Resting-State Functional Connectivity Reflects Structural Connectivity in the Default Mode Network

Abstract

Analyses of resting-state functional magnetic resonance imaging (fMRI) data have demonstrated temporal correlations in the blood oxygen level–dependent (BOLD) signal of widely separated brain regions (Biswal et al. 1995; Cordes et al. 2001). These temporal correlations are presumed to reflect intrinsic functional connectivity and have been demonstrated across several distinct networks serving critical functions like vision, hearing, language, and salience detection (Hampson et al. 2002; Beckmann et al. 2005; Seeley et al. 2007). One such network that has been studied extensively is the default mode network (DMN), a set of brain regions that typically deactivate during performance of cognitive tasks (Raichle et al. 2001). The DMN is detectable using task-free functional connectivity MRI and has been implicated in episodic memory processing (Greicius et al. 2003; Greicius et al. 2004; Buckner et al. 2005). Despite the growing number of resting-state functional connectivity MRI (fcMRI) studies, critical concerns regarding the underlying source of the BOLD signal correlations remain. Several groups have demonstrated that physiologic noise can contaminate the low-frequency oscillations on which fcMRI studies depend (Cordes et al. 2001; Birn et al. 2006; Shmueli et al. 2007). Given the various sources of noise that can contaminate these low-frequency BOLD signal oscillations, healthy skepticism persists as to how closely fcMRI estimates neural connectivity (Maldjian 2001).

Structural connections between brain regions are readily identified in nonhuman primates using tract-tracing methods. Most such studies are undertaken in macaques (Suzuki and Amaral 1994; Morris et al. 1999; Lavenex et al. 2002; Kobayashi and Amaral 2003; Parvizi et al. 2006). Similar studies in humans are not feasible because the most effective tracers require antemortem injections. Postmortem tract tracing in humans is severely restricted by the glacial pace of postmortem tracers and the limited distance they travel (Sparks et al. 2000). Thus, most assumptions about human brain connectivity are extrapolated from tracer studies in monkeys. Tracts connecting several nodes in the DMN have been demonstrated between the homologous regions in macaques. The medial temporal lobe (MTL) and the posterior cingulate cortex/retrosplenial cortex (PCC/RSC) are known to have monosynaptic connections (Suzuki and Amaral 1994; Morris et al. 1999; Lavenex et al. 2002; Kobayashi and Amaral 2003). These MTL–RSC tracts have been invoked in humans to explain why Alzheimer's disease patients show reduced metabolism in the PCC/RSC even in early stages where pathology is limited to the MTL (Greicius et al. 2004). Tracts between the PCC and the medial prefrontal cortex (MPFC) have also been demonstrated in macaques (Parvizi et al. 2006). The uncinate fasiculus has been shown in macaques to connect the anterior MTL with the more inferior, orbitofrontal aspect of the MPFC (Carmichael and Price 1995; Petrides and Pandya 2007).

The relationship between resting-state fcMRI and structural connectivity has not been explored in detail. A study of fcMRI in patients with agenesis of the corpus callosum showed little to no functional connectivity between the auditory cortices in the left and right hemisphere; a null finding suggesting that structural neural connections are required for functional connectivity (Quigley et al. 2003). Recent advances in diffusion tensor imaging (DTI) have allowed noninvasive in vivo measurement of white matter pathways in individual humans (Conturo et al. 1999; Mori et al. 1999; Basser et al. 2000; Catani et al. 2002; Behrens et al. 2003; Dougherty et al. 2005). One prior study (Koch et al. 2002) found that DTI structural connectivity measures between adjacent gyri correlated with task-free functional connectivity measures. This relationship has not been examined across nonadjacent regions in whole-brain networks. The poor temporal resolution of fMRI precludes determining whether functional connectivity between 2 nodes reflects a direct (potentially monosynaptic) connection or is conducted across multiple synapses via intervening nodes. Here, we combined DTI tractography and resting-state fMRI to test the general hypothesis that resting-state functional connectivity reflects structural connectivity. Based on the tract-tracing data in macaques, we hypothesized more specifically that direct connections between PCC and MTL and between PCC and MPFC would be detectable whereas direct connections between MPFC and MTL would not.

Methods

Functional MRI

Six subjects underwent an 8-min, task-free spiral in/out functional scan (Glover and Law 2001) with the following parameters: time repetition (TR) = 2.5 s; time echo (TE) = 40 ms; flip angle = 85°; slice thickness = 4 mm; skip = 1 mm; 30 slices. Subjects were instructed simply to close their eyes and hold still.

After discarding the first 4 volumes to allow for stabilization of the magnetic field, the images were concatenated across time into a single 4-dimensional image. The 4-dimensional image was then spatially smoothed with an 8-mm kernel, temporally filtered with a 100-s high pass filter, and subjected to independent component analysis (ICA) with FSL's melodic software (www.fmrib.ox.ac.uk/fsl/melodic2/index.html). ICA is a statistical technique that separates a set of signals into independent—uncorrelated and non-Gaussian—spatiotemporal components (Beckmann and Smith 2004). When applied to the T2* signal of fMRI, ICA allows not only for the removal of artifact (McKeown et al. 1998; Quigley et al. 2002) but also for the isolation of task-activated neural networks (McKeown et al. 1998; Gu et al. 2001; Calhoun et al. 2002). Most recently, ICA has been used to identify neural networks, including the DMN, during task-free or cognitively undemanding fMRI scans (Greicius et al. 2004; van de Ven et al. 2004; Beckmann et al. 2005). We allowed the software to estimate the optimal number of components for each subject. All components were then transformed into standard MNI space, and the DMN component was selected from among each subject's independent components using an automated template-matching procedure as we have done previously (see Fig. 1 in (Greicius et al. 2007). This procedure includes a frequency filter step that excludes any components in which 50% or more of the power in the frequency domain is above 0.1 Hz.

Open in a separate window

Figure 1.

Functional connectivity reflects structural connectivity in the DMN. (a) Task-free, functional connectivity in the DMN is shown in a group of 6 subjects. The PCC/RSC and MPFC clusters are best appreciated on the sagittal view. Prominent bilateral MTL clusters are seen on the coronal image (left side of image corresponds to left side of brain). (b) DTI fiber tractography in a single subject demonstrates the cingulum bundle (blue tracts) connecting the PCC/RSC to the MPFC. The yellow tracts connect the bilateral MTL to the PCC/RSC. Note that generally the tracts from the MPFC enter the more rostral aspect of the PCC/RSC ROI corresponding to the PCC proper, whereas the tracts from MTL enter the more caudal aspect of the PCC/RSC ROI corresponding to the RSC proper. Left and right columns show slightly different views of the same tracts to highlight the distinct entry points into the PCC/RSC. There were no tracts connecting the MPFC to the MTL.

After selecting the best-fit DMN component, the native space DMN component for each subject was aligned to the custom template space in which the DTI analyses were performed (described below). This was done by aligning T2-weighted images acquired in the same session and with the same prescription as the fMRI data to that subject's T1-weighted 3D SPGR anatomical image volume using a mutual information 3-D rigid-body coregistration algorithm from SPM2 (Ashburner and Friston 2003). The spatial normalization parameters that map each individual brain to the custom template were computed based on the T1-weighted images. A one-sample t-test was performed with SPM2 on the normalized DMN maps. Significant clusters in the resulting group-level map were determined using a joint probability threshold (Poline et al. 1997) of P < 0.001 (height and extent, corrected) and a minimum cluster size of 100 voxels (Fig. 1A). Regions of interest (ROIs) were selected from clusters in the MPFC, PCC/RSC, and MTL bilaterally. After visually selecting a one-voxel seed-point within each cluster on the functional map, the final ROIs were grown around the seed-point using an in-house 3-D flood-fill algorithm. The algorithm grew a spatially contiguous ROI by beginning with the seed point and filling all connected voxels with a t statistic that was within 20% of that of the seed voxel. For the smaller MTL ROI, we also limited the radius of the 3-D flood-fill algorithm to a maximum of 25 mm from the seed voxel to prevent the ROI from extending beyond the MTL region. These 4 group-level, spatially normalized ROIs obtained from the fMRI data were then warped back to each individual brain so that they could be used for the subsequent DTI tractography analyses. This was done by applying the inverse of the spatial normalization transformation. As fiber tracking becomes unreliable in gray matter, we ensured that our ROIs extended 2–3mm into the white matter.

Results

Figure 1A shows the DMN resting-state connectivity map from which the ROIs were taken. Single-subject tracts connecting the PCC/RSC to the MPFC and MTL are shown in Figure 1B. The MTL fibers generally enter the caudal, retrosplenial portion of the PCC/RSC ROI, whereas the MPFC fibers enter the more rostral (PCC proper) portion of this ROI. This rostral–caudal distinction—MPFC–PCC/RSC tracts entering the PCC proper and MTL–PCC/RSC fibers entering the more caudal RSC—was a consistent finding across individual subjects (Fig. 2). There were no tracts connecting the MPFC to the MTL.

Open in a separate window

Figure 2.

DTI fiber tractography in 3 additional subjects. In all cases, the MPFC fibers (blue tracts) enter the rostral (PCC proper) aspect of the PCC/RSC ROI, whereas the MTL fibers (yellow tracts) enter the more caudal aspect (RSC proper) of the PCC/RSC ROI. Left and right columns show slightly different views of the same tracts to highlight the distinct entry points into the PCC/RSC. There were no tracts connecting the MPFC to the MTL.

The group-level map of the tracts is shown in Figure 3 using a subject count threshold. Voxels with tracts from fewer than 5 subjects are not colored. The tracts connecting PCC/RSC to MPFC were detected in 22 of 23 subjects, and those connecting PCC/RSC to MTL were detected in all 23 subjects (to both left and right MTL ROIs). Although these tracts were detected in 22 to 23 of 23 subjects, the fiber density maps in Figure 3 show voxel values less than 22 due to the individual variability in the precise path of the tracts from one voxel to the next. As with the single-subject tracts, the group-level map shows that fibers from the MTL enter the more caudal, putative retrsoplenial, portion of the PCC/RSC ROI, whereas the fibers from the MPFC enter the more rostral aspect corresponding to the PCC proper. There were no tracts connecting the MPFC and MTL using this group-level method and at the single-subject level tracts connecting MPFC to MTL were found in none of the 23 subjects. Given the risk of false negatives in DTI, we repeated the search for MPFC–MTL tracts using substantially looser angle thresholds but again detected none.

Open in a separate window

Figure 3.

Tracts are consistent across subjects. (a) Sagittal views of the MPFC–PCC/RSC tracts (blue–green) and MTL–PCC/RSC tracts (red–yellow) where the color scales indicate the number of subjects that had a tract in a given voxel. The more medial sagittal views at x = 10 and x = 15 demonstrate the distinct entry points of the 2 tracts into the PCC/RSC with the MPFC fibers entering rostrally and the MTL fibers entering caudally. (b) Coronal views show the red–yellow MTL fibers coursing between the hippocampus and parahippocampal gyrus (whose anterior extent at y = −10 includes entorhinal cortex). There were no tracts connecting the MPFC to the MTL.

Discussion

We combined resting-state fcMRI with DTI to demonstrate that resting-state functional connectivity reflects, to a large degree, the underlying structural connectivity. This finding, coupled with work in patients with agenesis of the corpus callosum (Quigley et al. 2003) should help refute some of the (healthy) skepticism surrounding resting-state fcMRI analyses (Maldjian 2001). Although physiologic noise from several sources can certainly contaminate the connectivity maps (Cordes et al. 2001; Birn et al. 2006; Shmueli et al. 2007) the preponderance of evidence now suggests that resting-state fcMRI is a marker of neural connectivity.

Several lines of evidence suggest that the particular network investigated here, the DMN, plays a vital role in episodic memory processing. First, it includes bilateral hippocampus clusters both in humans and macaques (Greicius 2004; Vincent et al. 2007). Second, although deactivated by most cognitive tasks, it is activated by episodic memory tasks (Maguire and Mummery 1999; Buckner et al. 2005). Finally, it overlaps considerably with brain regions targeted by Alzheimer's disease (Greicius et al. 2004; Buckner et al. 2005), the quintessential disorder of episodic memory. By isolating tracts within the DMN connecting the RSC to the MTL, the current study adds to the evidence linking the DMN to episodic memory. These RSC–MTL pathways have been identified in macaque tracer studies but not previously shown explicitly in humans. We suspect that the tracts shown here are a subset of the descending cingulum bundle demonstrated in previous studies (Wakana et al. 2004; Concha, Beaulieu, et al. 2005; Concha, Gross, et al. 2005). Those earlier studies used manually selected ROIs drawn in white matter to define tracts and did not explicitly define terminal connections. By contrast, in the current study, we used functional ROIs in the cortex and hippocampus and defined terminal connections in cortex by using an “end-in” criterion. In addition to the RSC–MTL tracts demonstrated here, the descending cingulum bundle likely encompasses other tracts connecting the PCC/RSC with regions in the lateral temporal lobe and thalamus (Morris et al. 1999).

The caudal entry point of MTL fibers into the PCC/RSC (compared with the more rostral entry of MPFC fibers) is consistent with monkey studies showing that RSC (as opposed to PCC proper) is densely connected with MTL (Suzuki and Amaral 1994; Morris et al. 1999; Lavenex et al. 2002; Kobayashi and Amaral 2003). Our MTL ROIs contain hippocampus proper as well as the inferiorly adjacent parahippocampal and entorhinal cortex. On coronal sections (Fig. 3), the tracts appear to course along the white matter between the hippocampus and the parahippocampus (posteriorly) and entorhinal cortex (anteriorly). As such, we cannot distinguish which subregions of MTL are most densely connected with the RSC. We would predict, from work in macaques, that the bulk of the connectivity is via the entorhinal cortex but tracer studies have also shown direct projections between RSC and hippocampus (Kobayashi and Amaral 2003). Future studies creating ROIs from MTL subregions should enable still more detailed mapping of these tracts, but precise mapping will require higher resolution diffusion data to resolve the exact location at which fibers turn to enter the gray matter.

DTI tractography, as used here in conjunction with resting-state fMRI, has the potential to enrich our understanding of human functional networks. Although functional connectivity implies structural connectivity, it cannot distinguish between direct (potentially monosynaptic) and indirect (potentially multisynaptic) connections. We have demonstrated, for example, that within the DMN the PCC/RSC appears to have direct projections to the MPFC and MTL, whereas the MTL and MPFC appear not to have direct connections. The uncinate fasiculus has been shown to connect the MTLs to the inferior aspect of the MPFC (Wakana et al. 2004) but not to the more superior MPFC region examined here. One must be wary of false negatives, especially in tractography studies. However, the majority of known false negatives in tractography studies involve small, laterally running tracts that are lost when crossing larger tracts like the longitudinal fasiculii and corona radiata (Dougherty et al. 2005). Further, in a post hoc analysis, we loosened our path-tracing thresholds substantially and were still unable to detect tracts connecting our MTL and MPFC ROIs. As such, the absence of connections here between the medially located MPFC and MTL likely reflects a true negative, suggesting that functional connections between these 2 nodes are mediated via a third party such as the PCC/RSC.

The tractography results also point to the coarseness of functional connectivity maps when considered in isolation. The DMN identified here includes a large functional cluster that probably collapses 2 distinct functional regions, the PCC and the RSC, into a single PCC/RSC cluster. The DTI data, showing tracts from the MTL entering caudally into the RSC and tracts from the MPFC entering rostrally into the PCC, certainly suggest that these adjacent regions have distinct projections. In this instance, the DTI results have the potential to better outline the hierarchy of connectivity in the DMN and suggest that there is a flow of information from MTL to RSC to the adjacent PCC to the MPFC. In these data, the functional and structural connections from RSC to the adjacent PCC cannot be shown but must instead be assumed. Evidence from a recent ROI-based study (see figure 5b in Vincent et al. [2006]) supports this distinction between RSC and PCC connectivity by showing that RSC–MTL connections are stronger than PCC–MTL connections, though the latter were also detectable. Future work with both ICA and ROI-based studies should help further delineate the detailed functional and structural architecture of the DMN. It would be informative, for example, to see if forcing ICA to generate more components would result in a splitting of the DMN into 2 or more subnetworks. The hypothesized functional path from MTL to RSC to PCC to MPFC could also be assessed by directly contrasting ROI-based maps of PCC and RSC connectivity using a paired t-test.

In describing the relationship between functional and structural connectivity in this study, we have been careful to note that functional connectivity “reflects” structural connectivity. As our results demonstrate, although functional connectivity reflects structural connectivity to a large degree there is not a simple one-to-one mapping. This is most evident in the lack of MPFC–MTL tracts described above strongly supporting the intuitive notion that functional connectivity can exist in the absence of direct, monosynaptic connections. The converse, that structural connectivity can exist in the absence of functional connectivity, is less intuitive and not addressed by our data but nonetheless quite plausible. Perhaps the best example of this scenario relates to context-dependent changes in connectivity. Several groups have demonstrated that functional connectivity between regions with known or presumed structural connections can wax or wane depending on contextual variables such as awareness of learning (McIntosh et al. 2003) or task performance (Hampson et al. 2004; Rissman et al. 2004). In sum, although our data show that resting-state functional connectivity reflects structural connectivity to a large degree, each can exist without the other.

The main limitation of our study is that our survey of the DMN was incomplete in that we were not able to interrogate the structural connectivity of the bilateral angular gyri. Our group (Greicius et al. 2003) and others (Vincent et al. 2006) have shown that the bilateral angular gyri are key nodes in the DMN. Unfortunately, our ability to examine tracts running laterally to and from the angular gyrus is severely restricted owing to the problem of resolving crossing fibers in tractography analyses (Mori and Zhang 2006; Peled et al. 2006). Although it is quite likely that there are direct connections between the angular gyrus and the PCC (and perhaps to the MTL and MPFC), we assumed, a priori, that these smaller laterally running fibers would be impossible to trace through the larger perpendicularly oriented tracts like the anterior-to-posterior oriented superior longitudinal fasiculus and the superior-to-inferior oriented corona radiata (Dougherty et al. 2005). This assumption was confirmed in an analysis (data not shown) in which we were only able to find tracts connecting the PCC to the angular gyrus in 4 of 23 subjects, tracts connecting the MPFC to the angular gyrus in 4 of 23 subjects, and tracts connecting the MTL to the angular gyrus in 2 of 23 subjects. Future work using novel algorithms designed to overcome the problem of crossing fibers (Peled et al. 2006) will hopefully allow for a more complete investigation of structural connectivity in the DMN and other functional networks.

Subsequent, more detailed, applications of combined DTI tractography and functional connectivity MRI to both controls and patient groups should greatly enhance our understanding of functional (and dysfunctional) brain networks. To cite a specific example, independent groups have already shown that functional (Greicius et al. 2004) and structural (Zhang et al. 2007) measures of DMN connectivity have potential utility in distinguishing Alzheimer's disease patients from controls. Rather than examining these DMN measures in isolation, the current findings suggest that combining structural and functional approaches in the same subjects should ultimately allow for superior characterization of the DMN changes caused by Alzheimer's disease (Buckner et al. 2005).

Funding

Alzheimer's Association (NIRG-04-1060); National Science Foundation (BCS-0449927); National Institutes of Health (NS048302, EY015000, HD047520, NS058899).

Acknowledgments

References