On the Detection of High Frequency Correlations in Resting State fMRI

Abstract

Current studies of resting-state connectivity rely on coherent signal fluctuations at frequencies below 0.1 Hz, however, recent studies using high-speed fMRI have shown that fluctuations above 0.5 Hz may exist. This study replicates the feasibility of measuring high frequency (HF) correlations in six healthy controls and a patient with a brain tumor while analyzing non-physiological signal sources via simulation. Resting-state data were acquired using a high-speed multi-slab echo-volumar imaging pulse sequence with 136 ms temporal resolution. Bandpass frequency filtering in combination with sliding window seed-based connectivity analysis using running mean of the correlation maps was employed to map HF correlations up to 3.7 Hz. Computer simulations of Rician noise and the underlying point spread function were analyzed to estimate baseline spatial autocorrelation levels in four major networks (auditory, sensorimotor, visual, and default-mode). Using seed regions based on Brodmann areas, the auditory and default-mode networks were observed to have significant frequency band dependent HF correlations above baseline spatial autocorrelation levels. Correlations in the sensorimotor network were at trend level. The auditory network was still observed using a unilateral single voxel seed. In the patient, HF auditory correlations showed a spatial displacement near the tumor consistent with the displacement seen at low frequencies. In conclusion, our data suggest that HF connectivity in the human brain may be observable with high-speed fMRI, however, the detection sensitivity may depend on the network observed, data acquisition technique, and analysis method.

1. INTRODUCTION

Resting-state fluctuations have emerged as an adjunct to task-based fMRI, allowing for simultaneous mapping of functional connectivity across dozens of resting-state networks (RSNs), (Biswal et al., 1995; Fox et al., 2005; De Luca et al., 2006; Raichle et al., 2007; Abou-Elseoud et al., 2010; Schopf et al., 2010; Allen et al., 2011; Li et al., 2011; Lowe, 2012). These network fluctuations are dominated by low frequency components (< 0.1 Hz) arising from the slowly varying vascular dynamics underlying the blood oxygenation level dependent (BOLD) signal. However, the slow time scales, considerable spatio-temporal non-stationarity (Chang and Glover, 2010; Feinberg et al., 2010; Smith et al., 2012; Liu and Duyn, 2013; Allen et al., 2014), and physiological noise present challenges for reliable measurements and segregation of RSNs in single subjects. Advances in high-speed fMRI (Rabrait et al., 2008; Witzel et al., 2008; Feinberg et al., 2010; Lin et al., 2010; Moeller et al., 2010; Posse et al., 2012; Setsompop et al., 2012), which enable unaliased sampling of physiological signal fluctuations, have increased sensitivity for mapping functional connectivity and detecting dynamic changes (Feinberg et al., 2010; Smith et al., 2012; Posse et al., 2012) compared with conventional echo-planar imaging (EPI) techniques.

Recently, several studies using high-speed fMRI techniques have reported different potential resting-state networks (RSNs) at high frequencies (up to 5 Hz) (Chu et al., 2013; Lee et al., 2013; Kalcher et al., 2014; Lee et al., 2014; Chen and Glover, 2015; Gohel and Biswal, 2015), suggesting that functional integration between brain regions at rest occurs over broader frequency bands than previously thought (Lee et al., 2014; Gohel and Biswal, 2015). Due to the limited signal to noise ratio (SNR) inherent to high frequency fluctuations, these studies used extended regions of interest (ROIs) for seed-based connectivity analysis (SCA). (Chen and Glover, 2015) have shown that resting state functional connectivity extends up to 0.5 Hz and that the BOLD contrast of these fluctuations decreases with increasing frequency, but cautioned that the origin of these signal sources remains unclear. A recent study using multi-echo simultaneous multi-slice (multi-band) EPI was unable to detect high frequency BOLD resting state networks (Olafsson et al., 2015). The authors conclude that this observation casts doubt on findings related to resting state BOLD networks at frequencies higher than 0.2 Hz and that emerging reports of high frequency resting BOLD networks may require closer examination. These conflicting results and differing networks observed may in part be due to image degradation at high acceleration factors, e.g. due to increasing slice cross-talk and worsening g-factor (Moeller et al., 2010), that differ between the different high-speed fMRI data acquisition approaches, and due to differences in data and model driven analyses in the above studies. As the underlying mechanism of current observations remains unknown, further investigation into the characteristics of high frequency fluctuations and characterization of possible sources is required.

Characterizing the spatial and temporal correlations of an observed signal, whether physiological or acquisition method related is an important step for correlation analyses, regardless of the frequency band being investigated. The point spread function (PSF), which depends on the data acquisition method, can create spatial correlations that will differ for 2D and 3D acquisition techniques. Additional spatial autocorrelations arise when using an extended seed region and will provide baseline correlations across said region, regardless of actual signal content. This effect is confounded by the intrinsic PSF, but has an independent origin. Correction for temporal autocorrelations (Worsley and Friston, 1995; Kruggel et al., 2002) have been applied in previous studies in order to threshold correlations at fast acquisition rates (Lee et al., 2013).

Confounding signals of non-physiological and physiological origin must also be accounted for in fMRI studies. Scanner artifacts, e.g. due to RF or gradient instabilities, can induce signal correlations within a data set that are typically at discrete frequencies. Physiological confounds arise from a variety of sources. Head motion within the scanner can cause movement artifacts, obscuring networks as well as creating false-positive connections (Satterthwaite, et al., 2012; Van Dijk et al., 2012) despite state-of-the-art motion correction in post-processing. Motion effects are particularly pronounced at the edges of image slices, creating high sensitivity to even subtle movements. Cardiac pulsations induce correlations at multiple frequencies (fundamental and higher harmonics) that are coherent across the brain, inducing false-positive correlations unrelated to the neuronal activity being investigated. Additionally, respiration induces signal pulsations around 0.3 to 0.5 Hz, which can induce false-positive correlations in a similar fashion. Frequency filtering of the data itself can also change noise correlations, particularly in the case of narrow pass band filters that are associated with increased filter ringing.

The optimal approach to account for these confounding signal sources is still a topic of debate. Previous studies of high frequency connectivity have used regression of physiological signal fluctuations (Chu et al., 2013; Lee et al., 2013; Lee et al., 2014; Chen and Glover, 2015; Gohel and Biswal, 2015), frequency band filtering combined with ICA (Lee et al., 2014; Olafsson et al., 2015), or frequency band filtering without removing physiological signal pulsation (Kalcher et al., 2014). Although the present study does not utilize regression of any kind, regression via a general linear model (GLM) has shown to be effective for removing non-neuronal sources of correlation, such as low frequency drifts, aliased physiological noise (cardiac), and residual movement in low frequency resting state connectivity analysis using conventional EPI (Monti, 2011). However, the effectiveness of physiological noise regression in short TR data has not yet been compared with frequency filtering and may introduce false-positive correlations (Chen et al., 2016). Challenges include the substantially lower temporal SNR than EPI and the presence of regional differences in the cardiac waveform (Posse et al., 2013). Additionally, global signal regression is often used to increase confound tolerance. While it has been used in previous high frequency connectivity studies (Lee et al., 2013), it is still controversial in studies of low frequency connectivity and may eliminate possible neuronal sources as well as inducing negative correlations within a dataset (Fox et al., 2009; Murphy et al., 2009; Uddin et al., 2009; Chang and Glover, 2010; Anderson et al., 2011; Chai et al., 2012; Griffanti et al., 2015).

In this study using a recently developed ultra-high speed multi-slab echo-volumar imaging (MEVI) technique with 136 ms temporal resolution (Posse et al., 2012) and sliding window correlation analysis (Posse et al., 2013), we have three primary objectives. (i) Investigate high frequency connectivity reported in previous studies and characterize the contribution of spatial autocorrelation to the high frequency fluctuations between 0.5 – 3.7 Hz in four major resting state networks (auditory - AUN, sensorimotor - SMN, default-mode – DMN, and visual – VSN). To accomplish this we analyze subjects in 4 separate frequency bands: traditional low frequency (0–0.3 Hz, LF Band), between the respiratory and fundamental cardiac frequencies (~0.5–1 Hz, HF Band 1), between the fundamental and first harmonic cardiac frequencies (~1.5–2.5 Hz, HF Band 2), and a broadband high frequency band with cardiac components removed (0.5–3.7 Hz, HF Broadband). Additionally, we analyze subjects with two distinct seed types at all frequency bands: bilateral Brodmann area based seeds, and unilateral single voxel seeds. (ii) Examine the dependence of high frequency correlations on seed selection (extent, symmetry, and proximity of sub-clusters). (iii) Analyze high frequency connectivity in a patient with a brain tumor that displaces resting state networks, to provide additional evidence for the physiological origin of observed high frequency correlations.

2. MATERIALS AND METHODS

2.1 Subjects

Resting state scans were acquired in 7 right-handed healthy controls (4 female, 3 male) aged between 21–50 years using a scan duration of 5 minutes. Scans were also acquired in a 38-year old female patient with a left hemispheric brain tumor and a 1.5 year history of headaches (Posse et al., 2013). The clinical MRI showed loss of gray-white matter differentiation with multiple areas of gyral expansion, consisting of a small lesion in the left superior frontal gyrus and a large lesion in the left parietal lobe, which was suspected to be a primary glial tumor (e.g. multiple oligodendroglioma or multiple astrocytic tumors). All subjects were instructed to keep their eyes open, clear their mind, relax and fixate on a cross-hair presented on a computer screen. Institutionally reviewed informed consent was obtained.

2.2 Data Acquisition

Data were collected using a 3T Siemens TIM Trio clinical scanner equipped with 12-channel (6 subjects) and 32-channel (1 subject) head array coils. A high-speed MEVI pulse sequence, which combines the efficiency of volumetric encoding with high BOLD sensitivity, while mitigating geometrical distortion and susceptibility related signal losses inherent in single-shot whole brain acquisition methods, was used for data acquisition (Posse et al., 2012). Multiple adjacent slabs were excited sequentially in a single TR and encoded using repeated EPI modules with interleaved phase encoding gradients and fly-back along the k_z – direction, four-fold acceleration using generalized autocalibrating partially parallel acquisition (GRAPPA) along the in-plane phase encoding direction (anterior-posterior), 6/8 partial Fourier encoding along the in-plane phase encoding direction (anterior-posterior), and oversampling along the slab-direction (10%). The MEVI scan parameters were as follows - TR: 136 ms, TE_eff: 28 ms, α: 10º, two slabs in AC/PC orientation, slab thickness: 42 mm, inter-slab gap: 10%, raw data acquisition matrix size for each slab (including 6/8 partial Fourier and 4xGRAPPA): 64×12×8. Field of View (FOV) per slab: 256×256×48 mm³, voxel size: 4×4×6 mm³, 2200 scan repetitions, scan time: 5 min and 16 s. The in-plane image reconstruction was performed on the scanner using a Hanning raw data filter, images were transferred to an external Linux workstation in DICOM format with abbreviated header information, and through-plane reconstruction was performed in real-time using our custom TurboFIRE (Turbo Functional Imaging in REal-time) fMRI analysis tool as described previously (Posse et al., 2013). The through-plane reconstruction combined the inner slices from each of the two slabs that overlap the gap and discarded the outer slice of each slab, resulting in a final image matrix size of 64x64x13 voxels.

Data comparing high frequency connectivity with 3- and 4-fold GRAPPA acceleration were collected in one healthy control, using TR/TE: 162/33.5 ms for 3-fold GRAPPA acceleration and a shorter scan duration with 1000 scan repetitions each.

2.3 Data Preprocessing

Data analysis was performed using the TurboFIRE real-time fMRI analysis software tool (version 5.14.5.1) (Gembris et al., 2000; Posse et al., 2001) and custom MATLAB (Mathworks Inc., Naticket, MA, USA) scripts. Raw image data were motion corrected and converted to analyze format using TurboFIRE. Frequency bandpass filtering was performed in MATLAB. Further preprocessing steps in TurboFIRE included spatial normalization into MNI space using the SPM99 EPI template (Gao and Posse, 2003), segmentation into 144 brain regions in reference to the Talairach Daemon Database (Lancaster et al., 1997; Lancaster et al., 2000) that segregated left and right hemispheric regions, and spatial smoothing with an 8mm isotropic Gaussian spatial filter. Low frequency connectivity data were preprocessed using the same steps without frequency filtering, as described previously (Posse et al., 2013).

2.4 Bandpass Filter Design

In an initial exploratory analysis, high frequency correlations across the spectrum from 0.5 to 3.7 Hz (HF Broadband) were assessed by removing low frequency fluctuations, respiratory, and cardiac signals using multi-band finite impulse response (FIR) bandpass filters. For quantitative analysis, high frequency connectivity in two frequency bands of interest (HF Band 1 and HF Band 2) was assessed using single passband FIR filters (Fig. 1B, C). Equiripple FIR passband filters were designed and implemented in MATLAB using “designfilt” and “filtfilt”, respectively, which are part of the signal processing toolbox (MATLAB, 2016). Filters were designed with a stop band attenuation of −60 dB and passband ripple of 0.1 dB (filter order was minimized using “designfilt” to ~ 200). The precise placement and bandwidth of the FIR pass bands was specified independently for each subject based on a time-frequency analysis of the unfiltered data using a 256 point Kaiser window with a beta value of 10 (Fig. 1), which measured the drift of the heart rate during the scan and the full width at half maximum (FWHM) of the cardiac peak in each time window.

Figure 1.

Spectra of frequency bands of interest. Integrated power spectra (top) and time-frequency power spectra (bottom) of unfiltered (A) and bandpass filtered data (B, C) from a central region of the brain with strong physiological pulsation. (B) 0.5 Hz to fundamental frequency of cardiac pulsatility (HF Band 1). (C) Fundamental frequency to 1^st harmonic of cardiac pulsatility (HF Band 2). Stop bands were made wide enough to account for fluctuations in respiratory (0.3 Hz) and cardiac (1.25 Hz) frequencies during the scan, and a 1.05 Hz machine artifact. Color scale: power/frequency [dB/rad/sample]. A.U.: Arbitrary units.

The lower and upper bounds of the cardiac frequency bands were determined for each subject by first identifying the time windows with the minimum and maximum cardiac frequencies in the time-frequency power spectrum and then using a 10 % amplitude threshold of the cardiac peak in these windows. The broadening of successive cardiac harmonics was taken into account when determining the passband frequencies for each frequency band. The first 800 images were ignored in subsequent analysis steps to account for memory limitations in TurboFIRE, filter initialization, and possible gradient heating related drifts in the initial period of data acquisition.

2.5 Resting State Connectivity Analysis

Analysis of connectivity was performed in TurboFIRE using windowed seed-based connectivity analysis (wSCA), which employs sliding window correlation analysis (Gembris et al., 2000) with a running mean and standard deviation (Posse et al., 2013). This approach, as our previous studies on low frequency connectivity have shown, minimizes the effects of confounds without the need for regression (Posse et al., 2013; Vakamudi et al., 2014). As described by (Leonardi et al. 2015), sliding window correlation analysis provides high-pass filter characteristics as well as low-pass filtering of fluctuations in correlation, both with cut-off frequency 1/(window width). For the present study, we chose filter windows considerably longer than the periodicity of the lowest correlation frequency of interest. As a consequence, the confidence interval for detecting significant correlation within an individual window varies only very slowly for the range of window widths as well as for the averaged data in this study (Leonardi et al., 2015; Zalesky and Breakspear, 2015). Averaging the correlations from successive windows suppresses the effect of correlation fluctuations in individual windows, which removes artifacts due to local confounding signal changes, such as residual head movement effects or signal spikes.

Bilateral seed regions were initially selected as Brodmann areas corresponding to the four RSNs of interest (AUN: BA 41, SMN: BA01 and BA02, DMN: BA07, and VSN: BA17 and BA18) as well as an extended white matter area in the centrum semiovale. White matter seeds were delineated in reference to the MNI atlas built into TurboFIRE. Blurring due to the underlying PSF and spatial filters were accounted for by selecting regions at least 3 voxels away from gray matter areas. Windowed SCA was performed using 15 s sliding window with cumulative meta-statistics (running mean and standard deviation of windowed correlation maps (Welford 1962)), which was optimal for confound suppression. However, an 8 s window, which was used in the initial analysis of HF Broadband data, yielded very similar results. A spatially constrained cluster analysis, limited to the seed region and without using correlation thresholding, was performed on these maps in order to identify the maximum and mean correlations within each of the seed regions. Mean correlations from the cluster analysis were used for quantitative comparison with simulation. Single voxel seeds at the peak voxel location were used to generate additional connectivity maps. The same procedure was applied to the patient data. A spatially unconstrained cluster analysis at correlation threshold = 0.3, was performed with these single voxel seeds to quantify extent of networks. Clusters were identified as contiguous voxels with correlation values above threshold.

An additional analysis was performed using the first half of the data to identify the single voxel seed and the second half of the data to analyze that seed. This was done to investigate the possibility of a circular seed selection bias. Preprocessed low frequency connectivity data were analyzed using wSCA with a 15 s sliding window as described previously (Posse et al., 2013).

2.6 Modeling Spatial Autocorrelation

Simulations to estimate baseline spatial autocorrelations were performed in order to assess the significance of the in vivo results. To estimate these spatial autocorrelations, random noise signals with a Rician distribution (Gudbjartsson and Patz, 1995) were generated in MATLAB for 2200 time points, representing voxel time courses in the same matrix as the in vivo data sets (64×64×13 matrix). A simulation iteration was generated for each in vivo data set analyzed. A correlation analysis was performed in the in vivo data sets using a single voxel seed in white matter to estimate the experimental PSF. To account for the spatial smoothing inherent in the MEVI data, which combines the effects of T2* signal attenuation and the raw filter, and the spatial filter in the SCA, a 3D Gaussian was fit to the peak in the in vivo correlation maps. The fitted standard deviations of the Gaussian along the three spatial axes were consistent across subjects: σ_x=σ_y=1.126, σ_z=1.316 voxels. This Gaussian was applied to the simulation data, centered on each voxel individually and convolved. These convolutions were then summed at each time point to model the effective PSF. Using custom MATLAB software, seed regions from the in vivo analysis were loaded in the simulated data. Correlations across the simulation matrix were computed for each seed region and maps were generated for 1400 time points to match the in vivo analysis. Brain masks from in vivo scans were applied. The maximum and mean correlations were calculated for each subject specific seed. To assess the significance of differences between simulated results and in vivo results, correlations were Fisher Z-transformed and a paired T-test across subjects was performed.

Additionally, the effects of the symmetry of the seed, the size of seed, and spatial separation of two distinct, symmetric seed clusters on autocorrelations were investigated by independently varying relevant parameters: voxel number, eccentricity and distance, respectively. While voxel number was varied, eccentricity and distance were held at 0. While eccentricity was varied, voxel number was held at approximately 500 and distance was held at 0 (slight differences in voxel number due to digitization and eccentricity changes were not significant enough to change the behavior of the graph). While distance was varied, eccentricity was held at 0 and voxel number was held at four values (1, 4, 28, 80) over four separate simulations.

3. RESULTS

3.1 Artifacts

When designing the data analysis pipeline for this study, several artifacts resembling functional connectivity were identified. A narrowband 1.05 Hz machine artifact (width: <0.05 Hz), when not filtered, manifested as a spatially coherent correlation pattern that resembled the DMN. This artifact frequency corresponds with the 6^th harmonic of the repetition time (1/(7*TR)), however, its origin is unknown. Correlation analysis seeded near the interface between the two imaging slabs led to strong, extended correlations throughout networks that may reflect signal interference between slabs. Average scan to scan head movement in all subjects was at or less than 1 % of the mean voxel dimension, with the exception of subject 5 where 1.5 % average scan to scan head movement was measured. The peak scan to scan head movement in all subjects was at or less than 10 % of the mean voxel dimension. In all subjects, sliding window correlation analysis with running mean strongly reduced spurious connectivity at tissue edges due to residual movement effects and in white matter compared with conventional correlation across the entire time series (Fig. 2). Comparable confound suppression was observed across a range of window widths between 4 and 30 s.

Figure 2.

Comparison of conventional correlation analysis and sliding window correlation analysis with cumulative meta-statistics. High frequency (HF Band 1) correlations in the auditory network (AUN). (A) T₁-weighted MRI in radiological display. (B) Correlation without sliding window across the entire time series and (C) sliding window correlation analysis with running mean. The correlation maps are threshold at zero to show spurious connectivity in white matter. Data were filtered using a passband from 0.5 Hz to the fundamental frequency of cardiac pulsatility.

3.2 High frequency correlations

High frequency correlations in vivo in all subjects using the multi-band filter (HF Broadband) showed a remarkable similarity to previously observed low frequency connectivity patterns in this data when using Brodmann area seeds (Fig. 3), despite the residual signal amplitude being almost 2 orders of magnitude lower than that of low frequency resting state signal fluctuations. The spatial extent and Z-scores, however, were smaller, which reflects the relatively low SNR of the high frequency data. Interestingly, correlations at high frequencies in the auditory network were stronger than in the other networks, which may be related to our earlier observation that low frequency resting state connectivity in auditory cortex can be detected with sliding window widths as short as 1 s (Posse et al., 2013). For all subjects in this initial analysis, the AUN in the HF Broadband maps resembled the structure of low frequency RSN connectivity previously observed (e.g. Allen et al., 2011). The HF Broadband maps were qualitatively similar to those obtained in HF Bands 1 and 2.

Figure 3.

Comparison of correlations in different frequency bands in a single subject (S4). (A) T₂-weighted MRI in radiological display (B) Low frequency RSN maps (auditory, sensorimotor, default-mode and visual) using single voxel seeds overlaid on raw EVI data. (C) Corresponding HF Broadband (multi-band filter) correlations using bilateral Brodmann area seeds. (D) HF Broadband correlations using unilateral single voxel seeds. (E) HF Band 1 correlations using unilateral single voxel seeds. (F) HF Band 2 correlations using unilateral single voxel seeds.

When measuring high frequency connectivity in discrete frequency bands, sensitivity in HF Band 1 was higher than in HF Band 2. When using Brodmann area seeds in HF Band 1, three of the subjects (2, 3, and 4) did show extended correlations within the DMN and two of the subjects (2 and 3) showed above noise connectivity in SMN (Fig. 4A). In HF band 1, using unilateral single voxel seeds, subjects showed little to no consistent correlations outside the PSF in all but the auditory network, although subjects 2 and 3 showed some connectivity in the SMN and DMN (Fig 4B). In subject 5 and 6, the DMN maps in the first frequency band showed residual edge artifacts corresponding to the slightly higher head movement in this subject. When using Brodmann area seeds, the connectivity in HF Band 2 was qualitatively similar to that in HF Band 1, albeit with lower mean correlation coefficients (Fig. 4C). In HF Band 2, the AUN was the only network to show bilateral correlations with unilateral seeds (subjects 2, 4, and 5), with the exception of subject 3, which showed extended correlations in the other three networks (Fig. 4D).

Figure 4.

High frequency correlation maps measured in 6 healthy subjects (S1–S6) in a single slice. (A) bilateral Brodmann area seeds in HF Band 1 (0.5 Hz to the fundamental frequency of cardiac pulsatility), (B) single voxel seeds in HF Band 1, (C) bilateral Brodmann area seed in HF Band 2 (fundamental frequency to 1^st harmonic of cardiac pulsatility), and (D) single voxel seeds in HF Band 2. The approximate slice location in MNI space for the different seeds is indicated in (A). The correlation maps are threshold at zero to show the observed level of non-significant connectivity across the brain. Seed areas/voxels were selected in 4 networks (AUN, SMN, DMN, VSN) and in white matter (WM).

The extent of the single voxel correlation maps for the LF Band, HF Band 1, and HF Band 2 is quantified across subjects in Fig. 5, which is based on the last half of the data to avoid circular bias in seed selection (as outlined in the last paragraph of section 2.5). In general, the lower the frequency band, the greater the extent. Fig. 6 shows multiple slices from S4 to better visualize the spatial extent and distribution of these networks.

Figure 5.

Quantitative comparison of correlation cluster extent in different frequency bands using single voxel seeds. Spatially unconstrained cluster analysis was thresholded at a correlation coefficient of 0.3 and performed on correlation maps generated using single voxel seeds. Extent values were averaged across all subjects for the four networks of interest (AUN, SMN, DMN, and VSN) for each of the three frequency bands (LF Band, HF Band 1, and HF Band 2). Extent is much higher in the low frequency band and is the lowest in HF Band 2, as is expected. The AUN shows the smallest difference between HF Band 1 and HF Band 2 extents.

Figure 6.

High frequency correlations in multiple slices in a single representative subject (S4). Four networks of interest (AUN, SMN, DMN, and VSN) were analyzed using (A) bilateral Brodmann area seeds in HF Band 1 (0.5 Hz to the fundamental frequency of cardiac pulsatility), (B) single voxel seeds in HF Band 1, (C) bilateral Brodmann area seed in HF Band 2 (fundamental frequency to 1^st harmonic of cardiac pulsatility), and (D) single voxel seeds in HF Band 2. The approximate slice location in MNI space for the different seeds is indicated on the left.

The comparison of the 3- and 4-fold GRAPPA accelerated scans in an additional single subject showed similar pattern of connectivity in HF Band 1 in the AUN (Supp. Fig 1), despite the considerable reduction in g-factor related noise amplification with 3-fold GRAPPA acceleration (Wiggins et al., 2006). The mean correlations across bilateral Brodmann area 41 seeds in HF Band 1 for 3- and 4-fold GRAPPA acceleration were similar (0.45 and 0.48, respectively).

3.3 Simulations

The systematic investigation of the dependence of spatial autocorrelations on seed parameters showed a negative relationship between mean spatial autocorrelations and both the size and eccentricity of an ROI (Fig. 7A, B) and a significantly weaker relationship between mean spatial autocorrelations and the separation of two ROIs (measured from midpoint) dominated initially by a steep drop off followed by small fluctuations around a constant baseline (Fig. 7C). The initial drop off is more prominent with smaller ROI sizes as they are more sensitive to the PSF and can sample smaller separations. The small fluctuations following the drop off are due to the Rician noise profile of the data and digitization effects (Supp Fig. 2), which also effects in vivo data..

Figure 7.

Simulations showing the dependence of spatial autocorrelations on ROI structure. Effect of (A) the number of voxels within an ROI, (B) the shape of an ROI (eccentricity), and (C) the separation of two discrete ROIs on mean spatial autocorrelation for 4 distinct ROI sizes (1, 4, 28, and 80 voxels). Note the difference in scale for each graph.

Comparing in vivo bilateral Brodmann area results and simulations in HF Band 1 (Fig. 8, 9), the AUN and DMN both showed correlation amplitudes significantly above simulation (p = 0.011 and 0.003 respectively). The SMN showed a trend towards significance (p = 0.077) while the VSN was not significantly above simulation (p = 0.552). Within HF Band 2, only the AUN showed correlation values near significance above simulation (p = 0.058), as opposed to the SMN, DMN, and VSN, which were not significantly different from the simulations (p = 0.953, 0.726, and 0.597 respectively). As expected, none of the networks showed significant correlations below the simulated baseline correlations. As expected, the correlations in LF Band were significantly above simulation for all networks observed (p < 0.018), (Fig. 9).

Figure 8.

Qualitative comparison of in vivo and simulation correlation maps using bilateral Brodmann area seeds. Correlation maps for a representative subject (S4) in vivo in the (A) HF Band 1 (0.5–1 Hz) and (B) HF Band 2 (1.5–2 Hz), and (C) the corresponding simulation analysis.

Figure 9.

Quantitative comparison of Fisher Z-transformed mean correlations across all 6 healthy subjects for simulation and in vivo in different frequency bands. All low frequency networks (LF Band: 0–0.3 Hz) showed significantly higher correlations than simulations (p<0.018). Only the AUN and DMN were significant in HF Band 1 (0.5–1 Hz), (p<0.011), and only the AUN was near significance in HF Band 2 (1.5–2 Hz), (p=0.055).

3.4 Patient with brain tumor

The patient showed high frequency AUN correlations consistent with healthy controls when analyzed with extended Brodmann area seeds (Fig. 10). In HF Band 1, correlations on the healthy side (0.49) were higher than those on the tumor side (0.46), both of which were above simulation baseline (0.38). This correlation pattern does not show a tumor induced displacement due to the bilateral, symmetric nature of the seed region (Fig. 10C). When analyzed with a single voxel seed selected using the cluster maximum on the healthy side, a 12 mm anterior displacement of the contralateral component of the AUN in the vicinity of the tumor was detected (Fig. 10D), consistent with the displacement seen in the low frequency connectivity data (Fig. 10B). The seed contralateral to the tumor showed distributed connectivity anteriorly and along the midline (Fig. 10E), while the seed within the tumor showed little connectivity (Fig. 10F).

Figure 10.

Comparison of low frequency resting state connectivity and high frequency correlations in a patient with a low grade glioma in the left parietal lobe. (A) T₂W-MRI in radiological display. (B) Low frequency connectivity using a single voxel seed in right auditory cortex (LF Band). High frequency (HF Band 1) correlation analysis of the auditory network with C) Brodmann area seed regions and (D) single voxel seeds. Single voxel analysis shows an anterior displacement of the HF correlation on the tumor side by ~12 mm. (E) HF Band 1 correlation using single voxel seed contralateral from the tumor. (F) Single voxel seed within the tumor shows strongly reduced HF Band 1 correlations compared with the contralateral side.

4. DISCUSSION

4.1 High frequency correlations

Our results suggest that high frequency resting state network connectivity may exist within the auditory network in both high frequency bands (HF Band 1 and 2) and default-mode network in HF Band 1. The other networks (SMN and VSN) did not show correlations significantly above baseline, although there was a trend within the SMN in HF Band 1. While structurally cohesive correlation maps were measured in the SMN and VSN in both high frequency bands (Fig. 4A, C), this does not necessarily correspond with actual physiological connectivity, as can be seen in the cohesive maps generated via the pure Rician noise simulations (Fig. 8). The non-significant correlations in the VSN may reflect differences in the resting state condition in our study using eyes open condition compared with other studies (e.g. Lee et al., 2013).

When analyzed with unilateral seeds in HF Band 1, only the AUN showed consistent bilateral connectivity in all subjects (Fig. 4B, D), which may reflect SNR limitations in the filtered data. Subjects 2 and 3, however, did show extended correlations within the DMN and SMN. Correlations within HF Band 2 did not show such consistent patterns and were often confined to PSFs, with the exception of AUN, which showed bilateral correlations in half the subjects (2, 4, and 5), and the DMN which showed extended correlations in subject 3 (Fig. 4D). The amplitude of the observed correlations was much smaller compared to the corresponding low frequency connectivity measured in our previous study (Posse et al., 2013), although the spatial distribution was similar, consistent with earlier studies (e.g. Allen et al., 2011). Additionally, the extent of the AUN was greater than other networks in the high frequency bands and had the lowest drop off between HF Band 1 and HF Band 2 (Fig. 5). The strong AUN connectivity is in proximity to the insula, which is highly vascularized and has been shown to contain elevated high-frequency spectral power in the cardiac band (e.g. Kalcher et al., 2014). However, our highly selective frequency filtering approach makes it unlikely that physiological pulsatility significantly contributed to the AUN connectivity.

4.2 Comparison with previous studies

Existing studies have yet to reach a consensus on the properties, origin, and existence of high frequency connectivity. Different studies have observed different RSNs, such as VSN (Lee et al., 2013; Kalcher et al., 2014; Wang and Deshpande 2015) and SMN (Lee et al., 2013; Kalcher et al., 2014), as well as DMN and the attention network (Gohel and Biswal, 2015). As noted above, a recent study has been unable to replicate these results (Olafsson et al., 2015). These studies used a variety of data acquisition techniques (MREG, Inverse Imaging, Spiral, MB-EPI/SMS-EPI), analysis methods (e.g. ICA and SCA), a range of TRs (100–645 ms), confound suppression approaches (e.g. global signal regression, physiological noise regression), which may induce false-positive correlations (Chen et al., 2016). Each study also analyzed a unique set of frequency bands, including 0.5–0.8 Hz and 1.5–5 Hz (Lee et al., 2013), 0.1–0.25 Hz, 0.25–0.75 Hz , and 0.75–1.4 Hz (Kalcher et al., 2014), 0.1–0.5 Hz (Wang and Deshpande, 2015), 0.198–0.5 Hz and 0.5–0.75 Hz (Gohel and Biswal, 2015), and 0.19–0.38 Hz (Olafsson et al., 2015). Additionally, a recent study by (Chen and Glover, 2015) cautions about the origin and existence of these high frequency networks based on the observation that the BOLD effect in the observed networks show a sharp intensity drop-off with increasing fluctuation frequency. Our study introduces a novel technique, which combines a highly sensitive high-speed (TR: 136 ms) segmented 3D-acquisition method (MEVI) (Posse et al., 2012) with rigorous frequency filtering and a seed-based analysis approach that does not use regression, yet is tolerant to confounding signal changes (Posse et al., 2013). We were able to observe cohesive correlation maps within the VSN, SMN, and DMN using Brodmann area seeds at both high frequency bands, and additionally detect high frequency correlation within the AUN in both high frequency bands, even with single voxel seeds. However, we were unable to verify the significance of these correlations in the VSN and SMN as discussed in section 4.3.

4.3. Significance of correlations

Simulations were used in order to assess the significance of our results with respect to the seed-based approach utilized in this study. Our results show that the shape and structure of ROIs in a seed-based analysis have significant effects on the spatial autocorrelations within the region, which can resemble connectivity. In general, compact, symmetric seed regions will lead to higher spatial autocorrelations. In ROIs with fewer voxels, any single voxel time course will more strongly bias the average signal, leading to a higher correlation. This effect is enhanced by the underlying PSF, further increasing the contribution of the more central voxels to the average signal. These effects cause smaller ROIs to appear to have higher correlations, even with no underlying connectivity in the source data. The limiting case of this effect is the single voxel seed, which, assuming no underlying connectivity, generates an isolated PSF.

The dependency of an ROI's autocorrelation on its symmetry and the relative separation between different components is due to the nature of the underlying PSF, which drops off sharply with distance. This PSF is anisotropic for the MEVI sequence, exhibiting stronger blurring along the slice direction than in-plane (Posse et al., 2012). As a result, the total signal from the central voxel within an ROI will be maximized by an ellipsoidal region, with a slight eccentricity in the Z-direction. Further, the patterns shown in Fig. 7C shows that separation between two isolated seed areas or adjacent sub-regions of an extended seed area is dominated by the PSF drop off for small separations. For larger ROIs (voxel number > 4), typical of this study, this effect is considerably smaller in magnitude compared with the effects of ROI shape and size.

The above relationships between ROI structure and spatial autocorrelations can result in unique correlation profiles that originate from the same non-physiological source. Independently accounting for these baseline spatial autocorrelations for each ROI is necessary to distinguish the signal from noise. This is a particularly important step when performing analyses at high frequencies due to significantly lower SNRs (Chen and Glover, 2015). When comparing simulations with the MEVI in vivo results, the AUN and DMN were the only high frequency RSNs with mean correlations significantly above baseline spatial autocorrelations in HF Band 1, while the AUN was the only RSN with mean correlations significantly above baseline in HF Band 2. The existence of the AUN at high frequencies is further supported by the wSCA with unilateral peak voxel seeds, which shows bilateral connectivity patterns (Fig 4B, D).

To further explore whether these significant correlations represent true network connectivity or are merely confounds that have not been adequately removed, we examined correlations in a patient, which showed a spatial displacement of AUN in low frequency analyses (Fig. 10). When seeded on the healthy side in the high frequency data, we observed a corresponding anterior displacement in connectivity on the tumor side by ~12 mm. If correlations were due to a non-physiological artifact, we would expect it to manifest consistently across healthy controls and the patient, as is the case for the 1.05 Hz machine artifact discussed previously. Additionally, when seeded within the actual tumor itself, correlation patterns were constrained to a PSF. This confirms that the tumor does not exhibit connectivity to healthy tissue, as expected. Overall, this implies that the correlation patterns observed are connected with the underlying AUN, adding to the evidence that we may be observing true high frequency functional connectivity.

The question arises why the AUN shows much stronger correlations than the other RSNs. While we have previously observed the AUN to be strong in low frequencies (Posse et al., 2013), that does not explain the absence of significant connectivity within some of the other networks. The proximity of the AUN to highly vascularized regions, such as the insula, that exhibit high cardiac activity, and the low spatial resolution of MEVI may have introduced sensitivity to residual cardiac pulsatility. While data were filtered as robustly as possible, aliased cardiac pulsation beyond the 2^nd harmonic may still be contaminating results. Additionally, the auditory network has been shown to be affected by scanner noise, showing reduced activation in traditional task-based activation (Shah et al., 1999) and decreased connectivity in low frequency resting state analyses (Rondinoni et al., 2013). The effects of scanner noise on high frequency connectivity have yet to be studied. On the other hand, the existence of high frequency components of BOLD fMRI signals up to 5 Hz have been documented in the visual and motor system enabling estimation of significant dominant directions of information flow using time-domain Granger causality analysis (Lin et al., 2015). A recent study at 7 Tesla using visual stimulation at 0.75 Hz provided further support for fast fMRI signal changes that are consistent with the balloon/Windkessel hemodynamic model, with order of magnitude larger amplitude than predicted from the linear canonical model (Lewis et al., 2016). It is thus of interest to study whether stimulus induced signal changes at high frequencies can be observed in the auditory system.

4.4 Limitations

The low spatial resolution partial brain images acquired in this study limit the precision of spatial normalization and the delineation of seed regions. The low spatial resolution also introduces sensitivity to partial volume effects from blood vessels in highly vascularized brain regions. Rigorous frequency filtering was used to mitigate this possible confound. Despite the short TR (136 ms) in the present study, the resolution of cardiac pulsatility in most subjects is limited to the second harmonic, after which successive harmonics are aliased into lower frequencies. This effect is expected to be small due to the low amplitude of higher harmonics. Neither scrubbing nor despiking were used in this study, which might differentially affect the confound suppression of conventional and windowed SCA approaches, although windowed SCA reduces the effects of confounds. These approaches will be explored in a future study. Head movement, which can confound high frequency connectivity and introduce inter-subject variance in frontal cortex (Yuan et al., 2016), was minor in the present study. We are currently investigating whether the addition of regression of head movement will reduce spurious connectivity in peripheral regions.

The data analyzed in this study were robustly filtered to ensure the suppression of physiological noise up to the second cardiac harmonic of cardiac pulsatility in all subjects. Regression of physiological noise might further suppress residual cardiac pulsatility, but the performance needs to be carefully characterized to avoid the possibility of overfitting low SNR data. Like any seed-based approach, wSCA is highly sensitive to the choice of the seed locations, therefore, seeds were carefully chosen to avoid the edges of the MEVI slabs. Tailoring the seed selection to the functional neuroanatomy of individual subjects may further improve sensitivity. Additionally, the simulations utilized in this study only model thermal noise and do not account for spatial variations of the PSF due to GRAPPA reconstruction related spatial variations in g-factor related noise amplification. However, as shown in Supp. Fig. 1 the HF Band 1 correlations were not significantly different between the 3- and 4-fold GRAPPA accelerated scans, suggesting that the effects of spatial autocorrelations were similar in the two scans. The parameters of the Gaussian model were determined experimentally, as described above, and thus capture the broadening of the PSF due to partial Fourier reconstruction.

4.5 Future Work

Further studies are needed to probe the origin and properties of these observed high frequency correlations. Both fundamental cardiac and relative harmonic strength can vary greatly depending on the region of the brain, capnic state, and disease process being analyzed (Wagshul et al., 2011). Acquisitions with shorter TRs will allow for the unaliasing and filtering of higher cardiac harmonics, although at the expense of spatial resolution and volume coverage. We will also use image reconstruction with extended 16-bit dynamic range to further increase sensitivity. Setting correlation thresholds based on the temporal degrees of freedom taking into consideration temporal autocorrelations as described in (Lee et al., 2013) is under investigation. As described in the methods section, none of the networks showed significant correlations below the simulated baseline correlations, which suggests that our simulations provide thresholds for true-positive correlations that capture most of the spatial and temporal correlation structure in the in vivo data. Further analysis of the temporal auto-correlation structure of the MS-EVI data is in progress (Mutihac et al. 2011), but beyond the scope of the present study.

In the present study, we did not examine the temporal stability of high frequency connectivity. In future studies, we will assess the full spatio-temporal dynamics of high frequency signal correlations. To further characterize the sources of these correlations it will be of interest to measure the BOLD contrast in high frequency bands and to conduct stimulation studies to probe the frequency response of the sensory systems in more detail. Combining low and high frequency resting state connectivity might thus improve the specificity of mapping resting state connectivity beyond what is feasible with current EPI-based approaches (Smith-Collins et al., 2015; Wang and Deshpande, 2015).

5. CONCLUSIONS

This preliminary study establishes resting state connectivity at frequencies from 0.5 – 1 Hz and from 1.5 – 2.5 Hz in the auditory network and from 0.5 – 1 Hz in the default-mode network using subject specific bandpass filtering, which may permit the observation of functional connectivity dynamics at even shorter time scales than currently feasible at frequencies < 0.3 Hz. Simulations verified the significance of correlations in the AUN and the DMN, but not in the SMN and VSN. While this does not rule out the possibility of SMN and VSN at high frequencies, their existence cannot be determined with the current SNR and temporal resolution limits of this study. A patient with a brain tumor was analyzed to observe displacement of high frequency correlation maps as a result of the lesion thus supporting the physiological nature of observed high frequency correlations. Further studies are required to comprehensively assess the behavior and origin of these high frequency correlations.

Supplementary Material

1. Supplementary Figure 1.

Comparison of 3- and 4-fold GRAPPA acceleration in HF Band 1. Three slices in the AUN were analyzed using a single voxel seed showing similar levels of correlation between GRAPPA 3 (top) and GRAPPA 4 (bottom). Only 1000 scans were gathered in this experiment, resulting in nosier data than typically observed in HF Band 1 correlation maps.

2. Supplementary Figure 2.

Dependence of spatial autocorrelations on ROI separation and voxel size. Single voxel simulation results from Fig. 7C with the in vivo resolution (4x4x6 mm³, solid line) and the simulation run with 100 times the spatial resolution (0.04x0.04x0.04 mm³, dashed line). The initial drop, dominated by the PSF (~1–15 mm separation) is similar at both resolutions. The fluctuations apparent from 20 mm separation and onwards are dependent on spatial resolution and noise instantiation of each simulation.

Highlights.

• High frequency correlations in auditory cortex consistent across subjects

• Weaker but significant correlations observed in default mode network

• Spatial displacement of high frequency correlations in vicinity of brain tumor

• Ultra-high speed multi-slab EVI with sliding-window seed-based analysis to mitigate confounding signals

• Simulations of Rician noise to assess significance of seed-based connectivity