NeuroImage 16, 985 992 (2002) doi:10.1006/nimg.2002.1141 T* 2 Dependence of Low Frequency Functional Connectivity S. J. Peltier* and D. C. Noll *Department of Applied Physics and Department of Biomedical Engineering, University of Michigan, Ann Arbor, Michigan 48109 Received November 27, 2001 Resting state low frequency (<0.08 Hz) fluctuations in MR timecourses that are temporally correlated between functionally related areas have been observed in recent studies. These fluctuations have been assumed to arise from spontaneous blood oxygenation level-dependent (BOLD) oscillations. This work examines the T* 2 characteristics of the low frequency fluctuations (functional connectivity) and compares them to those of task activation induced signal changes. Multi-echo spiral data were fit using a mono-exponential decay model to generate T* 2 and intensity (I 0 )parameter timecourses. Resultant correlation maps show that both functional connectivity and BOLD activation modulate T* 2, not I 0. Regression analysis also finds that both have a linear dependence on echo time. Thus, functional connectivity and task activation MR signal changes appear to arise from the same BOLDrelated origins. 2002 Elsevier Science (USA) Key Words: functional connectivity; multi-echo imaging; signal characteristics; BOLD. INTRODUCTION Recent studies in functional MRI have shown slowly varying timecourse fluctuations in resting state data that are temporally correlated between functionally related areas. These low frequency oscillations ( 0.08 Hz) seem to be a general property of symmetric cortices, and have been shown to exist in the motor, auditory, visual, and sensorimotor systems, among others (Biswal, 1999; Lowe, 1998; Xiong, 1999; Hyde, 1999, Cordes, 2000). Thus, these fluctuations agree with the concept of functional connectivity: a descriptive measure of spatio-temporal correlations between spatially distinct regions of cerebral cortex (Friston, 1993). Several recent studies have further shown decreased lowfrequency correlations for patients in pathological states such as callosal agenesis (Lowe, 1997) or cocaine use (Li, 2000). Thus, low frequency functional connectivity is potentially important as an indicator of regular neuronal activity within the brain. Multi-echo studies have shown the ability to quantify T* 2 and I 0 values during functional studies (Gati, 1997; Speck, 1998; Posse, 1999). Changes in the initial signal amplitude (I 0 ) arise from inflow saturation effects or hardware instability, while changes in the transverse relaxation time (T* 2 ) directly measure the BOLD-related signal change (during task activation). Thus, separating functional connectivity into its T* 2 and I 0 dependencies is a potential indicator of its BOLD origins. BOLD percent signal change has a linear dependence on echo time, to first order (Menon, 1993). In a recent study, we demonstrated that some functional MRI noise components can also have echo time dependencies (Peltier, 2000). While signal drift and respiration effects are largely echo time-independent, cardiac effects can have a significant echo time dependence, perhaps due to pulsatile flow. This echo time-dependent source of signal variation can reduce the detection of true activation. Characterization of the echo time dependence of functional connectivity is therefore also important to evaluate these correlated fluctuations as a potential confound in functional MRI. The present work seeks to ascertain the T* 2 characteristics of the low frequency fluctuations and compare them to those of normal BOLD activation. A multiple echo time acquisition pulse sequence was used to acquire task and resting state data sets, which were then used to form T* 2 and I 0 data sets. Correlation maps of both task activation and resting state functional connectivity were analyzed to compare the echo time dependence of BOLD task activation and the spontaneous low frequency oscillations. METHODS Acquisition A series of multi-echo fmri experiments were performed on a 1.5 T Signa-LX scanner (GE Medical Systems, Milwaukee, WI) using a modified spiral acquisition. The sequence implemented a spectrospatial excitation pulse containing four refocused spiral echoes with 20.48 ms spacing. An initial TE of 7.7 ms was used, to give echo times of 7.7, 28.18, 48.66, and 69.14 ms. Pulse sequence parameters were TR/FA/Slice thickness of 520 ms/45 /5 mm. A 20-cm FOV was ac- 985 1053-8119/02 $35.00 2002 Elsevier Science (USA) All rights reserved.
986 PELTIER AND NOLL quired, with a 56 56 matrix size, yielding a nominal in-plane resolution of 3.6 mm 2. Five axial slices were acquired in each TR, with a total of 400 vol being collected. Six subjects were studied under conditions of activation and rest. Four sets of data, two task sets, and two resting state sets were acquired for each subject. To minimize the effect of possible attentional and/or physiological changes on the data, one set of resting data was acquired before and one set after the two sets of activation data for each subject. A sequential finger-tapping motor paradigm (20.8 s fixation, 20.8 s task, 5 repeats) was implemented for the activation data, with subjects performing one-handed finger-tapping with their dominant hand. The paradigm cues were visually presented to the subjects using IFIS (Integrated Functional Imaging System Psychology Software Tools, Pittsburgh, PA). Resting state data were acquired while the subjects were inactive (lying still, with visual fixation cross being presented), and matched to the duration of the activation data. The cardiac rhythm of the subjects was recorded during all runs using a pulse oximeter. Postprocessing Motion detection was performed on all data in order to detect gross head movement. This analysis was performed using 3-D rigid-body registration in AIR (Automated Image Registration) (Woods, 1998). Following prior work (Lowe, 1998; Cordes, 2000), a cutoff of 0.4 mm was used for displacement in x, y, or z. One subject exceeded this motion threshold and was excluded from the study. The data of the remaining five subjects was then analyzed using the unregistered data, to avoid the introduction of artifactual spatio-temporal correlation through the motion correction process (Lowe, 1998). A method of systematic noise removal was employed on all data following acquisition. Linear trends were removed from the data, to eliminate the effect of gross signal drifts, which are ostensibly due to scanner instabilities or gross physiological changes in the subject. Physiological noise variations in the data due to the cardiac rhythms were then removed using the regression analysis method proposed by Hu et al. (1995). This approach removes the effects of the first and second order harmonics of the externally collected physiological waveforms. Furthermore, any cardiac signal energy below 1.84 Hz (e.g., the primary harmonic) is removed when the low-pass filter is applied (see below). In addition, the timecourse of the highest variance voxel in the sagittal sinus was further used as a regressor (Lund, 2001), in order to remove any cardiac effects that were not in synchrony with the external cardiac measurement. These noise-corrected data was used for all subsequent analysis. Applying a mono-exponential model, I I 0 exp( TE/T* 2 ), and using the multi-echo data, T* 2 and I 0 timecourses were calculated for all voxels. The natural log of the magnitude data for each set of echoes was fit to a first-order polynomial, in order to calculate T* 2 and I 0. The choice of implementing monoexponential fitting versus multiexponential fitting was motivated by the limited number (four) of echoes. Functional connectivity maps (low frequency timecourse correlation maps) were generated for all resting state data, as follows. First, every timecourse in the third echo (TE 48.66) of the task activation data sets was correlated with the motor task reference waveform (square wave with 5.2 s of hemodynamic offset) to form activation correlation maps. Seed clusters were then selected for each slice by identifying the block of four voxels containing the highest average correlation coefficient. This method of using seed clusters has been implemented in other functional connectivity studies (Lowe, 1998; Cordes, 2000). The resting state data was low-pass filtered with a 0.08 Hz cutoff frequency in order to remove the effects of respiration (which typically occurs in 0.1 0.5 Hz) as done in other studies (Biswal, 1999; Lowe, 1998; Xiong, 1999; Hyde, 1999, Cordes, 2000). The seed cluster timecourses in the filtered resting state data had their linear trends removed and were averaged together to create a single low-frequency reference waveform. This low-frequency reference waveform was cross-correlated with all other low-pass filtered voxels to form functional connectivity maps. This process was performed for all slices in all sets of resting state data. Analysis The generated functional connectivity maps were compared to the corresponding task correlation maps and structural images to assess the spatial agreement of the primary and associated motor areas. Qualitative assessment of echo time dependence was investigated using the task activation and functional connectivity correlation maps, for all echo times. Coincidence maps were formed between the T* 2 - weighted and parameter (T* 2,I 0 ) maps to quantitate the distribution of functional connectivity between the T* 2 and I 0 resting state correlation maps. For each slice in every resting state data set, the functional connectivity correlation maps for the T* 2 -weighted (third echo, TE 48.66 ms), T* 2, and I 0 data were thresholded at a given correlation coefficient (for a range of correlation thresholds from 0.3 to 0.9, with a step size of 0.025). The coincident voxels were then found by taking the intersection of the corresponding thresholded maps (T* 2 -weighted and T* 2,T* 2 -weighted and I 0 ). The resulting numbers of coincident voxels were summed over all slices and all subjects. Echo time dependence of task activation and functional connectivity was then examined quantitatively by calculating the normalized signal change versus
T* 2 DEPENDENCE OF FUNCTIONAL CONNECTIVITY 987 FIG. 1. (A) Comparison of the significant activity found in a task activation study (top) and the corresponding resting state functional connectivity study (bottom). The structural overlays are the T* 2 -weighted (third echo, TE 48.66 ms) correlation maps thresholded at 0.4, 0.35 for task, connectivity, respectively. The waveforms are the averaged timecourses from the seed cluster (indicated by a box) for each data set. The task timecourse has been smoothed by 8 points for viewing purposes; the low-frequency waveform was formed using a 0.08 Hz lowpass filter. (B) Comparisons for a typical slice from each of the five subjects.
988 PELTIER AND NOLL FIG. 2. Comparison of correlation maps for (A) task activation (upper) and functional connectivity (lower) at all echo times with the T* 2 -weighted data, and (B) with the calculated T* 2,I 0 data for task activation (upper) and functional connectivity (lower). In (B), the T* 2 -weighted data is taken from the third echo (TE 48.66 ms) of the multi-echo data. echo time in both active and resting states. First, the left and right motor cortices were anatomically defined for all slices (Yousry, 1997). Significant voxels were then selected from the activation data and resting state data. The significant activation voxels were identified as those in the motor cortex contralateral to the exercised hand whose correlation coefficient with the task waveform was 0.5 (p 10 8 ) in the activation exper-
T* 2 DEPENDENCE OF FUNCTIONAL CONNECTIVITY 989 polynomial fitting was performed on the subject-averaged data to test the potential linear and quadratic echo time dependence of both activation and functional connectivity. RESULTS FIG. 3. (A) Coincidence of T* 2 -weighted functional connectivity correlation map (third echo) with T* 2 (left) and I 0 (right) correlation maps. All correlation maps were thresholded (cc 0.35) before forming the coincidence. (B) The number of coincident voxels between the T* 2 -weighted functional connectivity maps and the T* 2 - weighted (solid), T* 2 (dashed), I 0 (dot-dashed), and random (dotted) connectivity maps versus correlation threshold. iments in both the third and fourth echoes. The significant rest voxels were identified as those in the motor cortex ipsilateral to the exercised hand whose correlation with the low frequency reference waveform was 0.35 (p 0.05) in the rest experiments in both the third and fourth echoes. Two echoes were used to avoid the inclusion of any false positives. The mean and the corresponding reference signal (motor task reference waveform or low-frequency reference waveform) were then fit using least-squares to the timecourses of the selected voxels, for all echo times. Normalized signal change was then calculated as twice the rms signal amplitude of each fit timecourse, normalized by the mean (scaled by 2 for the low-frequency reference, to account for the sinusoidal-like waveform). A regression analysis using first- and second-order least-squares Figure 1A shows the activation and low frequency connectivity correlation maps for a typical subject, along with the associated seed cluster timecourses. As shown, the activation correlation map has the expected response to the motor task, with the contralateral motor cortex being significantly activated, and the ROI waveform varying synchronously with the task paradigm. The functional connectivity map also shows activation in the contralateral motor cortex (in addition to the seed cluster), as well as the ipsilateral cortex and supplementary motor areas. This is in agreement with previous functional connectivity studies (Biswal, 1999; Lowe, 1998; Xiong, 1999; Hyde, 1999; Cordes, 2000). Figure 1B demonstrates this behavior is found across all the subjects. The task activation and functional connectivity correlation maps were examined at each echo time. In Fig. 2A, a comparison shows that patterns in both the activation and functional connectivity become more significant in later echo times. Conversely, at 7.7 ms, there is virtually no significant task activation or patterns in the functional connectivity map. This behavior is also seen in Fig. 2B, which compares the correlation maps at the third echo time (48.66 ms) with corresponding correlation maps formed from the T* 2 and I 0 timecourses. It is seen that the activation and functional connectivity patterns seen in the T* 2 -weighted maps are also apparent in the T* 2 maps, but not in the I 0 maps. This result is consistent across all subjects. The coincidence maps of the resting state T* 2 - weighted and T* 2, and T* 2 -weighted and I 0 functional connectivity correlation maps were examined for the T* 2 and I 0 distribution of the patterns. Figure 3A shows the resultant coincidence maps for the connectivity data of Fig. 2B, using a correlation threshold of 0.35. Significant patterns are seen in the T* 2 coincidence map, whereas the largest grouping in the I 0 coincidence map is the seed cluster. The total coincident voxels found across all subjects and all slices versus correlation threshold are shown in Fig. 3B. It is seen that approximately 50% (or higher) of the voxels in the T* 2 -weighted connectivity patterns are found in the T* 2 maps across all thresholds, whereas the I 0 coincidence is less than 20% (at most). To establish a lower bound on coincident voxels, a random set of data was generated (using Gaussian random noise timecourses with the mean and standard deviation taken from the actual subject data, and smoothed using a Gaussian kernel with FWHM of 1 voxel). Functional connectivity maps were formed (using the same ROIs as the subjects), and
990 PELTIER AND NOLL TABLE 1 T* 2 -weighted coincidence (%) Subject T* 2 Coincidence I 0 Coincidence 1 58.79 15.36 2 47.40 15.96 3 44.15 11.70 4 51.97 13.54 5 42.06 9.98 Mean 48.87 13.31 SD 6.69 2.50 Note. Percentage of coincidence of T* 2 -weighted functional connectivity patterns with the T* 2 and I 0 functional connectivity patterns. All connectivity maps were thresholded at (r 0.35) to form the coincidence. FIG. 4. Plot of the normalized signal change vs echo time due to the fit task activation (left) and low frequency functional connectivity (right) reference waveforms. Results are the mean and standard error over all slices and all subjects. the coincident maps calculated. As shown, the I 0 coincidence results do not differ greatly from the random noise coincidence results, especially at thresholds of 0.35 and above. The percentage of coincidence at a correlation threshold of 0.35 for each subject is shown in Table 1. It is seen that the overall average coincidence is 49% for T* 2, but only 13% for I 0. The normalized signal change results at each echo time (TE) for both the activation and connectivity data sets are shown in Table 2 for each subject, with the result over all subjects displayed in Fig. 4. It is seen that both the activation and resting state normalized signal change increase proportionally with echo time. Regression analysis on the subject-averaged data found that both activation and functional connectivity have a significant linear dependence on echo time (p 10 4 for activation, p 0.002 for connectivity). No significant quadratic dependence or intercept was found in either case, as shown by the t-scores of the fit parameters. This agrees with the first-order BOLD signal change relationship of: S S TE R * 2, (1) where R* 2 1/T* 2 (Menon, 1993). Table 3 displays the ratio of the normalized signal change (task over connectivity) for each subject. It is seen that the ratio between task and activation remains largely the same across echo times. This behavior, as well as the ratio magnitude (2.91 0.7 across all subjects), agrees with findings from other studies (Hyde, 1999, 2001). Thus, functional connectivity is seen as a potential BOLD-dependent confound that scales with the signal. TABLE 2 Task activation Functional connectivity Subject Echo 1 Echo 2 Echo 3 Echo 4 Echo 1 Echo 2 Echo 3 Echo 4 1 0.0071 0.0233 0.0397 0.0473 0.0030 0.0102 0.0182 0.0260 2 0.0066 0.0216 0.0390 0.0526 0.0044 0.0084 0.0145 0.0193 3 0.0077 0.0246 0.0480 0.0661 0.0023 0.0076 0.0167 0.0257 4 0.0099 0.0307 0.0482 0.0576 0.0021 0.0086 0.0162 0.0221 5 0.0068 0.0215 0.0397 0.0548 0.0017 0.0061 0.0117 0.0167 Parameter t scores Fit Intercept Linear Quadratic R 2 Activation Quadratic 0.4160 5.024 0.9561 0.948 Linear 1.0495 17.6968 0.946 Functional connectivity Quadratic 0.2763 2.6016 0.4521 0.910 Linear 0.1040 13.4154 0.909 Note. Echo time results for all subjects. (Top) Results for normalized signal change task activation and functional connectivity. (Bottom) Polynomial regression results, using both a quadratic and linear fit on the activation and connectivity data (t 2.2 corresponds to p 0.05)
T* 2 DEPENDENCE OF FUNCTIONAL CONNECTIVITY 991 Subject TABLE 3 Ratio (task/connectivity) Echo 1 Echo 2 Echo 3 Echo 4 1 2.3667 2.2843 2.1813 1.8192 2 1.5000 2.5714 2.6897 2.7254 3 3.3478 3.2368 2.8743 2.5720 4 4.7143 3.5698 2.9753 2.6063 5 4.0000 3.5246 3.3932 3.2814 Mean 3.1858 3.0374 2.8227 2.6009 SD 1.2785 0.5798 0.4417 0.5221 Note. Ratios of normalized signal change (task activation over functional connectivity). DISCUSSION The correlation map analysis revealed that functional connectivity and task activation signal modulate T* 2, not intensity (I 0 ) (Figs. 2B and 3). The significant functional connectivity patterns found in the calculated T* 2 data coincide with 49% of the patterns found in the corresponding T* 2 -weighted data, as opposed to 13% coincidence with the corresponding I 0 data. Regression analysis showed that both are linearly echo time dependent (Table 2; Fig. 4); this agrees with preliminary results found by Hyde and Biswal, 1999. There are other possible theories concerning the underlying source of functional connectivity, such as synchronized cardiac fluctuations or spontaneous oscillations in cerebral flow, however, these are thought to be less likely. Cardiac noise, while capable of having a complex dependence on TE, was minimized in this study through the use of a rapid acquisition rate, which prevents aliasing of the primary cardiac components into the spectral band of interest, as well as postprocessing to remove the effects of the secondary harmonics. Spontaneous flow oscillations would not be expected to be localized to both motor cortices and the supplemental motor area simultaneously. Also, if functional connectivity had an underlying source different than that of activation, which is known to be a BOLD effect, then their ratio of amplitudes would most likely fluctuate with echo time, and not remain as constant as shown in Table 3. Spontaneous spatiotemporally correlated neuronal fluctuations, the presumed source of functional connectivity, are a potential BOLD-dependent source of variance in functional MRI. These fluctuations may exist in both the resting and active states and, from the activation/connectivity ratio results (Table 3), would introduce a consistent level of variance independent of TE. Since the fluctuations are low frequency, they fall within the frequency band of typical MRI activation, and can thus reduce the statistical detection of activation. As this physiological noise mechanism increases with increasing field strength, these fluctuations will have an increased effect on the contrast-to-noise (CNR) ratio of a study as opposed to other sources of noise (e.g., thermal), and may prove to be a limiting source of noise in functional MRI (Hyde, 1999; Krueger, 2001). CONCLUSIONS This study performed a multi-echo analysis of low frequency functional connectivity in the motor cortex. Motor task activation and resting state data were collected for five subjects. T* 2 and I 0 maps were calculated, and activation and functional connectivity maps were formed for comparison. A regression analysis was also performed to fit the reference signal (task paradigm or low frequency reference waveform) to all data at all echo times to explore the echo time dependence of both task activation and functional connectivity. Functional connectivity is seen to be a T* 2 modulatory effect, as is task activation. Furthermore, both functional connectivity and task activation signal changes are linearly dependent on echo time. Hence, functional connectivity and regular task activation both seem to arise from the same BOLD-related origins. ACKNOWLEDGMENT This work was supported by NIH Grant NS32756. REFERENCES Biswal, B., Yetkin, F., Haughton, V., and Hyde, J. 1994. Functional connectivity in the motor cortex of resting human brain using echo-planar MRI. Magn. Reson. Med. 34: 537 541. Cordes, D., Haughton, V., Arfanakis, K., Wendt, G., Turski, P., Moritz, C., Quigley, M., and Meyerand, M. E. 2000. Mapping functionally related regions of brain with functional connectivity MR imaging. Am. J. Neuroradiol. 21: 1636 1644. Friston, K. J., Frith, C., Liddle, P., and Frickowiak, R. 1993. Functional connectivity: The principal components analysis of large (PET) data sets. J. Cereb. Blood Flow Metab. 13: 5 14. Gati, J., Menon, R. S., Ugurbil, K., and Rutt, B. K. 1997. Experimental determination of the BOLD field strength dependence in vessels and tissues. Magn. Reson. Med. 38: 296 302. Hu, X., Le, T. H., Parrish, T., and Erhard, P. 1995. Retrospective estimation and correction of physiological fluctuation in functional MRI. Magn. Reson. Med. 34: 201 212. Hyde, J., and Biswal, B. 1999. Functionally related correlation in the noise. In Functional MRI (C. T. Moonen and P. A. Bandettini, Eds.), pp. 263 275. Springer-Verlag, Berlin. Hyde, J., Biswal, B., and Jesmanowicz, A. 2001. High-resolution fmri using multislice partial k-space GR-EPI with cubic voxels. Magn. Reson. Med. 46: 114 125. Krueger, G., and Glover, G. 2001. Physiological noise in oxygenation sensitive magnetic resonance imaging. Magn. Reson. Med. 46: 631 637. Li, S. J., Biswal, B., Li, Z., Risinger, R., Rainey, C., Cho, J., Salmeron, B., and Stein, E. 2000. Cocaine administration decreases functional connectivity in human primary visual and motor cortex as detected by functional MRI. Magn. Reson. Med. 43: 45 51.
992 PELTIER AND NOLL Lowe, M. J., Rutecki, P., Turski, P., Woodard, A., and Sorenson, J. 1997. Auditory cortex fmri noise correlations in callosal agenesis. Neuroimage 5(2): S194. Lowe, M. J., Mock, B., and Sorenson, J. A. 1998. Functional connectivity in single and multislice echoplanar imaging using resting state fluctuations. Neuroimage 7: 119 132. Lund, T. E., Hanson, L. G. 2001. Physiological noise reduction in fmri using vessel time-series as covariates in a general linear model. In Proc. ISMRM, 9, Glasgow, p. 22. Menon, R. S., Ogawa, S., Tank, D., and Ugurbil, K. 1993. 4 Tesla gradient recalled echo characteristics of photic stimulation-induced signal changes in the human primary visual cortex. Magn. Reson. Med. 30: 380 386. Peltier, S. J., and Noll, D. C. 2000. Analysis of fmri signal and noise component TE dependence. Neuroimage 11: S623. Posse, S., Wiese, S., Gembris, D., Mathiak, K., Kessler, C., Grosse- Ruyken, M., Elghahwagi, B., Richards, T., Dager, S., and Kiselev, V. 1999. Enhancement of BOLD-contrast sensitivity by single-shot multi-echo functional MR imaging. Magn. Reson. Med. 42: 87 97. Speck, O., and Hennig J. 1998. Functional imaging by I 0 and T* 2 - parameter mapping using multi-image EPI. Magn. Reson. Med. 40: 243 248. Woods, R., Grafton, S., Holmes, C., Cherry, S., and Mazziotta J. 1998. Automated image registration: I. General methods and intrasubject, intramodality validation. J. Comput. Assist. Tomogr. 22: 141 154. Xiong, J., Parsons, L., Gao, J., and Fox, P. 1999. Interregional connectivity to primary motor cortex revealed using MRI resting state images. Hum. Brain Mapp. 8: 151 156. Yousry, T. A., Schmid, U., Alkadhi, H., Schmidt, D., Peraud, A., Buettner, A., and Winkler, P. 1997. Localization of the motor hand area to a knob on the precentral gyrus. A new landmark. Brain 120: 141 157.