Abstract
The function of spontaneous brain activity is unknown. Here we test the novel hypothesis that patterns of spontaneous activity reflect not only synaptic homeostasis or synchronization of neuronal populations, but code representational patterns evoked by stimuli and tasks. We compared in visual cortex the spatial patterns of spontaneous activity to the patterns evoked by ecological visual stimuli (faces, bodies, scenes) and low-level visual features (e.g. phase-scrambled faces). We identified regions that preferred particular stimulus categories during localizer scans, measured multivariate spatial patterns for each category during task scans, and then spatially correlated these stimulus-evoked patterns to the pattern measured in each frame of resting-state scans. The mean correlation coefficient was essentially zero for all regions/stimulus categories, indicating that resting activity patterns were not biased toward particular stimulus categories. However, the spread of correlation coefficients, i.e. both positive and negative, was significantly greater for a stimulus category over the ROIs preferring that category (e.g. the body category in body-preferring ROIs). Therefore, the putative representational content of spontaneous activity was related to stimulus-evoked spatial activity patterns. This content also governed the temporal correlation or functional connectivity of spatial patterns of spontaneous activity between individual regions. Resting spatial activity patterns related to an object category (e.g. bodies) fluctuated preferentially between ROIs preferring the same category. Moreover, activity patterns related to different categories fluctuated independently within their respective preferred ROIs. These results support the general proposal that spontaneous multi-voxel activity patterns are linked to stimulus-evoked patterns, consistent with a representational function for spontaneous activity.
Significance Statement Neurons throughout the brain are spontaneously active. Although this activity was once thought to reflect only noise, the remarkable spatiotemporal regularities of spontaneous activity have motivated functional explanations. Here we address the hypothesis that spontaneous activity codes the representational structure evoked by stimuli. We show that the spatial pattern of stimulus-evoked activity across high-level areas of visual cortex that prefer visual categories such as bodies or scenes is related to the spatial pattern of spontaneous activity across the same areas. This spatial structure partly governs the spontaneous spatiotemporal interactions between regions known as functional connectivity, resulting in correlated fluctuations of spatial activity patterns that are specific to particular stimulus representations. These results support a representational framework for understanding spontaneous activity.
Introduction
Spontaneous neural activity is observed throughout the brain, yet its function remains mysterious. An important clue, however, comes from work that has uncovered striking similarities between spontaneous activity and the activity evoked by a task (1–12). For example, the temporal correlation of spontaneous activity between brain regions (functional connectivity, FC) closely resembles the spatial topography of task-evoked activity (13–18), links distributed brain regions into functional networks, and can be used to predict task activation (5, 6).
The remarkable spatiotemporal regularities of spontaneous activity and widespread findings that abnormalities in inter-regional correlations of spontaneous activity in humans are associated with neurological and psychiatric disorders (e.g. (19, 20)) have motivated a search for functional explanations. One hypothesis is that spontaneous activity has a role in the synaptic homeostasis of structural connections (21). Another idea is that fluctuations of spontaneous activity between regions constitute a spatiotemporal prior that facilitates the recruitment of task circuitries during behavior (22, 23). Both hypotheses have been mainly concerned with explaining the FC between regions.
A novel hypothesis is that spontaneous activity and inter-regional FC has a role in representing behaviorally relevant information. Genetically determined circuitries generate spontaneous activity that is shaped in the course of development by experience through Hebbian statistical learning (1). Conversely, spatial and temporal patterns of spontaneous activity constrain task-evoked patterns As a result of this cyclic process, both spontaneous and task-evoked activity code similar representations of internal and external states (24, 25). The same process determines the spontaneous interactions between regions, which reflect connectivity patterns that are coded as synaptic efficacies in cortical networks (25, 26). Figure 1 illustrates schematically the representation hypothesis of spontaneous brain activity.
The ‘representation hypothesis’ is supported by human and animal work. In animals, imaging of neural activity at a scale that extends across many cortical columns has shown that the macro-scale spatial pattern of spontaneous neural activity within a sensory area in an anesthetized animal mirrors the pattern of activity evoked by stimulation of a specific visual feature (7, 8). In humans, fMRI studies of early visual cortex have shown that FC between individual voxels respects the stimulus-evoked selectivity of voxels for polar angle, eccentricity, and low-level stimulus features (9–12). Recent work has also shown that voxel-wise resting FC in visual cortex is better approximated by the FC evoked by movies than by more artificial stimuli such as rotating checkerboards or static pictures of stimuli (26, 27).
However, these studies have not considered the spatial patterns of activity within a region, or the temporal correlation of these spatial patterns across cortical regions. Yet to-date, the best evidence in humans that evoked activity in cortex codes for behaviorally relevant information such as stimulus categories, retrieved memories, or cognitive processes (e.g. attention) has come from multivariate spatial pattern analyses of fMRI signals in visual and associative cortex (28–35). Critical tests of the ‘representation’ hypothesis of spontaneous interactions therefore include predictions about the spatial pattern of activity within a region and the interaction of that spatial pattern with the patterns from other regions.
First, if spontaneous activity carries information about stimulus categories such as faces, bodies, or scenes, then spontaneous spatial activity patterns, i.e. patterns of activity observed at rest in the absence of any stimulation, in functionally specialized occipital regions such as the extrastriate body area (EBA, (36)) should be more related to the spatial patterns of activity evoked by the preferred stimulus category (e.g. bodies) than by other categories. This predicted relationship between spontaneous and task-evoked spatial activity patterns putatively reflects the entrainment of task patterns into spontaneous activity in the course of development and experience. Second, if the FC between regions at least partly reflects correlated fluctuations of the spontaneous representational content of those regions, then the resting inter-regional temporal correlation of patterns of spatial activity should systematically depend not only on whether the regions prefer the same visual stimulus category, but also on whether the correlated spatial activity patterns code the preferred visual category.
Results
The first goal of the experiment was to compare multi-vertex activity patterns measured in the resting state with fMRI to stimulus-evoked activity patterns reflecting a variety of stimulus categories, including those that are more or less ecological (e.g. photographs of faces, tools, and scenes vs. phase-scrambled images of those stimuli). This comparison was conducted in regions of higher-order visual cortex that activated more strongly to stimuli from a particular category (e.g. bodies) relative to other categories (e.g. chairs and tools), and was also conducted in early visual cortical regions (37). To measure spontaneous activity, we ran a set of resting-state scans in which human observers fixated a central point on a blank screen. This activity was measured first to prevent possible learning effects from the other conditions (Fig. S1).
Localization of regions with visual category preferences
To identify category-specific visual regions, we ran a set of localizer scans in which multiple stimuli belonging to one of five stimulus categories (faces, bodies, scenes, man-made objects (chairs and tools), and phase-scrambled versions of these stimuli) were presented in a blocked design (Figs. S1 and S2). We used standard contrasts (as in (38)) to identify category-preferring regions. For instance, activity evoked by body stimuli was subtracted from activity evoked by man-made objects (tools and chairs) to localize body-preferring regions such as EBA (see Figs. S3A and S3B, and Table S1 for all category-preferring regions). Separate contrasts identified regions more active for whole-objects (face+scene+bodies+(tools+chairs)) than for low-level visual features (phase-scrambled objects). Phase-scrambled objects activated more strongly in regions of early visual cortex (V1-V3 based on the maps of (37)), while whole objects activated more strongly lateral and ventral occipital cortex, including some category-preferring regions (Figs. S4B and S4B, Table S1).
Representational similarity analysis of task-evoked patterns
During task scans (Fig. S1), we randomly presented individual stimuli belonging to each category to extract the stimulus-evoked spatial pattern in a particular ROI for each stimulus and corresponding category. Two general linear models (GLMs) were conducted to estimate the stimulus-evoked patterns. One model used stimulus-specific β weights to estimate the spatial activity pattern evoked by each individual stimulus, and the other used category-specific β weights to estimate the stimulus-evoked spatial activity pattern associated with each category.
To show that our stimuli and procedure generated spatial patterns that were consistent with the literature, we conducted a representational similarity analysis (RSA). Importantly, the representational similarity analysis was conducted using the task scans, which were completely independent of the localizer scans used to determine category-preferring ROIs. Figure 2A shows the spatial similarity of multi-vertex patterns evoked by individual stimuli within several classical category-preferring ROIs. In left EBA the highest representational similarity was found between human bodies, and the next highest between pictures of mammals, which included their bodies. In the right fusiform face area (FFA (39)), faces and other animate stimuli (bodies, mammals) generated more similar patterns than stimuli from the inanimate categories chair, tool, and scene, with the most consistent representational similarity found between face exemplars (32, 40). In the scene-preferring region right parahippocampal place area (PPA (41)), the activity patterns evoked by different scenes were well correlated, with low correlations between and within all other categories.
We conducted a second representational similarity analysis using the pattern evoked by a stimulus category, as estimated by the category regressor in a GLM, rather than using the patterns evoked by individual stimuli. Instead of conducting this analysis separately within each localizer-defined ROI, we grouped each set of category-preferential ROIs for an individual into a single joint-ROI. For instance, the body joint-ROI included left and right EBA, left and right fusiform body area (FBA), and so forth, and the scene joint-ROI included constituent regions such as PPA, the transverse occipital sulcus (TOS), and retrosplenial cortex (RSC) (see Table S1 for a complete listing). The results of this analysis (Fig. 2B) were consistent with the literature. Both in body and face joint-ROIs, the highest representational similarity was found between animate categories (face, bodies, mammals) as compared to other categories (32, 40). For instance, the similarity of body- and face-evoked spatial activity patterns in the face joint-ROI was ρ=0.73, while the similarity of face- and scene-evoked patterns in the face joint-ROI was ρ=0.41. Table S2 indicates the representational similarity between the task-evoked spatial patterns corresponding to the face, body, and scene categories within each joint-ROI.
Task-rest pattern similarity analysis in category-preferring regions
We next tested the first prediction of the representation hypothesis, namely that spatial patterns of spontaneous activity at rest in category-preferring regions should be more related to the task-evoked spatial pattern for preferred than non-preferred categories. For each category, the category-evoked spatial pattern was spatially correlated on each resting frame with the spatial activity pattern in a joint-ROI to determine a resting timeseries of correlation coefficients and a corresponding frequency distribution of coefficient values. The upper 90% value (U90 value) of the distribution was used as a summary measure of the relationship between the stimulus-evoked and resting spatial activity patterns.
Figure 3A illustrates this procedure in a single subject using a region that prefers scenes (PPA) and the corresponding category-evoked spatial pattern for scenes. The scene-evoked pattern in PPA is a multi-vertex set of normalized activation values (Fig. 3A, left). This evoked pattern is spatially correlated (ρ) with the spontaneous patterns of activity on each frame of the resting state scans. A timeseries of ρ-values is generated (Fig. 3A, middle), as well as a corresponding frequency distribution of ρ-values (Fig. 3A, the histogram in blue). A U90 value for the joint-ROI and category is then determined from the distribution. The insets in the middle panel show resting frames in which the spontaneous activity (real data) was not correlated (ρ=0.003, outlined in green), positively correlated (ρ=0.81, outlined in magenta), or negatively correlated (ρ=-0.74, outlined in cyan) with the scene-evoked spatial activity pattern. The same procedure was used to generate a U90 value for each of eight categories (faces, bodies, mammals, chairs, tools, scenes, grid-scrambled, phase-scrambled) in each joint-ROI (body, face, scene).
Figure 3B shows the distributions of correlation coefficients across all subjects for each preferred category within its corresponding joint ROI (green hue). For example, the leftmost graph shows the distribution using the body-evoked activity pattern within the body joint-ROI. A second distribution, generated using the spatial activity pattern evoked by phase-scrambled objects within the same body joint-ROI, has been superimposed (black hue). Theoretically, the two distributions might differ in the mean, variance, skewness, or some other parameter. For each joint-ROI, the distribution of correlation coefficients for both the preferred stimulus category and the phase-scrambled category were symmetric and centered on zero. However, the spread of the distribution was higher for the preferred category-evoked spatial activity pattern, meaning that larger correlation coefficients, both positive and negative, were observed for the preferred than phase-scrambled category (red arrows in Fig. 3B). Therefore, a larger U90 value indicates the presence of larger positive matches and negative matches of the resting spatial pattern to the category-evoked spatial pattern. Similar findings were obtained for face (middle panel) and scene activity patterns (rightmost panel) in the corresponding joint-ROIs, as compared to phase-scrambled patterns.
The categorical specificity of spontaneous activity patterns in each joint-ROI was tested by comparing U90 values for different categories. Figure 3C shows mean U90 values for a joint-ROI’s preferred category, defined from its localizer contrast (green symbol; e.g. body in the Body joint-ROI), “non-preferred” categories (red symbols), grid-scrambled category (blue symbol) and phase-scrambled category (black symbol) averaged across subjects. We first conducted an overall repeated measures analysis of variance ANOVA on U90 values with joint-ROI (body, face, and scene) and Category (8 levels) as factors. The main effects of joint-ROI (F(2, 30)=50.4, p<.0001) and Category (F(7, 105)=3.46, p=.002), and the interaction of joint-ROI by Category (F(14, 210)=4.37, p<.0001) were all significant. The interaction indicated that the variation of U90 values across categories depended on the joint-ROI. Separate repeated measures ANOVAs for each joint-ROI with Category (8 levels) as a factor indicated a highly significant main effect of Category in each joint-ROI (Body: F(7,105)=7.25, p<.0001; Face: F(7,105)=3.25, p=.004; Scene: F(7,105)=3.93, p=.0008). Therefore, for each joint-ROI, the spread of stimulus-evoked-to-rest spatial similarity values significantly depended on the category of the stimulus-evoked spatial pattern.
To compare the U90 value for the joint-ROI’s preferred category vs. each other category, we conducted paired t-tests with a Holm-Bonferroni correction for multiple comparisons. Significant, multiple-corrected comparisons are indicated in the figure by plus signs. In the Body joint-ROI, the U90 value for bodies was significantly larger than for chairs, scenes, and phase-scrambled stimuli. In the Face joint-ROI, the U90 value for faces was significantly larger than for scenes. Conversely, in the Scene joint-ROI, the U90 value for scenes was significantly larger than all other categories.
Therefore, the ‘animate’ Body and Face joint-ROIs and the ‘inanimate’ Scene joint-ROI showed a significant double dissociation involving the corresponding categories, with U90 values in the Face and Body joint-ROIs significantly greater for the face and body categories, respectively, than for scene categories, and the U90 value in the inanimate Scene joint-ROI significantly greater for scenes than for either faces or bodies. The U90 values for face and body categories within each joint ROI were similar, reflecting the fact that both are animate categories and have greater cross-category representational similarity with each other than with scenes (Table S2). These results provide some support for the first prediction of the representation hypothesis, namely that spontaneous activity patterns in category-preferring regions are more related to the patterns for some categories than for others. However, ‘more related’ means a greater spread of extreme similarity values, both positive and negative, rather than a shift in the mean to more positive similarity values.
Task-rest pattern similarity analysis in regions preferring whole vs. phase-scrambled objects
The above results showed that the spatial pattern of spontaneous activity in regions of high-level visual cortex that respond preferentially to ecological visual categories was more related to the spatial pattern evoked by one category than another (e.g. bodies vs. scenes). We next asked whether a similar result would be found in regions that show stimulus preferences for low-level features as compared to more ecological categories such as face or body. This result would support a general conclusion that the stimulus preferences of a region largely drive the spatial pattern of spontaneous activity. We used the localizer scans to identify ROIs in which stimulus-evoked responses were stronger or weaker for phase-scrambled objects than for the union of the whole-object categories (face, body, mammal, chair, tool, and scene). The resulting ‘Phase-scrambled objects’ joint-ROI was located in medial posterior visual regions in early visual cortex (V1-V3 according to the Wang template (37) while the ‘Whole-objects’ joint-ROI was located in lateral and ventral visual cortex, in association visual cortex (Fig. S4B).
A representational similarity analysis in the Phase-scrambled objects joint-ROI showed high similarity between phase-scrambled, grid-scrambled, and scene stimuli, while the Whole-objects joint-ROI showed low similarity between those categories (Fig 4A). Figure 4B shows the results of a task-rest pattern similarity analysis based on U90 values in each joint-ROI, which support the general conclusion that task-rest pattern similarities are not necessarily stronger for more ecological stimuli. Instead, task-rest correspondences reflect stimulus preferences, which are different in high- and low-level visual cortical regions (Fig. 4B). An ANOVA with ROI-type (Whole-objects, Phase-scrambled objects) and Stimulus-type (whole-objects, grid-scrambled, phase-scrambled objects) as factors indicated that the critical interaction of ROI-type by Category (F(2,30)=14.2, p<.0001) was significant. A significant interaction was also found for a 2 x 2 sub-ANOVA restricted to the categories whole objects and phase-scrambled objects (F(1,15)=23.5, p<.0001).
Within each joint-ROI, we compared the U90 value for the preferred category vs. the two “non-preferred” categories using paired t-tests with a Holm-Bonferroni correction for the four comparisons over the two joint-ROIs. In the Phase-scrambled joint-ROI, U90 values were significantly higher for both scrambled stimulus categories than for the whole-objects category, and in the Whole-objects joint-ROI, the U90 value for the whole-object category was significantly greater than for the phase-scrambled object category. The grid-scrambled pattern, which contains both high-level and low-level features (e.g. a high density of contour terminators), showed U90 values both in early visual and higher-order visual cortex that were not distinguishable from the regions’ preferred stimulus category.
These results demonstrate a second double dissociation relating the dependence of U90 values on both the category-evoked spatial activity pattern and the joint-ROI in which similarity of the evoked pattern to spontaneous patterns was evaluated. They are consistent with the interpretation that spontaneous activity patterns in visual cortex are strongly affected by the stimulus preferences of the region, irrespective of whether those preferences favor more or less ecological categories.
U90 values correlate with activation strength
Another prediction of the representation hypothesis is that task-evoked patterns will entrain spontaneous activity patterns in the course of development and experience. Therefore, one expects a positive relationship between the magnitude of the stimulus specific response and the strength of the relationship between category specific patterns and spontaneous activity patterns (U90 values). Figure 5 shows the correlation across category between task activation magnitude and U90 values of task-rest pattern similarity for each joint-ROI. Figure S5 shows the mean activation strengths for different categories during the Task scans. The magnitude of the stimulus-evoked response in a joint-ROI was generally strongest for the preferred category (Fig. S5). Since joint-ROIs were defined from localizer scans that were independent of the task scans, this result indicates the stability of the ROI assignments.
There was a positive and significant correlation between task activation values and U90 values in each joint-ROI (Fig. 5). The greater the activation strength of a category, the greater the U90 value, a relationship that held for joint-ROIs irrespective of their stimulus preferences. This relationship also was significant when the correlation coefficient between activation strength and U90 value was computed separately for each participant, and a group 1-sample t-test was conducted. Correlation coefficients were highly significant in all joint-ROIs (Body, p=.0004, Face, p=.0047, Scene, p=.0014; Whole-object, p=.0002; Phase-scrambled, p<.0001).
Pattern-based functional connectivity at rest
FC analyses typically evaluate the correlation between the timeseries of activity for single voxels or between voxel-averaged timeseries. However, recent task-based studies have also measured the inter-regional temporal correlation of spatial activity patterns (42–44). We used a similar approach to determine whether resting fluctuations of the multi-vertex spatial pattern for a category in each constituent ROI of a joint-ROI fluctuated synchronously or independently across the constituent ROIs. For instance, we determined whether spatial patterns for scenes in regions such as PPA, TOS, and RSC, which were previously combined to form the Scene joint-ROI (Table S1), fluctuated synchronously at rest. Synchronous fluctuations would indicate temporal variations of an inter-regional brain state specific for a particular category. We computed the temporal correlation of spatial activity patterns across the constituent ROIs of the Body and Scene joint-ROIs. The Face joint-ROI was not included in these analyses, since that ROI only included two regions and one of them overlapped with a constituent body ROI. In contrast, the Scene and Body joint-ROIs contained multiple ROIs that were all disjoint.
For each body- and scene-preferring constituent ROI, separate body and scene “pattern-to-rest” correlation timeseries were computed based, respectively, on the spatial correlation of the activity pattern on each resting frame with the body-evoked activity pattern and the scene-evoked activity pattern. This procedure is illustrated in Figure S6 using data from segments of resting scans in one subject. Pattern-to-rest-correlation timeseries are shown for two scene-preferring regions (right PPA and right TOS) and two body-preferring regions (right EBA and right FBA). Each pattern-to-rest correlation timeseries shows the similarity values over time of resting spatial patterns to a category-evoked pattern in a single ROI. For example, the left lower dark blue timeseries in Figure S6 shows the similarity of the resting pattern on each frame to the body-evoked pattern in the body constituent region EBA. The left lower graph shows that the pattern-to-rest correlation timeseries for body-preferring regions (right EBA and right FBA) that were computed using body-evoked activity patterns are positively correlated. In contrast, the left upper graph shows that the pattern-to-rest correlation timeseries for scene-preferring regions (right PPA and right TOS) that were computed using body-evoked activity patterns are uncorrelated. Conversely, when pattern-to-rest correlation timeseries were computed using the scene-evoked activity pattern, the opposite results are found. Now the pattern-to-rest correlation timeseries for scene-preferring regions show positively correlated fluctuations (right upper graph), while the timeseries for body-preferring regions are weakly correlated (right lower graph).
The data from all resting scans of all subjects were analyzed and the results are summarized in Figures 6A and 6B. Figure 6A shows three resting ‘pattern-based FC’ matrices. A pattern-based ‘body’ FC matrix (leftmost matrix) was constructed by computing all pairwise inter-regional correlations between the pattern-to-rest correlation timeseries computed from body-evoked spatial patterns. Similarly, a pattern-based ‘scene’ FC matrix (middle matrix) was computed using the pattern-to-rest correlation timeseries computed using scene-evoked patterns. Qualitatively, body-preferring ROIs showed stronger positively correlated spontaneous fluctuations for body-evoked than scene-evoked patterns, and scene-preferring ROIs showed stronger positively correlated spontaneous fluctuations for scene-evoked than body-evoked patterns.
The graphs in Figure 6B summarize the pattern-based Body and Scene FC matrices by averaging the inter-regional correlations for body-preferring regions (the lower left block of each matrix in Fig. 6A outlined in blue) and scene-preferring regions (the upper right block of each matrix outlined in red). The left and middle graphs show respectively the results when pattern-to-rest timeseries were computed using body-evoked and scene-evoked spatial patterns. A 2-factor ANOVA on the mean pairwise pattern-based FC values with ROI-type (body, scene) and Category-evoked-pattern (body, scene) as factors yielded a main effect of ROI-type (F(1,15)=5.41, p=.035), reflecting the larger FC values in body-preferring regions and, critically, a significant interaction of ROI-type by Category-evoked pattern (F(1,15)=8.96, p=.009). This effect was further supported by paired t-tests of specific contrasts. When pattern-to-rest timeseries were computed using a body-evoked spatial pattern, inter-regional correlations were significantly higher in body- than scene-preferring ROIs. Conversely, when pattern-to-rest timeseries were computed using a scene-evoked spatial pattern, inter-regional correlations were significantly higher in scene- than body-preferring ROIs.
Therefore, spontaneous fluctuations of spatial patterns of activity were more strongly correlated for the spatial patterns corresponding to the regions’ more preferred stimulus. This result indicates that resting-state FC is modulated by the representational content of spontaneous activity. In addition, a paired t-test indicated that in the pattern-based Body FC matrix, the average FC was less in the scene-body block than in the body-body block of the matrix (p<.001; Fig. 6A, leftmost matrix, orange vs blue outlined blocks). Similarly, in the pattern-based Scene FC matrix, the average FC was less in the scene-body block than in the scene-scene block of the matrix (p=.035; Fig. 6A, middle matrix, gray vs red outlined blocks). Therefore, pattern-based FC was greater between regions preferring the same category than between regions preferring different categories.
A related question was whether putative body and scene representations fluctuated independently at rest. The rightmost matrix in Figure 6A shows a Preferred-Category pattern-based FC matrix in which the pattern-to-rest correlation timeseries in body-preferring regions were computed using body-evoked spatial patterns and the pattern-to-rest correlation timeseries in scene-preferring regions were computed using scene-evoked spatial patterns. Accordingly, the lower left and upper right blocks of the Preferred-Category matrix match, respectively, the lower left block of the Body pattern-based FC matrix and the upper right block of the Scene pattern-based FC matrix.
The ‘scene-body’ block of the Preferred Category matrix outlined in green is of primary interest. The correlation between scene and body regions was uniformly low under conditions in which the inter-regional correlation involved timeseries from scene and body regions that respectively indicated the fluctuations of scene- and body-evoked spatial patterns (see rightmost graph, Fig. 6B, for average correlation values for scene-body blocks. Therefore, periods in which a body-evoked pattern was maximally present in body-preferring ROIs were largely independent of periods in which a scene-evoked pattern was maximally present in scene-preferring ROIs. Paired t-tests indicated that correlations in scene-body blocks from the Preferred-Category matrix were significantly lower than the correlations from scene-scene (p=.009) and body-body region blocks (p < .0001).
Finally, a standard FC matrix (Fig. S7, left panel) was constructed by computing vertex-averaged resting timeseries for each region, followed by pairwise correlation of the regional timeseries. As in previous work (45-48), vertex-averaged FC was category specific, with stronger FC between body-preferring regions and between scene-preferring regions, than between body- and scene-preferring regions. Pattern-based FC matrices were moderately-to-strongly correlated with the vertex-averaged FC matrix. The largest correlation was with the preferred-category matrix rather than the matrices generated using a single category (body-evoked pattern, r=0.57; scene-evoked pattern, r=0.48; preferred-category, r=0.65).
Category selectivity of U90 values in constituent ROIs
Supplementary Figure 8 shows the category selectivity of U90 values for the individual constituent ROIs within a joint-ROI. For each joint-ROI, we conducted a two-factor ANOVA with Category and Constituent-ROI as factors and U90 value as the dependent measure. A main effect of Category with no interaction between Category and Constituent-ROI was observed for both the body joint-ROI (F(7,63)=2.45, p=.028) and the scene joint-ROI (F(7,63)=2.41, p=.03), indicating a consistent profile of U90 values over categories across the constituent ROIs of each joint-ROI. Nevertheless, variability in the category profiles over the constituent ROIs is evident. Since there were many fewer vertices in each constituent ROI than in the associated joint-ROI, some variability in category selectivity over constituent ROIs is expected due to noise.
Additionally, we argue in the discussion that the category selectivity of U90 values for a joint-ROI is aided by the category selectivity of the pattern-based FC between its constituent ROIs.
Discussion
The goal of the experiment was to test representational theories of spontaneous activity by determining whether in regions of human visual cortex there is a link between spatial patterns of spontaneous activity, measured using resting-state fMRI, and the spatial patterns evoked by more or less ecological visual stimuli such as bodies or phase-scrambled bodies.
We obtained two main results. First, resting spatial activity patterns in regions of visual cortex were more closely related to the activity patterns evoked by the regions’ more preferred stimulus categories. This relationship did not reflect a greater average similarity of resting patterns to the patterns for more preferred categories, but instead a greater spread of resting similarity values for more preferred categories, i.e. both larger positive and larger negative similarity values. This result was demonstrated statistically by two significant double dissociations. Body- and face-preferring regions showed larger U90 values, indexing the spread of similarity values, for faces and bodies than for scenes, while scene-preferring regions showed larger U90 values for scenes than for faces and bodies (Fig. 3). Regions preferring whole objects vs. phase-scrambled objects showed a similar significant double dissociation (Fig. 4). This result was further strengthened by the positive correlation between U90 values and stimulus specific activation magnitudes. The more strongly a stimulus activated a region, the higher the spread of spatial similarity values between the stimulus-evoked spatial pattern and the spontaneous activity pattern (Fig.5). The latter result is consistent with the notion that task-evoked patterns entrain spontaneous activity patterns in the course of development and experience, and therefore serve as a prior for task activation (Fig. 1)
The second main result was that spatial activity patterns coding for a category fluctuated more synchronously at rest between cortical regions preferring that category. For example, in the resting-state the spatial pattern evoked by bodies was more positively correlated between body-preferring ROIs than between scene-preferring ROIs. The spatial pattern coding for scenes showed the opposite result (Fig. 6). Finally, the spatial patterns coding for scenes and bodies fluctuated largely independently within the preferred regions for those categories.
To our knowledge, this is the first demonstration that spatial patterns of spontaneous activity within regions of human cortex, and fluctuations of those patterns between regions, code for stimulus and category specific information. In this respect, our work is more related to seminal work conducted in cats (7) and monkeys (8, 49) than to previous work in humans, which has focused on measurements of voxelwise functional connectivity (26, 27, 45, 46, 50-52).
Spontaneous activity patterns for objects and features in visual cortex
Spontaneous activity patterns were not more similar on average to preferred stimulus evoked-patterns, but showed the greatest variation with respect to those patterns (Fig. 3). Interestingly, animal studies thus far have found the same result with some caveats. For instance, Kenet et al. (7) recorded voltage sensitive dye imaging in anesthetized cat visual cortex and found a significant spatial correlation (r=0.6) between spontaneous activity patterns and orientation selective stimulus evoked patterns. Positive and negative values of the correlation distribution were higher (as in our experiment), rather than the mean, as compared to a control distribution obtained by flipping the orientation selective map. This finding was replicated recently in anesthetized monkey visual cortex (8). In auditory monkey cortex, spontaneous spatial covariations of gamma activity recorded with cortical grids from auditory cortex resemble tonotopic maps derived with auditory stimuli. Also in this case, the correlation involves both positive and negative high correlation values as compared to a control distribution obtained by shuffling the tonotopic map (49).
In our experiment, the control distribution was not spatially shuffled because this control might not preserve the local structure of the vascular architecture that is the anatomical basis of the measured BOLD signal. Therefore, we instead compared task-rest correlation distributions for two different stimuli (e.g. bodies vs. scenes). Although the animal and human experiments differed in many ways, in all experiments the reported match between spontaneous and task-evoked activity was not a shift in the mean, but rather a higher frequency of more extreme matches/mismatches of the spatial patterns (e.g. Compare our Fig.6A to (8) Figs.1-2).
An issue for future work is whether spontaneous activity patterns more resemble the average pattern evoked by a category or the patterns evoked by individual exemplars from the category. In rat hippocampus, spontaneous replay in anesthetized and awake animals of sequences of activity during navigation, a phenomenon qualitatively similar to what reported here (e.g. (53); see also (54)), seem to reflect individual experiences rather than averages. Hippocampal replay sequences have been mainly conceptualized as reflecting a mechanism for consolidating information in long-term memory.
We interpret our findings on spontaneous activity as constituting a prior for task processing, e.g. object recognition. An important rationale for postulating a representational function of resting activity is that limits on the information processing capacity of the brain may be mitigated by the incorporation of useful prior information. Appropriate priors will generally depend on context and therefore will change dynamically. The perceptual priors appropriate to walking alone through a forest vs. eating a family meal at the dinner table are quite different. These putative dynamic changes are thought to reflect generative models of the expected input via top-down pathways (55).
Resting scans are usually conducted under conditions in which subjects lie in a dark tube while fixating a cross in an otherwise blank display, which would not seem a fertile context for a perceptual prior. However, some aspects of an appropriate prior may not heavily depend on context. For example, recent work in monkey inferotemporal cortex indicates that individual faces can be coded by face cell assemblies whose firing rate is distributed along a small number of orthogonal dimensions (56). Therefore one possibility is that the spontaneous activity patterns found in the present work reflect fluctuations along canonical low dimensional configurations.
Synchronous fluctuations of representational content
The results for pattern-based FC show that the putative representational content of spontaneous activity, as indexed by multivertex spatial patterns, fluctuates more synchronously for the preferred category of ROIs that are linked by a common stimulus preference. Pattern-based FC between body-preferring regions was significantly larger when computed using a body- than scene-evoked pattern, while pattern-based FC between scene-preferring regions was significantly larger when computed using a scene-evoked than body-evoked activity pattern (Fig. 6).
An interesting, novel result from the pattern-based FC analysis was that at rest different putative representational states, as indexed by spatial activity patterns, fluctuated largely independently. Coherent fluctuations associated with body-evoked patterns in body-preferring ROIs occurred independently of fluctuations associated with scene-evoked patterns in scene-preferring ROIs, as shown by the very low correlations between scene and body regions in the analysis of the Preferred Category FC matrix (Fig. 6). Therefore, resting activity across category-selective regions of visual cortex cannot be described in terms of a single representational state. To our knowledge, the current results on resting pattern-based FC have not previously been reported and provide new constraints on theories of the function of FC.
Although the putative representational content of resting FC is often unspecified, an important exception comes from studies of early visual cortex, which have shown that resting FC respects the tuning of single voxels for polar angle, eccentricity, and low-level stimulus features (9–12). Most task-based studies of representation in higher-order visual and associative regions, however, have not involved measurements of voxelwise tuning functions but instead have identified task-evoked representations through measurements of regional spatial patterns. Therefore, pattern-based FC (42–44) could provide insights into the putative representational FC of spontaneous activity in high-level brain regions that are complementary to those provided by approaches based on the tuning properties of single voxels.
We suggest two additional ways in which pattern-based FC might inform studies of resting-state organization. First, pattern-based FC may help fractionate existing resting-state networks and identify the functional factors associated with that fractionation. For example, pattern-based resting FC between regions that prefer a particular category might depend on selectivity for features within the category, such as gender for face-preferring regions.
Second, pattern-based FC might uncover resting FC organizations that differ substantially from the normative whole-brain structure that has been described over the past decade (6, 15, 16, 57), although this structure does vary over individuals (58–60). In the current work, pattern-based FC was measured within category-preferring regions. Because regions that co-activate tend to show greater resting FC (61), and because previous studies have shown that regions preferring the same category show preferential FC (45, 46, 50-52), novel FC organizations were not expected. Accordingly, pattern-based FC matrices were moderately-to-strongly correlated with inter-regional vertex-averaged FC matrices. However, divergent FC organizations may be more likely in studies that use task-evoked activity patterns based on frequently occurring processes that combine different domains: for example, activity patterns based on integration of voice and face information during person-to-person interactions, visuomotor coordination during object manipulation, or biologically significant stimulus-reward or response-reward contingencies. Cross-domain pattern-based FC that cuts across standard networks might reflect synergies (62-64) or routines for implementing frequently occurring processes.
Pattern-based FC and correspondence of resting and evoked activity patterns
We suggest that the synchronous fluctuations of representational content evidenced by pattern-based FC (Fig. 6) is partly responsible for the correspondence of stimulus-evoked activity patterns and spontaneous activity patterns that was observed in joint-ROIs (Figs. 3 and 4). In the spatial analysis, the largest positive or negative similarity values for a resting frame in a joint-ROI will occur when the similarity values in the constituent ROIs on that frame are simultaneously large and have the same sign. Otherwise, across constituent ROIs the similarity values will tend to cancel or average to a lower value. Therefore, a larger spread of similarity values in a joint-ROI is more likely to be observed if the similarity values in the constituent ROIs fluctuate in a temporally correlated or coherent fashion.
This mechanism explains the pattern of U90 values across the different joint-ROIs shown in Figure 3C. U90 values averaged across categories were highest in the face joint-ROI, intermediate for the body joint-ROI, and lowest for the scene joint-ROI, i.e. were inversely related to the number of constituent regions in each joint-ROI (2 for face, 5 for body, and 9 for scenes). The greater the number of constituent ROIs in a joint-ROI, the more the overall U90 value for the joint-ROI was decreased by sub-optimal coherence. As noted in the results section, in a two-factor ANOVA on U90 values with joint-ROI (body, face, scene) and Category (8 levels) as factors, the main effect of joint-ROI was significant. We additionally computed a single U90 value for each joint-ROI for each participant by averaging over categories. Paired t-tests on these averaged U90 values, with a Bonferroni-Holm correction for multiple comparisons (3 tests), indicated significantly larger U90 values within the Face than Body joint-ROIs (p<.0001), Face than Scene joint-ROIs (p<.0001), and Body than Scene joint-ROIs (p=.01).
In addition, since the spatial activity patterns for a non-preferred category do not fluctuate as coherently across constituent ROIs as those for a preferred category (Fig. 6), the resulting similarity values across frames in the joint-ROI for that non-preferred category will show less variation from zero, resulting in smaller U90 values. Therefore, within a joint-ROI, differences in U90 values between categories should be more reliable for joint-ROIs comprised of more constituent ROIs. This suggestion is also consistent with the results in Figure 3C, with the fewest significant differences found for the face joint-ROI and the most for the scene joint-ROI. Although other factors besides the number of constituent ROIs are clearly important in determining the category selectivity of U90 values in a joint-ROI, the larger point is that the temporal coherence of preferred categories increases the incidence of extreme spatial matches and mismatches between evoked and spontaneous patterns in a category specific fashion.
Low- and high-level visual correspondences at rest
The spread of spatial similarity values between resting activity patterns and stimulus-evoked patterns was determined by how well a stimulus activated the region, irrespective of whether the stimulus was more or less ecological. In many higher-level visual ROIs, stimulus preferences favored a particular whole-stimulus category (e.g. bodies) over another whole-stimulus category (e.g. scenes) or over the phase-scrambled category. Conversely, in early visual cortex, preferences favored stimuli that weighted low-level features, resulting in larger U90 values for scrambled than whole-stimulus categories. The larger U90 values for scrambled stimuli in early visual cortex do not contradict an overall framework in which resting activity patterns reflect the statistical distribution of features in the environment. Rather, this result suggests that resting activity patterns in regions that primarily extract low-level visual features are relatively independent of the patterns associated with higher-order features/statistics that define categories of more ecological stimuli.
Grid-scrambled objects showed greater U90 values than whole-objects in the Phase-scrambled joint-ROI and equivalent U90 values to whole-objects in the Whole-object joint-ROI (Fig. 4). This latter equivalence may have reflected the fact that the union of different category-preferential regions in the Whole-object joint-ROI eliminated or reduced the importance of features selective for a particular ecological category. Instead, the U90 value reflected features common to different ecological categories that were also present in grid-scrambled objects.
Therefore, the distribution of spatial matches between resting and evoked activity patterns can be driven by a variety of stimulus features that reflect local (e.g. contour-related features) or global (e.g. faces) stimulus characteristics depending on the tested regions.
Limitations
Stimuli were not controlled for low-level variables that might have differentially activated visual regions. As noted, grid-scrambled stimuli may have included contour terminators to a greater extent than many whole-object stimuli, increasing the activation of early visual cortex. However, this factor was not explicitly controlled or manipulated. Also, stimuli were presented in a non-naturalistic context. Wilf et al. (27) have shown that in early visual cortex, resting FC patterns are better accounted for by movies than by standard retinotopic stimuli, while Strappini et al. (26) have shown that in higher-level visual cortex, resting FC patterns are better accounted for by movies than by static pictures of stimuli similar to those used here. Therefore, the present results may have underestimated the spatial correspondences between resting and evoked activity patterns.
Materials and Methods
More detailed information concerning Participants, Stimuli, Scanning Procedure, Imaging Parameters, fMRI pre-processing, and Definition of ROIs Is included in SI: Materials and Methods.
Participants
The study included 16 healthy young adult volunteers and was approved by the Institutional Review Board (IRB) of Washington University in St. Louis School of Medicine.
Stimuli
Images from seven ‘whole-object’ categories (human faces, human bodies, mammals, chairs, tools, scenes, and words), phase-scrambled images, and grid-scrambled images were presented on task scans. Word stimuli were included for exploratory analyses and results are not considered here. Phase- and grid-scrambled stimuli categories were constructed from the whole-object stimuli excluding words. Phase-scrambled stimuli preserved the spatial frequency amplitude spectrum of the whole-objects stimuli, and grid-scrambled stimuli included basic visual properties of the whole-objects images such as line segments and connectors. Localizer scans included human faces, human bodies, objects (chairs and tools), scenes, words, false font character strings and phase-scrambled images. The categories for localizer and task scans differed slightly since the former was only used to define regions of interest (ROIs).
Scanning Procedure
Subjects participated in two sessions on separate days (Fig. S1). Session one included 3 resting scans, 2 localizer scans, and 8 task scans. Session two include 2 resting scans, 2 localizer scans, 8 task scans, and 2 post-task resting scans. During resting scans participants maintained fixation on a centrally presented cross. Localizer scans contained blocks consisting of 20 presentations (300 ms duration, 700 ms interstimulus interval, ISI) of exemplars from a single category. In the task scans, exemplars from all categories were presented in random order (duration=300 ms, jittered ISI of 3.7 to 8.7 sec). In both localizer and task scans, subjects performed a minimal cognitively engaging task by pressing a button if the presented image changed its size or position.
Imaging Parameters and fMRI pre-processing
Structural and fMRI images were obtained from a Siemens 3T Prisma MRI scanner. FMRI scans involved a gradient echo-planar sequence sensitive to BOLD contrast (TE = 26.6 ms, flip angle = 58°, 2.4 x 2.4 x 2.4 mm voxels, 48 contiguous slices, TR = 1.0 s, and multiband factor of 4). FMRI data were pre-processed as described in (65).
Defining ROIs from localizer activation contrasts
ROIs were defined for each subject from univariate vertex-wise statistical contrasts of the localizer conditions. One set of contrasts isolated ROIs that preferred a particular category (face, body, or scene) relative to the object category (chairs + tools). Vertices from all ROIs that preferred a particular category (e.g. bodies) were grouped into a single ‘joint-ROI’, excluding all vertices located in early visual areas (V1 to V3) (26), as estimated from (37). A second set of contrasts identified ROIs that preferred whole-objects relative to phase-scrambled objects (face + body + scene + object > phase-scrambled) or the reverse (phase-scrambled > face + body + scene + object)(see Table S1 for ROI descriptive statistics, Fig. S3B and Fig. S4B for group-mean locations of ROIs from set1 and set2 contrasts).
Task scans: multi-voxel activation patterns
For each joint-ROI from each subject, the multi-vertex activation pattern for each stimulus category (except the word category) in the task scans was estimated via a GLM that included a category regressor for all stimulus presentations involving the category. In addition, the GLM included a target regressor for trials in which a stimulus was perturbed in size or position, and baseline and linear trend regressors for each scan. The category and target regressors were each convolved with an assumed hemodynamic response function, yielding a stimulus-evoked BOLD multi-voxel pattern for each category (e.g. the pattern outlined by the red square in Fig. S3A) and for target trials.
For each joint-ROI from each subject, a subject’s representational similarity matrix (RSM) was computed by spatially correlating the obtained categorical β weights (e.g. the average scene-evoked spatial pattern outlined by the red square in Fig. S3A) across categories. A group-averaged categorical RSM in a joint-ROI was computed by averaging all 16 subjects’ RSM within a joint-ROI with Fisher-Z transformations and reverse Fisher-Z transformations (Fig. 2B). Additionally, for each constituent ROI in each joint-ROI from each subject, the multi-vertex activation pattern for each of the stimulus exemplars (24 exemplars for each of 6 object categories) in the task scans was estimated via an exemplar-specific GLM similar to the above GLM. A group-averaged RSM for the stimulus exemplars was then computed for each constituent ROI in each joint-ROI (Fig. 2A). Finally, in order to determine the task-evoked magnitude for each stimulus category, a β weight matrix was separately computed using spatially non-normalized BOLD timeseries from the task scans.
Determining similarity of resting multi-vertex patterns and stimulus-evoked patterns
For each participant’s individual joint-ROI and the associated constituent ROIs, we determined the degree to which the multi-vertex pattern for a stimulus-evoked activity pattern for a category matched the multi-vertex pattern on each resting frame. The procedure is illustrated in Figure 3A for a single subject using real data. In the first step, as described above, the multi-vertex pattern evoked by a category in a region was determined (e.g. the ‘Scene’ activity pattern outlined by the red square in Fig. 3A, ‘Task BOLD’). Then, framewise intrinsic activity patterns were obtained from resting-state scans and the average stimulus-evoked pattern for a category was spatially correlated with the resting activity pattern on a frame (Fig. 3A, ‘Resting-state BOLD’). A high positive correlation coefficient indicates that the multi-vertex resting activity pattern on a given frame was very similar to the pattern evoked by the category (e.g. the resting frame with a ‘Scene’-like resting activity pattern outlined by the magenta square in Fig. 3A). A near zero correlation coefficient indicates that the multi-vertex resting activity pattern on a given frame was not similar to the pattern evoked by the category (e.g. the resting frame with a not-‘Scene’-like resting activity pattern outlined by the green square in Fig. 3A). Finally, a high negative correlation coefficient indicates that the multi-vertex resting activity pattern on a given frame was very similar to the inverse of the pattern evoked by the category (e.g. the resting frame with a ‘Scene’-inverted resting activity pattern outlined by the cyan square in Fig. 3A). This procedure was repeated across all resting frames, resulting in a ‘pattern-to-rest’ correlation timeseries (one correlation coefficient per resting frame) for a particular category in a particular ROI, as shown by the timeseries in Figure 3A.
From each timeseries, we constructed a corresponding distribution of correlation coefficients (Fig. 3A, distribution shown in blue). The upper 90% value of each distribution, hereafter termed the U90-value, was then determined. The U90 value computed for a category and ROI served as a measure of the relationship between resting activity patterns and the patterns evoked by a category mean. The U90-value was used as an alternative measure of variance since the U90-value refers to a spatial correlation coefficient value, which indicates the degree of pattern similarity between the task-evoked and resting state activity pattern. Using the U90-value instead of the variance of the distribution as a summary measure does not change the results of the analysis. For analyses that involved the Whole-Objects joint-ROI rather than joint-ROIs that preferred a particular object category such as faces, U90-values for the six whole-object categories (face, body, mammal, chair, tool, scene) were averaged together to form a whole-objects U90 value.
Statistical analysis of U90 values
U90 values were analyzed via repeated measures ANOVAs and post-hoc paired t-tests. For example, the statistical significance of an overall dependence of U90 values for a joint-ROI on the stimulus category was determined by conducting repeated-measures ANOVAs with Category-Type as factors. Paired t-tests were then conducted to test specific contrasts, with a Bonferroni-Holm correction for multiple comparisons.
Pattern-based resting functional connectivity
For the following pattern-based FC analysis, we used constituent ROIs from the joint-ROI that preferred a particular category (face, body, or scene) relative to the object category (chairs + tools). Since only two face constituent ROIs were found and one of those ROIs largely overlapped with a Body constituent ROI in ventral temporal cortex, Face constituent ROIs were not included in the FC analysis. Therefore, pattern-based FC was computed over 14 ROIs: 5 from the Body joint-ROI and 9 from the Scene joint-ROI.
Three pattern-based FC matrices were computed. Figure S6 illustrates the procedure for computing the cells of an FC matrix using the pattern-to-rest correlation timeseries from two scene regions (TOS, PPA) and two body regions (EBA, FBA). Figure 6A shows the resulting matrices. First, for each participant, pattern-to-rest correlation timeseries for all 14 ROIs were generated using only the body-evoked pattern for each ROI (i.e. the spatial activity pattern evoked by bodies in that ROI during the Task scans). The correlation between the timeseries for all pairings of the 14 ROIs was then computed (i.e. body-ROI-to-body-ROI, scene-ROI-to-scene-ROI, and body-ROI-to-scene-ROI pairings). For example, the leftmost graphs in Figure S6 show the correlation between TOS and PPA (top graph) and the correlation between EBA and FBA (bottom graph) using the pattern-to-rest correlation timeseries generated in each ROI from the body-evoked pattern. The correlation coefficients were then entered into the corresponding cells of the pattern-based FC matrix, “Using Body pattern-to-Rest correlation timeseries only”, shown in Figure 6A.
A similar procedure was used to generate a pattern-based FC matrix using only the scene-evoked pattern for each ROI. Pattern-to-rest correlation timeseries for all 14 ROIs were generated using only the scene-evoked pattern for each ROI (i.e. the pattern evoked by scenes in that ROI during the Task scans). Then the correlation between the pattern-to-rest correlation timeseries for all pairings of the 14 ROIs was computed. For example, the middle graphs in Figure S6 show the correlation between TOS and PPA, and between EBA and FBA using the pattern-to-rest correlation timeseries generated in each ROI using the scene-evoked pattern. Finally, the correlation coefficients were entered into the corresponding cells of the pattern-based FC matrix, “Using Scene pattern-to-rest correlation timeseries only”, shown in Figure 6A.
To generate the third pattern-based FC matrix in Figure 6A (“using Preferred Pattern-to-Rest correlation timeseries), the pattern-to-rest correlation timeseries in a Body ROI was generated using the body-evoked pattern for that ROI and the pattern-to-rest correlation timeseries in a Scene ROI was generated using the scene-evoked pattern for that ROI. Then the correlation between the correlation timeseries for all pairings of the 14 ROIs was computed and entered into the appropriate cells of the pattern-based FC matrix.
Finally, a vertex-averaged FC matrix was computed by first averaging the resting BOLD timeseries across all vertices of an ROI to generate a vertex-averaged timeseries, and then temporally correlating these averaged timeseries for all pairs of ROIs. Vertex-averaged FC matrices, which correspond to the standard regional FC matrices found in the literature, eliminate any information carried by the spatial pattern of BOLD activity within ROIs.
Pattern-based FC values were analyzed via repeated measures ANOVAs and paired t-tests. For example, we statistically evaluated whether the magnitude of pattern-based FC depended on both the category of the stimulus-evoked spatial activity pattern and the preferred category of the ROIs by conducting a repeated-measures ANOVA with the Category-evoked pattern (body, scene) and ROI-Type (body, scene) as factors. Paired t-tests were conducted to test differences between specific evoked patterns/ROI combinations. For example, pattern-based FC values for body ROIs vs. scene ROIs were compared for correlation timeseries generated using body-evoked spatial patterns.
Acknowledgements
This work was supported by the National Institutes of Health RO1 MH096482 to MC.