To examine spatiotemporal dynamics of low frequency fluctuations in rat cortex.
To examine spatiotemporal dynamics of low frequency fluctuations in rat cortex.
Gradient-echo echo-planar imaging images were acquired from anesthetized rats (repetition time = 100 ms). Power spectral analysis was performed to detect different frequency peaks. Functional connectivity maps were obtained for the frequency peaks of interest. The images in the filtered time-series were displayed as a movie to study spatiotemporal patterns in the data for frequency bands of interest.
High temporal and spectral resolution allowed separation of primary components of physiological noise and visualization of spectral details. Two low frequency peaks with distinct characteristics were observed. Selective visualization of the second low frequency peak revealed waves of activity that typically began in the secondary somatosensory cortex and propagated to the primary motor cortex.
To date, analysis of these fluctuations has focused on the detection of functional networks assuming steady state conditions. These results suggest that detailed examination of the spatiotemporal dynamics of the low frequency fluctuations may provide more insight into brain function, and add a new perspective to the analysis of resting state fMRI data. J. Magn. Reson. Imaging 2009;30:384–393. © 2009 Wiley-Liss, Inc.
LOW FREQUENCY FLUCTUATIONS (LFFs) in the blood oxygenation level dependent (BOLD) functional MRI (fMRI) signal from areas known to be strongly connected anatomically are correlated even in the absence of a task or stimulus (1). These spontaneous fluctuations appear to originate from the same source as the task-related BOLD response, leading to the hypothesis that correlated fluctuations reflect coordinated variations in neural activity (2). The possibility that correlated LFFs reflect the spontaneous “functional connectivity” of the brain has excited the interest of the fMRI and neuroscience communities, and the number of studies using the technique is growing rapidly.
Most research performed on LFFs falls into three categories. The first involves mapping functional networks in the brain. Networks associated with different systems including visual, motor, auditory, memory, and language have been detected in resting state fMRI studies (3, 4). In addition, researchers have identified two widely present networks, one containing areas that are typically deactivated during a task (the “default mode” network) and another containing areas that are active during a wide variety of tasks and that may be related to attention (5).
Studies in the second category focus on the effect of different interventions or pathological conditions on functional connectivity and LFFs. Anesthesia, hypercapnia, and cocaine administration have been reported to disrupt functional connectivity (6–8). Also, changes in functional connectivity have been observed in conditions including autism, Alzheimer's disease, schizophrenia, and blindness (9–12).
The third group of studies addresses the origin of the LFFs. Physiological processes such as respiration can introduce spatially localized correlations in functional MRI data, and several groups have examined the effects of physiological noise on functional connectivity (13–16). Networks of correlated areas persisted despite significant attenuation of physiological noise, indicating that the noise is not the primary source of the correlation (15, 17). Recent work by several groups has provided additional support for the neural basis of functional connectivity. A strong reduction of functional connectivity between hemispheres has been reported after surgical severance of the corpus callosum, the primary interhemispheric fiber tract. Intrahemispheric connectivity, however, was retained (18). It has also been shown that functional connectivity and the relationship between correlated networks is related to behavioral variability and the ability to learn new muscle synergies, more evidence that the correlated signal fluctuations are closely related to functional changes in the brain (19, 20). Additional support comes from combined electroencephalography (EEG) -fMRI studies, although the findings are not consistent between the studies (21–24), possibly due to differences in the areas examined or recording methods used.
To better understand the relationship between neural activity and functional connectivity, several groups have recently turned to animal models (25–30). The use of animal models allows for more invasive experimental designs, which can give deeper insight into the LFFs. These studies have provided additional evidence of a link between coordinated neural activity and functional connectivity. Anesthesia-dependent changes in delta power correlation were shown to co-vary with changes in functional connectivity in rats, although the measurements were not performed simultaneously due to technical limitations (30). Shmuel and Leopold reported correlations between spontaneous neural activity and BOLD fluctuations in the visual cortex of monkeys in the absence of any stimulus (27). These results are an extension of previous work showing very slow (< 0.1 Hz) coherent oscillations in the band limited power of local field potentials obtained from monkey visual cortex (31).
Current functional connectivity analysis techniques such as cross-correlation of time-courses (1) or independent component analysis (ICA) (32) provide only steady-state information about networks where the relationship between the areas involved is assumed to be maintained over the entire length of the scan. These analysis methods do not give any information about the spatiotemporal dynamics of the LFFs. From EEG and magnetoencephalography (MEG) studies, it is known that behaviorally relevant changes in network activity occur on much shorter time scales than the minutes required for the acquisition of functional connectivity data. If the LFFs used to map functional connectivity have a neural basis, it should be possible to identify dynamic changes in network activity using MRI. In support of this idea, the envelope of EEG is shown to predict fluctuations in cerebral blood flow measured using laser Doppler flowmetry in rats (33). The spatiotemporal dynamics of the LFFs might, therefore, enhance our interpretation of functional connectivity data and provide us with clues to the origin of the low frequency fluctuations.
In this study, we report a preliminary investigation into the spatiotemporal dynamics of LFFs in the rodent model. The use of an animal model is conducive to future experiments combining functional connectivity MRI and electrophysiology. In addition, acquisition with very short repetition time (TR) and high spatial resolution is possible with the high-field animal scanners, which is necessary for obtaining spatiotemporal information from the data while avoiding aliasing. This will also allow us to address questions that remain unanswered for animal studies, including the relative contribution of different frequency components and the extent of physiological noise contamination. In previous animal studies, acquisition parameters did not allow the visualization of fine spectral details needed to address this issue. In addition, most animal studies were conducted using a fairly long repetition time and no physiological correction was performed. The contributions of respiration and the cardiac cycle have not been examined in the rodent model, and could be more severe than in humans due to the widespread use of high-field scanners (34).
In this study, we describe the acquisition of functional connectivity data from the rat with high temporal resolution that allows separation of the primary cardiac and respiratory components from the frequencies of interest. High spectral resolution allows the identification of two low frequency peaks in the data acquired from the rat cortex. The peaks exhibit different functional connectivity specificity. We further characterize spatiotemporal characteristics of both peaks and identify patterns of LFF propagation, which may not be deduced from typical functional connectivity analysis methods.
All experiments were performed in compliance with guidelines set by the National Institutes of Neurological Disorders and Stroke ACUC. Six adult male Sprague-Dawley rats (168–234 g) were initially anesthetized with 5% halothane and maintained at 1.5% halothane during the following surgical procedures. Each rat was orally intubated and placed on a mechanical ventilator throughout the surgery and the experiment. Plastic catheters were inserted into the right femoral artery and vein to allow monitoring of arterial blood gases and administration of drugs. Two needle electrodes were inserted just under the skin of each forepaw, one between digits 1 and 2, and the other between digits 3 and 4. After surgery, the rat was given an i.v. bolus of α-chloralose (80 mg/kg) and halothane was discontinued. Anesthesia was maintained with a constant α-chloralose infusion (27 mg/kg/h) (35, 36).
The rat was placed on a heated water pad to maintain rectal temperature at ∼37°C while in the magnet. Each animal was secured in a head holder with ear bars and a bite bar to prevent head motion and was strapped to a plastic cradle. End-tidal CO2, rectal temperature, tidal pressure of ventilation, heart rate, and arterial blood pressure were continuously monitored during the experiment. Arterial blood gas levels were checked periodically and corrections were made by adjusting respiratory volume or administering sodium bicarbonate to maintain normal levels when required. An i.v. injection of pancuronium bromide (4 mg/kg) was given once per hour to prevent motion.
One of the rats was euthanized at the conclusion of the experiment without being removed from the scanner and an additional series of BOLD-weighted images was acquired to serve as a control dataset.
All images were acquired with an 11.7 Tesla (T) / 31-cm horizontal bore magnet (Magnex, Abingdon, UK), interfaced to an AVANCE console (Bruker, Billerica, MA) and equipped with a 9-cm gradient set, capable of providing 30 G/cm with a rise time of 65 μs. Shimming was performed with a custom-built shim set and high power shim supply (Resonance Research, Billerica, MA). A contoured rectangular surface coil (2 × 3 cm) that attached to the head holder was used to transmit and receive the MR signal. Scout images were acquired in three planes with a fast spin-echo sequence to determine appropriate positioning for the functional study. A spin-echo, echo-planar imaging (EPI) sequence was used to acquire a series of images during forepaw stimulation (2 mA current, 300 μs pulses repeated at 3 Hz) to locate the slice containing primary somatosensory cortex (SI). Setup included shimming, adjustments to echo spacing and symmetry, and B0 compensation. A single-shot sequence with a 64 × 64 matrix was run with the following parameters: effective echo time 30 ms, repetition time 1.0–1.5 sec, bandwidth 200 kHz, field of view 1.92 × 1.92 cm. Whole-brain coverage was obtained with 10–11 2-mm thick slices, spaced 0.2 mm apart. The paradigm consisted of 60 images during rest, followed by 30 images during forepaw stimulation, and another 60 images during rest. Functional connectivity data were acquired with gradient echo EPI on the slice containing SI with following parameters: repetition time 100 ms, echo time 20 ms (15 ms for one rat), field of view 1.92 × 1.92 cm, 2-mm-thick slice, 3600 repetitions. No stimulation was given during these scans.
The area comprising the brain was segmented using an intensity threshold and manual removal of remaining voxels outside the brain. The datasets were spatially blurred using a 3 × 3 Gaussian filter with σ = 2 pixels. All analysis was performed using Matlab (MathWorks, Natick, MA), unless otherwise noted. The first 500 time points were discarded before power spectral analysis and after all the filtering operations described later to discard transient effects of data acquisition and filtering. Resultant time-courses were de-meaned, and quadratic detrending was performed.
Power spectra were obtained for time-courses from SI and secondary somatosensory cortex (SII) using the Welch method (eight sections with 50% overlap, Hamming window). Visual inspection suggested the presence of two low frequency peaks, LF1 (f < 0.05 Hz) and LF2 (0.11 Hz < f < 0.18 Hz), in three of the datasets obtained from live rats (described later in results). Only the lower peak (LF1) was visually detectable for the other three datasets from live rats. Peaks due to respiration and cardiac noise were also observed. Frequencies of respiratory and cardiac contributions were assessed by inspecting the spectra. No specific peaks were observed for dead rats.
Maps were created to assess the spatial localization of physiological noise and the low frequency peaks. We visually examined the spectra of time-courses obtained from different locations in the brain for LF1, LF2 (for the datasets showing LF2), cardiac, and respiratory peaks, and noted the frequencies of the corresponding peaks for each dataset (Table 1). Band-pass or low-pass filtering (third-order Butterworth filter) was used to isolate the time-courses corresponding to the individual peaks. The filtered time-courses were normalized to obtain percentage difference from mean signal intensity of the raw data. Maps showing the standard deviation of each contribution were obtained, reflecting spatial localization of contribution due to different peaks. Parameters used for one of the live rats (rat 1) were used for processing the data from the dead rat because no visible peaks were present. Frequency cutoffs of 0–0.05 Hz and 0.08–0.2 Hz were used for LF1 and LF2, respectively. Filters for physiological noise had their pass bands centered at respective frequencies, with a 3 db width of 0.2 Hz.
Six functional connectivity maps were obtained from each dataset using two different filters (0–0.05 Hz and 0.08–0.2 Hz) and three different seed locations (SI, SII and caudate-putamen [CP]). The seeds' time-courses were obtained by averaging filtered time-courses from a 3 × 3 manually chosen region in the areas of interest. Cross-correlation between the seed time-course and time-courses from all the voxels in the brain was calculated to obtain correlation maps. To estimate localization of connectivity, the number of voxels passing an arbitrary cross-correlation threshold of 0.5 was calculated for bilateral regions of interest (ROIs) covering SI, SII, and CP for each seed location. In addition, average time-courses were obtained for the ROIs defined on left and right SI, SII, and CP and 6×6 correlation matrix was obtained.
Image-by-image visualization was used as a primary tool to study spatiotemporal dynamics. The preprocessed data were filtered and resulting time-courses were normalized individually to unit variance. The resulting data were displayed as a movie for visual detection of any spatiotemporal patterns. Using different filters, we were able to observe dynamics of LF1, LF2, and physiological contributions. The filters specifications were the same as described above.
Power spectral analysis of time-series obtained from cortical areas including SI and SII revealed two distinct low-frequency peaks in the following frequency ranges for three rats: below 0.05 Hz (LF1), and between 0.11 and 0.18 Hz (LF2) (Fig 1a). Only a single peak was observed in the remaining three datasets, with most of the power in the LF1 range (Fig. 1b). Peaks corresponding to respiratory and cardiac noise were also observed (Fig. 1d). The frequency of the respiratory peak was ∼1 Hz as expected, whereas the cardiac peak was detected at ∼4.5 Hz. No obvious peaks were observed in the data obtained from the dead rat (Fig. 1c).
Figure 1a (columns 3 and 4) shows standard deviation maps for low frequency contributions for one of the datasets showing the two low frequency peaks. The LF2 appears with high magnitude in the cortex with high specificity. LF1 contribution is also high in the cortex compared with the subcortex. Figure 1b (columns 3 and 4) shows the maps for a dataset without a clear LF2 peak. The LF1 maps are similar to those seen in rats with two peaks. In contrast, the LF2 no longer appears in the cortex with high magnitude compared with the subcortex. Peak distribution for the dead rat does not show any specific patterns (Fig. 1c). The strongest respiratory contribution was observed primarily near the sagittal sinus, large blood vessels, and ventricles. The contribution from the primary cardiac peak was high near large veins for most rats and occasionally also near the sagittal sinus. Examples from one rat are shown in Figure 1d. Edges and areas with low coil sensitivity appear with high magnitude in many of the standard deviation images, which may be expected because percentage difference due to noise factors will be large in those areas due to low baseline intensity. Edges appear with reduced intensity due to partial volume effects.
The low pass (0–0.05 Hz) filter retained the LF1 peak, whereas the band pass (0.08–0.2 Hz) filter retained LF2. Figure 2 shows connectivity maps for a dataset with two peaks and a dataset with one peak, each with two different seed locations (SI and SII). The correlation maps for LF1 exhibit low sensitivity to the location of cortical seed regions and high correlation (> 0.5) is seen for the whole cortex and some subcortical areas. We observed this trend for LF1 for most seeds within cortex in all the datasets. In the datasets with two peaks, correlation maps based on LF2 exhibit higher specificity than the maps based on LF1. High correlation values are found primarily in the contralateral analogue of the seed region for a seeds placed in SI, rather than evenly distributed throughout the cortex. Some bilateral correlation for LF2 is also observed in the datasets with one peak (Figs. 2b, 3a). The average correlation value in the contralateral SI is greater than 0.4 for seeds placed in SI for maps created based on both LF1 and LF2 in datasets with both peaks (Fig. 3a, rats 1–3). As expected, in rats with only the lower peak, correlation is reduced for LF2 (Fig. 3a, rats 4–6).
Figure 3b shows the average cross-correlation coefficients between all ROIs for LF1 and LF2. LF2 generally exhibits lower cross-correlation values, and strong correlation is mostly confined to bilateral cortical areas.
To better quantify the specificity of each peak, the number of voxels with cross-correlation values of greater than 0.5 were measured in anatomically drawn ROIs in SI, SII, and CP. Figure 3c,d shows that LF1 maps are relatively independent of cortical seed location, as compared with LF2. The average number of voxels with cross-correlation greater than 0.5 in different regions shows much smaller variability for LF1 compared with LF2 for seeds placed in SI and SII. The maximum difference between voxels crossing the threshold (0.5) for seeds placed in SI and SII is less than 2 voxels for LF1 and approximately 29 voxels for LF2. The high magnitude of error bars seen in Figure 3c–e is due to variability in the number of voxels in different areas across rats because of variation in slice orientation. Correlation values observed for LF1 were in general higher in comparison with those obtained for LF2 (Fig. 3a,b).
As seen in Figure 3c,d, very few voxels in CP (less than 1, on average) have the LF2 cross-correlation values above 0.5 for seeds placed in SI and SII. Also, no voxels in SI and SII showed LF2 correlation coefficient above 0.5 for seeds placed in CP. In contrast, correlation coefficients for both cortical and subcortical pixels crossed this threshold for the same seeds for LF1. These observations suggest that LF2 correlation maps are more specific to the networks in which the seed is placed, in agreement with previous work that identified separate, well-delineated networks in SI and CP (28).
Image by image visualization of LF2 signal (from the rats exhibiting two clear peaks) revealed consistent, well-organized spatiotemporal patterns. The most prominent pattern, which was observed in all the datasets, is displayed in Figure 4a,b. This pattern looks like a propagating wave, with high signal intensities starting from SII and traveling along the cortex. Intensity fluctuations were approximately 2–4%. The waves were bilateral in most cases, but unilateral propagation was occasionally observed. The travel time of the waves from SII to MI was consistent across rats (mean = 4.7 s, standard deviation = 0.4, 0.9, and 0.8 respectively). The waves were not uniformly distributed in time but tended to occur in clusters (Fig. 5). No organized patterns were seen when 0.08–0.2 Hz component was visualized in the same way for the datasets with one peak only. The most prominent pattern for the LF1 signal, in contrast, was a slow change in intensity throughout the whole brain, moving from the surface to the center (Fig. 4c). Movies that more clearly show these spatiotemporal dynamics for two different rats are available online as supplementary material (Supp. Movie_1.avi and Movie_2.avi).
Application of the same method to the respiratory and cardiac signals revealed different patterns, primarily involving large vessels and areas near the ventricles (Fig. 6a,b). No organized patterns were detectable in the datasets obtained from the dead rat (Fig. 6c). Movies showing dynamics of respiratory and cardiac noise are available online as supplementary material (Supp. Movie_3.avi and Movie_4.avi).
Functional connectivity studies in the well-characterized rat model provide valuable insights into the acquisition and interpretation of human studies. In this experiment, both the spectral resolution and the sampling rate were higher than in typical human studies, providing clear separation of the low frequency peaks and preventing aliasing of first harmonics of the cardiac or respiratory cycles. The implications of this study are twofold: (i) It identifies spatiotemporal patterns in the resting state BOLD data which are not revealed by the conventional analysis (ii) It demonstrates the presence of multiple peaks in the low frequency band, showing different spatiotemporal characteristics and functional connectivity.
Our findings suggest that the two low-frequency peaks contain different information and may have different physiological origins. Both connectivity maps and propagation patterns are dependent upon the frequency range chosen for analysis. These results motivate in-depth investigation of frequency-band selection for functional connectivity studies. In an interesting parallel to these results, Obrig et al. reported the presence of two low-frequency peaks (∼0.04 Hz and ∼0.1 Hz) in blood oxygenation level measurements made using near infrared spectroscopy in human subjects (37), suggesting that the phenomenon described in this article may not be limited to anesthetized rodents.
It is interesting to note that LF2 is not observed as a well-separated peak in three of the datasets. However, some LF2 connectivity is also observed in the datasets with only one peak, suggesting presence of some LF2 signal (Figs. 2, 3). This weak presence might be reflected as broadening of the LF1 peak (Fig. 1b). The short repetition time used in this study to avoid aliasing of the primary cardiac component results in a low signal-to-noise ratio, which may limit our ability to detect and separate the two peaks. This study was also limited by the small number of animals (n=6), imaged on a different MRI system than the one currently in use. Because LF2 and propagating waves were observed in only half of the datasets, we retrospectively inspected the data acquired for a different experiment (four rats) for presence of LF2 and waves to assess the reproducibility of our observations. The rats were anesthetized and prepared in the same way as described for this experiment, but the experimental paradigm and acquisition parameters were different. The number of repetitions available (taken from the initial resting periods before a period of forepaw stimulation) was 1200, rather than 3600, reducing the spectral resolution. Also, shorter TEs (∼15 ms) were used, reducing the sensitivity to the BOLD effect. Despite these differences, we detected LF2 and the associated spatiotemporal patterns in three of four of these datasets. A movie showing examples of waves for those three datasets is available online as supplementary material (Supp. Movie_5.avi).
Low-frequency fluctuations in neural activity, cerebral blood flow, and blood oxygenation levels have been observed in both humans and animals (31, 37, 38). However, the links between these phenomena have barely been explored, and the inter-relations may be complex, particularly in anesthetized animals. In addition to changes in blood oxygenation level and cerebral blood volume, the signal intensity of the EPI images acquired in this study is also highly sensitive to changes in cerebral blood flow, due to the short repetition time. It is possible to envision, then, a situation in which signal fluctuations in one frequency range (e.g., LF1) reflect widely coherent oscillations in blood flow (possibly due to vasomotion), while signal fluctuations in another range (LF2) reflect changes in the oxygenation level of blood, possibly linked with neural activity, which are more localized to highly connected cortical areas. This hypothesis is purely speculative because the relationship between neural activity and cerebral blood flow is highly dependent upon anesthesia, and no studies measuring electrical activity and blood flow have been performed in the α-chloralose anesthetized rat to our knowledge.
If the peaks can be attributed to different aspects of the brain's function, it may be that the relative magnitude of the two peaks is highly sensitive to the level of anesthesia, accounting for the variability seen in this study. Previous work has shown that functional activation increases with time after anesthesia is changed from halothane to α-chloralose, and, therefore, the extent to which halothane washes out before imaging might be a factor that determines the relative contribution of LF1 and LF2 (39).
Propagation patterns for LF1 seem unlikely to reflect coordinated neural activity between areas of the brain because of the lack of functional specificity. The primary propagation pattern for LF1 moves inward from the surface of the brain, and high signal appears across the entire cortex at once.
The propagation patterns for LF2 are potentially more interesting. SI, SII, and MI are all strongly connected anatomically, and the bilateral onset of the waves is suggestive of coordinated activity. Previous studies have reported fluctuations in cerebral blood flow at approximately 0.1 Hz, attributed to vasomotion (38). The waves of cerebral blood flow occurred simultaneously in both hemispheres, and were relatively high in amplitude (∼20%) (38). Low-frequency oscillations in neural activity have also been observed (31) and the close coupling of cerebral blood flow and neural activity during stimulation suggests that they are related. Thus, the propagating waves observed in this study may reflect the spatiotemporal dynamics of low-frequency oscillations in neural activity. This idea is supported by recent work from Shmuel and Leopold (27), which showed that the BOLD signal from the visual cortex in monkeys was correlated with band limited power of gamma band using simultaneously recorded fMRI and electrophysiological signal. The correlation was strongest at a lag of approximately 6 s, and the pattern of correlation propagated away from the electrode with a time scale of seconds. While the animal preparation used in these studies was very different, the similarities between the propagation of correlated BOLD responses and the spontaneous waves observed here are striking.
Other possible explanations for the waves observed here are less convincing. The presence of propagating waves raises the specter of spreading depression, that is, spreading depolarization across the cortex (40). However, the waves observed in this study occur at a much shorter time scale in comparison, and all animals were physiologically stable.
The spatiotemporal patterns we observed are also not likely to be due to the aliased components of the physiological contributions. First, the high sampling rate resulted in alias-free sampling of the primary components of both cardiac and respiratory rhythms. Second, the power maps for these peaks do not overlap completely with the low frequency peaks of interest. Although there is some overlap between the power maps of low frequency signals and respiratory contribution, those maps do not look identical and there are cortical areas where low frequency peaks can be seen in absence of any significant physiological peaks. Neither can these patterns be attributed to scanner noise, because they were not observed in the data obtained from the dead rat. Also, it is unlikely that the scanner noise should be restricted to the cortical region only.
It is interesting to note that physiological noise does not appear to have a major impact on functional connectivity studies in rats. The connectivity maps for SI obtained in this study were qualitatively similar to those obtained by other groups using longer repetition times (28–30). However, localized contributions from respiratory and cardiac noise were observed in our data and may be a confounding factor in functional connectivity studies of other brain regions, particularly those near large vessels or the ventricles.
This study uses single slice data because the short TR did not allow the acquisition of more images. Consequently, our data cannot reveal the full spatiotemporal propagation of the traveling waves. Other areas of the brain may also be involved. The primary direction of travel of the LF2 waves may not be SII to MI, as it appears in this study, because single slice data can capture only a projection of the wave motion. Possible ways to address this issue in the future include the acquisition of single slice data with different slice positions and orientations, or multislice experiments using a longer TR. However, the TR must remain short enough to provide adequate temporal sampling for visualization of the waves, so image acquisition is likely to be limited to a small number of slices.
The presence of the propagating waves has implications in regard to interpretation of functional connectivity maps. Although functional connectivity between SI and SII is reduced due to time lag, the propagation of the signal from SI to SII (and MI) suggests some kind of communication or “functional connection.” Observation of such events in the resting state BOLD data is expected to have a significant impact on how we interpret functional connectivity.
The detection of propagating waves of MRI signal fluctuations that may reflect slow changes in electrical activity naturally leads to speculation about whether other dynamic neural events can be detected with MRI. While the task-related BOLD response has been shown to reflect changes in neural activity, particularly in local field potentials, previous attempts to link functional connectivity measurements with electrophysiology have proven hard to interpret. Even recent work in animals has pointed to potentially different sources for BOLD signal correlation. Lu et al presented indirect evidence linking LFFs with delta band power in somatosensory cortex of anesthetized rodents, while Shmuel and Leopold recently showed that the MRI signal reflected changes in the gamma band power of local field potentials using implanted electrodes in the visual cortex of anesthetized monkeys (27, 30). Studies have also been performed in human subjects to investigate neural correlate for LFFs and functional connectivity. Goldman et al reported negative correlation between BOLD and alpha band power in multiple cortical regions of awake humans using simultaneous acquisition of EEG and fMRI (23). Another simultaneous EEG-fMRI study by Mantini et al (22) suggests that multiple frequency bands are related to the LFFs, and that the frequency spectra are different for different functional networks. A recently published human study by He et al uses electrocorticography and fMRI in the patients with intractable epilepsy and indirectly suggests that both slow cortical potentials (< 4 Hz, overlapping with delta band) and gamma band power are related to LFFs in BOLD in wakefulness and rapid-eye-movement sleep (21). Only slow cortical potentials showed a correlation pattern similar to that of LFFs in other states of sleep. In another study, the results varied within scans from the same subjects, suggesting that the relationship between LFFs and the EEG signal may vary depending on the current state of the subject (24). In general, the area of the brain that is studied, the relative sensitivity of the electrical recording techniques, and the state of the subject (awake, asleep, or anesthetized) might explain some of the differences in the findings. These studies highlight the need for further research addressing the issue of the neural origin of the LFFs. Animal studies may provide the missing link. The ability of MRI to distinguish such time-varying features such as periods of elevated cortical bursting will be limited by the temporal resolution of the imaging technique and the vascular response, the relative signal to noise of the signal of interest, the spatial extent of the activity, and the amount of time for which the activity persists.
In conclusion, we have presented a novel approach for analyzing and visualizing resting state fMRI data which provides new insight into the resting state BOLD fluctuations. Ongoing work focuses on spatiotemporal dynamics of low-frequency fluctuations in human data. Current challenges include the compromise between spatial coverage, SNR, and sampling rate. The properties of spatiotemporal events observed in the data, when combined with multimodality data and information about different factors contributing to the BOLD signal, might provide clues to the origin of the LFFs. In addition, if and when neural origin of LFFs is confirmed, such techniques would provide us with a convenient and noninvasive method of probing the transient aspects of resting brain under normal and pathological conditions.
The authors thank the Laboratory for Functional and Molecular Imaging at the National Institutes of Health for use of the 11.7T animal scanner, with special thanks to Dr. Artem Goloshevsky for his assistance with data acquisition.