnetic deoxyhemoglobin.This complex process is modeled using the hemodynamic response function (HRF) h t ð Þ in a task-based blood oxygen level-dependent (BOLD) fMRI analysis using a general linear model.The stimulus function (t) is convolved with HRF to produce a predicted signal time course x t ð Þ, assuming that the BOLD signal is the output of a linear time-invariant system.19

FIGURE 3 :
FIGURE 3: Spatial and temporal characteristics of resting-state fMRI (rsfMRI) signals.(a) Ten representative resting-state networks (RSNs) obtained by performing a spatial independent component analysis of data consisting of 100 runs from 50 healthy subjects.AUD = auditory; CER = cerebellum; DMN = default mode network; EC = executive control; lFP = left frontoparietal; lVIS = lateral visual; mVIS = medial visual; rFP = right frontoparietal; oVIS = occipital visual; SM = sensorimotor.(b) Example magnitude spectrum of a minimally processed rsfMRI signal time series obtained with a repetition time of 0.72 sec.The shade indicates the target frequency range (0.01-0.1 Hz) for rsfMRI analysis, sometimes called the "neural band."(c) Example of RSNs and global mean signal time series from a patient's single run (same data as for [b]).

FIGURE 4 :
FIGURE 4: Impaired hemodynamic responses assessed using motor task blood oxygen level-dependent (BOLD) fMRI.A motor task fMRI data of a 20-year-old patient with moyamoya disease or bilateral middle cerebral artery occlusion.(a) Areas of activation within the primary motor cortices (M1) and cerebellum identified using a hemodynamic response function model-free analysis.(b) Magnitude spectrum obtained by applying a fast Fourier transform to the average signal time series within M1, showing that the data consist of five cycles of signal modulation.(c) Signal time course within M1 and the cerebellum.A prolonged negative response and delayed time-to-peak are seen in the bilateral M1, which reflects delayed hemodynamic responses due to delayed local perfusion.(d) Normalized root-mean-squared error (nRMSE) plot of a temporal shift analysis showing no time shift between the left and right cerebella and a difference of 6 sec between the left and right M1. Adapted from Amemiya et al.22

FIGURE 5 :
FIGURE 5: Global signal temporal correlation and time delay.(a, b) Voxel-wise temporal correlation and time delay measured using the global mean signal as a reference.The high correlation coefficient within the gray matter indicates that the global mean signal is a dominant component that accounts for most of the variance in the rsfMRI signal throughout the brain (a).(c) Map of dynamic susceptibility contrast (DSC) perfusion-weighted imaging time delay computed using the averaged whole brain signal as a reference, as for the rsfMRI time delay.Despite some differences between the maps in (b) and (c), the overall patterns are similar.Adapted from Amemiya et al. 55 (d) A comparison of the time delay obtained from rsfMRI data (e) and DSC perfusion-weighted imaging (f)shows that the time lag is amplified in the rsfMRI measurement in an area with earlier perfusion (blue voxels), while the slope is close to 1 in an area with delayed perfusion (red voxels).This is because, while the rsfMRI blood oxygen level-dependent signal is less likely to reflect an arterial contribution, the DSC perfusion signal does so.The rsfMRI time lag maps (b and e) show similar patterns across studies.Adapted from Tong et al.58

FIGURE 6 :
FIGURE 6: Impaired hemodynamic responses assessed using the resting-state fMRI (rsfMRI) signal time delay.(a) Processing steps for computing rsfMRI signal time delay.After preprocessing, which included realignment, section timing correction, spatial normalization, spatial smoothing, and removal of time series linear trends and band-pass filtering, the temporal shift that gave the best positive fit among the correlation coefficients between each voxel's signal time course and time-shifted (AE20 sec) global signal in the unaffected hemisphere was computed voxel-wise.(b) RsfMRI signal time delay (upper rows) and time-to-peak (TTP) on dynamic susceptibility contrast-enhanced perfusion-weighted images (middle rows) of two patients (a1 and a2) with acute stroke.Diffusion-weighted images (DWI) are also shown (bottom rows).a = advanced; MCA = middle cerebral artery; PCA = posterior cerebellar artery; Rd = right dominant; SCA = superior cerebellar artery.Adapted from Amemiya et al.54

FIGURE 7 :
FIGURE 7: Resting-state fMRI (rsfMRI) signal time lag vs. cerebrovascular reactivity to acetazolamide or a carbon dioxide challenge.(a) The rsfMRI time lag map (right) in a 65-year-old patient with bilateral atherosclerotic lesions shows a pattern similar to that of a cerebrovascular reactivity (CVR) map (middle) measured using single-photon emission computed tomography (SPECT) with an acetazolamide challenge, while the cerebral blood flow (CBF) map (left) shows no remarkable abnormality.Adapted from Nishida et al. 66 (b) An rsfMRI time lag map (left) shows a spatiotemporal patter similar to that of a carbon dioxide challenge.The scatter density plot (right) shows a high correlation between the two measurements.Adapted from Yao et al.67

FIGURE 8 :
FIGURE 8: Resting-state network (RSN) representation of the global signal time delay.(a) Results of a spatial independent component analysis (ICA) of real rsfMRI data (left columns) and a synthetic time series embedded with the measured time lag of the data (right columns).The independent components (ICs) in the two datasets have similar spatial distributions corresponding to the RSNs.Adapted from Tong et al. 68 (b) Processing steps of temporal ICA-based global signal separation employing cross correlation time lag analysis and general linear model (GLM) applied to a set of 50 healthy subjects.The correlation matrices were computed from the total (global + local) rsfMRI signal, which is referred to as functional connectivity (FC) in (c), and the multiple global signal components embedded with the time delay computed by cross-correlation analysis (d).The red squares on the diagonal line indicate synchronization within the RSNs, which is seen not only in (c) but also in (d), indicating that RSN synchronization results from the global signal time lag.Adapted from Amemiya et al.69

FIGURE 10 :
FIGURE 10: RsfMRI internal carotid arteries (ICAs) and global mean signals.(a) The maximum cross-correlation coefficient (MCCC) and (b) time delay (calculated from 90 resting-state scan sessions) between the global signal (GS) and superior sagittal sinus (SSS), left ICA (L-ICA), and right ICA (R-ICA) signals show that the GS negatively correlates with the ICA signals, with a time lag of several seconds.Adapted from Tong et al.71The findings were replicated in our data: the average minimally processed (not band-passfiltered) extracranial ICA signal (c-e) negatively correlated with the global mean signal (GMS) (f).(g) A cross-correlation analysis shows that the rsfMRI GMS lags behind the ICA signals by several seconds.(h) Average cross-correlation data of nine subjects plotted with standard deviations (SD).To facilitate the comparison, the arterial signals are inversed in (f).invRCA = inversed right carotid artery; invLCA = inversed left carotid artery; TR = repetition time.Adapted from Amemiya et al.70

FIGURE 9 :
FIGURE 9: Time lag of visual stimulation and rsfMRI data.The wide-field visual stimulation system activated the whole visual cortex from the primary to the high-order areas (a).(b) Mean time series of two visual tasks with different activation and resting periods averaged within the mask across all scans acquired with a repetition time (TR) of 0.75 sec.(c) Average time lag maps obtained using the mean time series as reference signals for each task-based fMRI dataset.These maps of neural activation-driven blood oxygenation level-dependent response time lag strongly correlated with the rsfMRI global mean signal (GMS) time lag shown in (d).(e) 2D histograms comparing the time lag data of task-based fMRI and rsfMRI in each voxel pooled across subjects.Similar time lag maps were also obtained by a time lag analysis of each RSN component in a dataset of another 50 subjects.(f) Composite map of all RSNs' time lag data.(g) GMS time lag map of the same dataset.Adapted from Amemiya et al.70

FIGURE 11 :
FIGURE 11: Effect of signal time delay on the temporal correlation in rsfMRI analysis.The top row (a-c) illustrates how the temporal correlation (Pearson's correlation coefficient) between two identical signals changes with the signal component's time delay (y-axis) and magnitude decrease (x-axis) in the presence of different levels of noise (x1, x2, and x3).The values are the average correlation coefficients of 100 simulations.The middle row (d-f) shows examples of the two synthesized time series (X1 and X2), while the bottom row (g-i) shows each component of the time series separately (Signal, N1, and N2).The green and cyan asterisks in the top row indicate the correlation coefficients of 0.85 and 0.72 obtained with 2.2-and 3.4-sec delays, respectively, or 63% and 79% decreases in signal magnitude (a), respectively.The same effect was obtained by doubling (b) and tripling (c) the noise of the two time series.Adapted from Amemiya et al.70

FIGURE 12 :
FIGURE 12: Preoperative and postoperative global signal metrics and single-photon emission computed tomography (SPECT) cerebral blood flow (CBF).Preoperative and postoperative data of a 67-year-old male patient with atherosclerotic right middle cerebral artery occlusion undergoing superficial temporal artery-middle cerebral artery (MCA) bypass surgery.The postoperative data were acquired 5 days after the surgery.A clear increase in the fMRI temporal correlation (Rc) and a decrease in the time delay (TDc) were observed within the anterior MCA territory (the arrows indicate areas with a reduced Rc, increased TDc, and reduced SPECT CBF preoperatively).The postoperative changes on SPECT were unclear.The data are shown in the neurological view.Adapted from Amemiya et al.86

FIGURE 13 :
FIGURE 13: Longitudinal hemodynamic changes in acute ischemic stroke assessed based on the global signal delay.(a) Data of a patient with right middle cerebral artery infarction.Areas with delayed perfusion are similarly shown on blood oxygenation leveldependent (BOLD) signal time lag and dynamic susceptibility contrast perfusion time-to-maximum (Tmax) maps 2 hours after the onset of stroke symptoms (left column), which persisted after failed intravenous thrombolysis (right column).(b) Successful recanalization of left posterior cerebral artery infarction (left column, 2.5 hours after symptom onset; right post thrombolysis).The areas with delayed perfusion were minimized after recanalization (left column).Adapted from Khalil et al.90