Abstract
Resting-state functional magnetic resonance imaging (rs-fMRI) offers the opportunity to non-invasively study individual-specific brain networks. We propose a multi-session hierarchical Bayesian model (MS-HBM) that differentiates between intra-subject (within-subject) and inter-subject (between-subject) network variability. Across datasets, sensory-motor networks exhibited lower inter-subject, but higher intra-subject variability than association networks. Therefore, by ignoring intra-subject variability, previous individual-specific network mappings might confuse intra-subject variability for inter-subject differences. Compared with other approaches, MS-HBM cortical parcellations generalized better to new rs-fMRI and task-fMRI data from the same subjects. Importantly, MS-HBM parcellations from a single rs-fMRI session (10 min) were comparable to a recent state-of-the-art algorithm using five sessions (50 min). Individual-specific MS-HBM parcellations were highly reproducible, yet captured inter-subject differences. While other studies have already established that individual-specific networks exhibit features not observed in population-average networks, here we demonstrate that these features are behaviorally meaningful. Using kernel regression, individual differences in the spatial arrangement of cortical networks could be used to predict cognition, personality and emotion. Thus, individuals with more spatially similar parcellations exhibited more similar behavior. Overall, our results suggest that individual-specific cortical network topography might serve as a fingerprint of human behavior, orthogonal to previously proposed functional connectivity fingerprints.
Introduction
The human cerebral cortex consists of specialized areas whose complex interactions form large-scale, spatially distributed functional networks. Recent advances in non-invasive brain imaging technologies, especially fMRI (Kwong et al., 1992; Ogawa et al., 1992), provide the opportunity to map these brain networks in-vivo. One prominent tool for identifying large-scale brain networks is resting-state functional connectivity (RSFC), which reflects the synchrony of rs-fMRI signals between brain regions, while a subject is lying at rest without any goal-directed task (Biswal et al., 1995; Greicius et al. 2003; Fox and Raichle, 2007; Buckner et al., 2013).
RSFC brain networks have been shown to correspond well to task-evoked activation patterns (Seeley et al. 2007; Smith et al., 2009; Buckner et al., 2011; Cole et al., 2014; Yeo et al., 2015a; Tavor et al., 2016). RSFC is also heritable (Glahn et al. 2010; Yang et al. 2016; Ge et al., 2017), correlates with gene expression across the cortical mantle (Hawrylycz et al. 2015; Richiardi et al. 2015; Krienen et al. 2016), and predicts individual differences in behavior (Hampson et al., 2006; van den Heuvel et al., 2009; Finn et al., 2015; Rosenberg et al., 2015; Smith et al., 2015; Yeo et al., 2015b). Consequently, RSFC has been widely utilized to estimate population-level functional brain networks by averaging data across multiple subjects (Beckmann et al. 2005; Damoiseaux et al. 2006; Fox et al. 2006; Dosenbach et al. 2007; Margulies et al. 2007; Power et al. 2011; Yeo et al., 2011; Lee et al. 2012).
Population-level atlases of large-scale networks have provided important insights into the broad functional organization of the human brain. However, the fact that RSFC can be used to predict the behavior of individual subjects suggest the presence of behaviorally relevant inter-subject functional connectivity variability (Mueller et al., 2013; Finn et al., 2015; Smith et al., 2015). Furthermore, the shape, location and topology of functional brain networks vary substantially across individuals (Harrison et al., 2015; Wang et al., 2015; Gordon et al., 2017a; Gordon et al., 2017b). Therefore, the estimation of individual-specific brain networks could provide an important step towards precision medicine (Beckmann et al., 2009; Bellec et al., 2010; Zuo et al., 2010; Varoquaux et al., 2011; Hacker et al., 2013; Wig et al., 2013; Chong et al., 2017; Gordon et al., 2017c; Braga and Buckner, 2017).
Previous individual-specific network mappings only accounted for inter-subject variability, but not intra-subject variability. However, inter-subject and intra-subject RSFC variability can be quite different across regions (Mueller et al., 2013; Chen et al., 2015; Laumann et al., 2015). For example, the motor cortex exhibits high intra-subject functional connectivity variability, but low inter-subject functional connectivity variability (Laumann et al., 2015). Therefore, observed RSFC variability in the motor cortex might be incorrectly attributed to inter-subject spatial variability of brain networks, rather than just intra-subject sampling variability, resulting in sub-optimal network mapping.
Here, we proposed a multi-session hierarchical Bayesian model (MS-HBM) for deriving functional parcellations of the cerebral cortex within individual subjects. The multiple layers of the hierarchical model allowed the explicit separation of inter-subject (between-subject) and intra-subject (within-session) functional connectivity variability. By applying the MS-HBM to three multi-session rs-fMRI datasets, we confirmed that sensory-motor networks exhibited greater intra-subject, but less inter-subject variability than association networks. Importantly, compared with three other approaches, the MS-HBM parcellations generalized better to new resting and task fMRI data from the same individuals. MS-HBM parcellations estimated from a single rs-fMRI session were comparable to those generated by a recent influential algorithm using five times the data (Gordon et al., 2017a; 2017b).
Having established that the MS-HBM generated high-quality individual-specific parcellations, we further characterized their reproducibility and behavioral relevance. We found that individual-specific MS-HBM were highly reproducible, yet captured inter-subject differences. Although it has been shown that individual-specific functional networks exhibited unique features not present in group-average networks (Laumann et al., 2015; Glasser et al., 2016; Gordon et al., 2017c), their behavioral relevance is currently unknown. Extending previous works showing that inter-region functional connectivity could be an effective fingerprint of human behavior (Finn et al., 2015; Smith et al., 2015), we showed that individual differences in the spatial configuration of cortical networks could be used to predict cognition, personality and emotion. Thus, individuals with more spatially similar cortical parcellations had more similar behavior.
Methods
Overview
We proposed a multi-session hierarchical Bayesian model (MS-HBM) to estimate functional network parcellations of the cerebral cortex in individual subjects. The model distinguished between inter-subject and intra-subject network variability. Subsequent analyses proceeded in three stages. First, to examine whether inter-subject and intra-subject variability could be reliably estimated across datasets, the MS-HBM was applied to three multi-session resting-state fMRI datasets. Second, the MS-HBM was compared with three other approaches using new rs-fMRI and task-fMRI data from the same subjects. Third, we examined the reproducibility of the MS-HBM parcellations, how well the parcellations captured inter-subject differences, and whether individual differences in cortical parcellations reflected individual differences in behavior.
Multi-session fMRI datasets
The Genomic Superstruct Project (GSP) test-retest dataset (Holmes et al., 2015) consisted of structural MRI and resting-state fMRI from 69 healthy young adults (ages 18 to 35). All imaging data were collected on matched 3T Tim Trio scanners (Siemens Healthcare, Erlangen, Germany) at Harvard University and Massachusetts General Hospital using the vendor-supplied 12-channel phased-array head coil. Each participant has two sessions, acquired on two different days separated by less than 6 months. One or two rs-fMRI runs were acquired per session. Each BOLD run was acquired in 3mm isotropic resolution with a TR of 3.0 seconds and lasted for 6 minutes and 12 seconds. The structural data consisted of one 1.2mm isotropic scan for each session. Details of the data collection can be found elsewhere (Holmes et al., 2015).
The Hangzhou Normal University of the Consortium for Reliability and Reproducibility (CoRR-HNU) multi-session dataset (Zuo et al., 2014; Chen et al., 2015) consisted of structural MRI and resting-state fMRI from 30 young healthy adults (ages 20 to 30). All imaging data were collected on 3T GE Discovery MR750 using an 8-channel head coil. Each participant was scanned a total of 10 sessions across one month (one session every three days). One rs-fMRI run was collected in each session. Each fMRI run was acquired in 3.4mm isotropic resolution with a TR of 2.0 seconds and lasted for 10 minutes. The structural data consisted of one 1mm isotropic scan for each session. Details of the data collection can be found elsewhere (Zuo et al., 2014; Chen et al., 2015).
The Human Connectome Project (HCP) S900 release (Van Essen et al., 2012b; Smith et al., 2013) consisted of structural MRI, resting-state fMRI and task fMRI of 881 subjects. All imaging data were collected on a custom-made Siemens 3T Skyra scanner using a multiband sequence. Each participant has two fMRI sessions on two consecutive days. Two rs-fMRI runs were collected in each session. Each fMRI run was acquired in 2mm isotropic resolution with a TR of 0.72 seconds and lasted for 14 minutes and 33 seconds. The structural data consisted of one 0.7mm isotropic scan for each subject. Details of the data collection can be found elsewhere (Van Essen et al., 2012b; Smith et al., 2013).
Processing of GSP and CoRR-HNU data
Structural data were processed using FreeSurfer. FreeSurfer constitutes a suite of automated algorithms for reconstructing accurate surface mesh representations of the cortex from individual subjects’ T1 images (Dale et al., 1999; Fischl et al., 2001; Ségonne et al., 2007). The cortical surface meshes were then registered to a common spherical coordinate system (Fischl et al. 1999a; 1999b). The GSP subjects were processed using FreeSurfer 4.5.0 (Holmes et al., 2015), while the CoRR-HNU subjects were processed using FreeSurfer 5.3.0.
Resting-state fMRI data of GSP and CoRR-HNU were initially pre-processed with the following steps: (i) removal of first 4 frames, (ii) slice time correction with the FSL package (Jenkinson et al., 2002; Smith et al., 2004), (iii) motion correction using rigid body translation and rotation with the FSL package. The structural and functional images were aligned using boundary-based registration (Greve and Fischl 2009) using the FsFast software package (http://surfer.nmr.mgh.harvard.edu/fswiki/FsFast).
Framewise displacement (FD) and voxel-wise differentiated signal variance (DVARS) were computed using fsl_motion_outliers (Smith et al., 2004). Volumes with FD > 0. 2mm or DVARS > 50 were marked as outliers. Uncensored segments of data lasting fewer than 5 contiguous volumes were also flagged as outliers (Gordon et al., 2016). BOLD runs with more than half of the volumes flagged as outliers were removed completely. For the CoRR-HNU dataset, no session (and therefore no subject) was removed. For the GSP subjects, only one run was removed (out of a total of 222 runs). No individuals in the GSP dataset lost an entire session, and therefore, all subjects were retained.
Linear regression using multiple nuisance regressors was applied. Nuisance regressors consisted of global signal, six motion correction parameters, averaged ventricular signal, averaged white matter signal, as well as their temporal derivatives (18 regressors in total). The flagged outlier volumes were ignored during the regression procedure. The data were interpolated across censored frames using least squares spectral estimation of the values at censored frames (Power et al., 2014). Finally, a band-pass filter (0.009 Hz ≤ f ≤ 0.08 Hz) was applied.
The preprocessed fMRI data was projected onto the FreeSurfer fsaverage6 surface space (2mm vertex spacing). The projected fMRI data was smoothed using a 6mm full-width half-maximum kernel and then downsampled onto fsaverage5 surface space (4mm vertex spacing). Smoothing on the fsaverage6 surface, rather than in the volume minimized the blurring of fMRI signal across sulci.
Processing of HCP data
Details of the HCP preprocessing can be found elsewhere (HCP S900 manual; Van Essen et al. 2012b; Glasser et al. 2013; Smith et al. 2013). Of particular importance is that the rs-fMRI data has been projected to the fsLR surface space (Van Essen et al. 2012a), smoothed by 2mm and denoised with ICA-FIX (Salimi-Khorshidi et al. 2014; Griffanti et al., 2014).
However, recent studies have shown that ICA-FIX does not fully eliminate global and head-motion related artifacts (Burgess et al., 2016; Siegel et al., 2016). Therefore, further processing steps were performed on the rs-fMRI data in fsLR surface after ICA-FIX denoising, which included nuisance regression, motion censoring and interpolation, and band-pass filtering. Volumes with FD > 0.2mm or DVARS > 75, as well as uncensored segments of data lasting fewer than 5 contiguous volumes were flagged as outliers. BOLD runs with more than half the volumes flagged as outliers were completely removed. Consequently, 56 subjects were removed. Furthermore, for this work, only subjects with all four runs remaining (N = 676) were considered.
Nuisance regression utilized regressors consisting of global signal, six motion parameters, averaged ventricular signal, averaged white matter signal, and their temporal derivatives (18 regressors in total). The outlier volumes were ignored during the regression procedure. The data were interpolated across censored frames using least squares spectral estimation (Power et al., 2014). A band-pass filter (0.009 Hz ≤ f ≤ 0.08 Hz) was then applied to the data. Finally, spatial smoothing was applied by iteratively averaging the data at each surface mesh vertex with its neighbors four times.
Population-level parcellation and functional connectivity profiles
We have previously developed an approach to derive a population-level parcellation of the cerebral cortex into large-scale resting-state networks (Yeo et al., 2011). The cortical networks were defined as sets of cortical regions with similar corticocortical functional connectivity profiles. Here we applied the same approach to the GSP, CoRR-HNU and HCP datasets. Our previous analyses (Yeo et al., 2011) identified 7 and 17 networks to be particularly stable. For simplicity, we will only consider 17 networks. Details of this approach have been previously described (Yeo et al., 2011). For completeness, we briefly described its application to the current datasets.
Recall that the preprocessed fMRI data from the CoRR-HNU and GSP subjects have been projected onto the fsaverage5 surface meshes. The fsaverage5 surface meshes consisted of 18715 cortical vertices. Following previous work (Yeo et al., 2011), the connectivity profile of a cortical region (vertex) was defined to be its functional coupling to 1175 regions of interest (ROIs). The 1175 ROIs consisted of single vertices uniformly distributed across the fsaverage5 surface meshes. For each rs-fMRI run of each subject, the Pearson’s correlation between the fMRI time series at each spatial location (18715 vertices) and the 1175 ROIs were computed. The 18715 × 1175 correlation matrix were then binarized by keeping the top 10% of the correlations to obtain the final functional connectivity profiles. Outlier volumes (flagged during preprocessing) were ignored when computing the correlations.
In the case of the HCP dataset, the preprocessed fMRI data have been projected onto the fsLR surface space. The fsLR_32K surface meshes consisted of 59412 cortical vertices. We defined the connectivity profile of a cortical region (vertex) to be its functional coupling to 1483 ROIs. The 1483 ROIs consisted of single vertices uniformly across the fsLR_32K surface meshes. For each rs-fMRI run of each subject, the Pearson’s correlation between the fMRI time series at each spatial location (59412 vertices) and the 1483 ROIs were computed. The 59412 x 1483 correlation matrix were then binarized by keeping the top 10% of the correlations to obtain the final functional connectivity profile. Outlier volumes (flagged during preprocessing) were ignored when computing the correlations.
To obtain a population-level parcellation from a group of subjects, each vertex’s connectivity profiles were averaged across all BOLD runs of all subjects. The averaged connectivity profiles were clustered using a mixture of von Mises–Fisher distributions (Lashkari et al., 2010; Yeo et al., 2011). The expectation-maximization (EM) algorithm operated by first randomly assigning the vertices (18715 in the GSP and CoRR-HNU datasets, or 59412 in the HCP dataset) to different networks. The algorithm then iterated between two steps (E-step and M-step) until convergence. In the M-step, the algorithm computed a network-level connectivity profile based on vertices assigned to the same network. In the E-step, the algorithm re-assigned the network membership of vertices based on the similarity between each vertex’s connectivity profile and the network-level connectivity profile. The clustering algorithm was repeated 1000 times with different random initializations and the estimate with the best model likelihood was selected.
Multi-session hierarchical Bayesian model (MS-HBM)
The previous section described an approach to estimate a population-level parcellation from a group of subjects. Figure 1 illustrates the MS-HBM model for estimating individual-specific cerebral cortex parcellations using multi-session fMRI data. Some of the model parameters (e.g., inter-subject variability) must be estimated from a training set of subjects. A new subject (possibly from another dataset) could then be parcellated without access to the original training data. Even though the model was defined on multi-session fMRI data, an effective workaround was provided for single-session fMRI data. The exact mathematical model is found in Supplemental Methods S1. Here we provide the intuition behind this model.
Let denote the (binarized) functional connectivity profile of cortical vertex n from session t of subject s. For example, Figure 1 illustrates the binarized functional 1 1 connectivity profile for a posterior cingulate cortex vertex and a precuneus vertex from the 1st session of the 1st subject. Based on the connectivity profiles of all vertices from all sessions of a single subject, the goal is to assign a network label for each vertex of the subject. Even though a vertex’s connectivity profiles were unlikely to be the same across different fMRI sessions, the vertex’s network label was assumed to be the same across sessions.
Consistent with previous work (Yeo et al., 2011), the von Mises–Fisher mixture model was utilized to encourage brain locations with similar functional connectivity profiles to be assigned the same network label (illustrated by arrow from network label to connectivity profile in Figure 1). For example, the connectivity profiles of PCC and precuneus were very similar, so they were more likely to be grouped into the same network (i.e., default mode network or DMN).
However, unlike the group averaged connectivity profiles, the functional connectivity profiles of individual subjects are generally very noisy. If the connectivity profiles of PCC and pCun were too noisy, the mixture model might not assign them to the same network. Therefore, an additional spatial smoothness prior was incorporated. More specifically, the spatial smoothness prior V (Potts model) encouraged neighboring vertices (e.g., PCC and pCun) to be assigned to the same network.
To model inter-subject spatial variability, the spatial prior 𝛩l,n denote the probability of network l occurring at a particular spatial location n. As an example, the spatial variability map of the default network (𝛩DMN) is shown in Figure 1 (bottom left), where warm color indicates high probability and cool color indicates low probability. Both PCC and pCun had high prior probabilities of being assigned to the default network.
To model inter-subject functional connectivity variability, let denote the group-level functional connectivity profile of network l. For example, Figure 1 (top left) illustrates the group-level DMN connectivity profile . Let denote the functional connectivity profile of network l and subject s. For example, Figure 1 (top right) illustrates the DMN connectivity profiles of two different subjects ( and ). The parameter ∊l controlled how much the individual-specific network connectivity profile can deviate from the group-level network connectivity profile , and therefore represented the amount of intersubject functional connectivity variability. For example, Figure 1 illustrates the inter-subject connectivity variability ∊ for the 17 networks considered in this paper. Hotter colors indicate higher connectivity variability. The default network was colored green, which indicated an intermediate amount of inter-subject functional connectivity variability.
To model intra-subject functional connectivity variability, let denote the functional connectivity profile of network l and subject s during session t. For example, Figure 1 illustrates the default network connectivity profiles of subject 1 during sessions 1 11 12 and 2 ( and ). The parameter σl controlled how much the session-specific network connectivity profile could deviate from the individual-specific network connectivity profile , and therefore represented the amount of intra-subject variability. For example, the intra-subject functional connectivity variability σ for the 17 networks are shown in Figure 1. Hotter colors indicate higher connectivity variability. The default network was colored blue, which indicated low intra-subject functional connectivity variability.
The functional connectivity profile could be thought of as the representative connectivity profile of vertices belonging to network l of subject s during session t. However, the connectivity profiles of two regions belonging to the same network (e.g., and ) might exhibit slightly different connectivity profiles. Suppose vertex n is assigned to network l. The parameter k controlled how much the connectivity profile of vertex n from session t of subject s could deviate from the individual-specific session-specific network connectivity profile . For simplicity, k was assumed to be the same across networks and all subjects.
Given a dataset of subjects with multi-session rs-fMRI data, the group-level network connectivity profiles , the inter-subject functional connectivity variability ∊l, the intrasubject functional connectivity variability σl, the spatial smoothness prior V and the intersubject spatial variability prior 𝛩l could be estimated. Given the estimated model parameters , the parcellation of a new subject could then be inferred. Here we utilized a variational Bayes Expectation-Maximization (VBEM) algorithm to learn the model parameters from the training data and to estimate individual-specific parcellations. Details of the VBEM algorithm can be found in Supplementary Methods S2.
Although the MS-HBM was formulated for multi-session fMRI data, most studies only collect a single run of fMRI data. We considered the ad-hoc approach of splitting the single fMRI run into two and treating the resulting runs as two separate sessions. Our evaluations (see Results) suggest that this workaround worked surprisingly well.
Characterizing inter-subject and intra-subject network variability
We first evaluate whether inter-subject and intra-subject variability can be reliably estimated across datasets. For the purpose of subsequent experiments, the GSP dataset was divided into training (N = 37) and validation (N = 32) sets. The CoRR-HNU dataset (N = 30) was kept unchanged. The HCP dataset was divided into training (N = 40), validation (N = 40) and test (N = 596) sets. Furthermore, different fMRI runs within the same session were treated as data from different sessions. For example, each HCP subject underwent two fMRI sessions on two consecutive days. Within each session, there were two rs-fMRI runs. For the purpose of our analyses, we treated each HCP subject as having four sessions of data. Future work might differentiate between intra-session and inter-session variability.
The group-level parcellation algorithm was applied to the GSP training dataset. The resulting group-level parcellation was then used to initialize the estimation of the group-level network connectivity profiles , the inter-subject functional connectivity variability ∊l, the intra-subject functional connectivity variability σl, and the inter-subject spatial variability prior Θl. For this analysis, the spatial smoothness prior V was ignored. The estimated intersubject functional connectivity (∊l) and intra-subject functional connectivity (σl) variability maps, as well as the inter-subject spatial variability (𝛩) maps were visualized in Figures 2 and 3. The procedure was repeated for the CoRR-HNU dataset and HCP training set, allowing us to evaluate whether inter-subject and intra-subject variability could be reliably estimated across datasets.
Comparison with alternative approaches
Having established that inter-subject and intra-subject functional connectivity variability are indeed different across cortical networks, we tested whether the MS-HBM produced better individual-specific parcellations than three alternative approaches. The first approach was to apply the population-level parcellation (Yeo et al., 2011) to individual subjects. We will refer to this approach as “Yeo2011”. For the second approach, recall that the population-level parcellation algorithm iteratively computed a network connectivity profile based on vertices assigned to the same network (M-step) and then re-assigned the network membership of vertices based on the similarity between each vertex’s connectivity profile and the network connectivity profile (E-step). Using the network connectivity profiles from the Yeo2011 population-level parcellation, we can estimate networks in an individual subject by assigning a network label to each vertex based on the similarity between the vertex’s connectivity profile (for that subject) and the population-level network connectivity profile (i.e., E-step). Since this approach is analogous to the ICA back-projection algorithm (Calhoun et al., 2009; Beckmann et al., 2009; Filippini et al., 2009; Zuo et al., 2010; Calhoun and Adali 2012), we will refer to this second alternative approach as “YeoBackProject”. Finally, we also implemented the influential individual-parcellation algorithm of Gordon and colleagues (Gordon et al., 2017a; Gordon et al., 2017b), where the binarized functional connectivity map of each cortical vertex was matched to binarized network templates derived from the group-level parcellation. We refer to this approach as “Gordon2017”. All algorithms were applied to the CoRR-HNU dataset and the HCP test set.
In the case of the CoRR-HNU dataset, the model parameters of all algorithms were estimated from the GSP dataset and then utilized to infer the parcellations of CoRR-HNU subjects. This is especially important for the MS-HBM because inter-subject and intrasubject variability might differ across datasets, so it was important to evaluate whether model parameters estimated from one dataset could be generalized to another dataset. More specifically, the training procedure for the MS-HBM was the same as the previous section, except that the GSP validation set was also used to tune the spatial smoothness prior V. Similarly, “free” parameters in Gordon2017 were tuned using the GSP validation set.
In the case of the HCP dataset, recall that the HCP data were in a different surface space from the GSP data, so the GSP model parameters could not be applied to the HCP subjects. Instead, the model parameters of all algorithms were re-estimated from the HCP training and validation sets, and then utilized to infer the parcellation of each subject in the HCP test set.
Quantitative evaluation measures
If an individual-specific parcellation captured the system-level organization of the individual’s cerebral cortex, then each network should have homogeneous connectivity and function. Therefore, the following resting-state connectional homogeneity and task functional inhomogeneity measures were used as parcellation evaluation metrics (Gordon et al., 2016; Gordon et al., 2017c; Schaefer et al., in press):
Resting-state connectional homogeneity. Resting-state connectional homogeneity was computed by averaging the Pearson’s correlations between the resting-state fMRI time courses of all pairs of vertices within each network (Schaefer et al., in press). The average correlations are then averaged across all networks while accounting for network size: where ρl is the resting-state homogeneity of network l and |l| is the number of vertices within network l (Schaefer et al., in press). For each subject from CoRR-HNU (N = 30) and HCP test set (N = 596), we used one session to infer the individual-specific parcellation and computed the resting-state homogeneity of the individual-specific parcellation with the remaining sessions. Because the HNU dataset has the most amount of data (100 min), we also parcellated each CoRR-HNU subject using one or more fMRI sessions, and evaluated the resting-state homogeneity with the remaining sessions. This allowed us to estimate how much the various algorithms would improve with more data. When comparing between parcellations, a two-sided paired-sample t-test (dof = 29 for CoRR-HNU, dof = 595 for HCP) was performed.
Task functional inhomogeneity. The HCP task-fMRI data consisted of seven functional domains: social cognition, motor, gambling, working memory, language processing, emotional processing and relational processing, each with multiple task contrasts (Barch et al., 2013). For a given task contrast, task inhomogeneity was defined as the standard deviation of (activation) z-values within each network (Gordon et al., 2017c; Schaefer et al., in press). A lower standard deviation indicates higher functional homogeneity within the network. The standard deviations are averaged across all networks while accounting for network size: where stdl is the standard deviation of task activation z-values for network l and |l| is the number of vertices in parcel l (Gordon et al., 2017c; Schaefer et al., in press).
For each subject in the HCP test set (N = 596), the first rs-fMRI run from the first session was used to infer the individual-specific parcellation. The individual-specific parcellation was then utilized to evaluate task inhomogeneity for each task contrast (Eq. (2)) and then averaged across all contrasts within a functional domain, resulting in a single functional inhomogeneity measure per functional domain. The number of task contrasts per functional domain ranged from three for the emotion domain to eight for the working memory domain. When comparing between parcellations, the inhomogeneity metric (Eq. (2)) was averaged across all contrasts within a functional domain before a two-sided paired-sample t-test (dof = 595) was performed for each functional domain.
Characterizing the MS-HBM parcellations
Having established that the MS-HBM was better than other approaches in generating individual-specific parcellations, we further characterized the reproducibility of individual-specific MS-HBM networks using the CoRR-HNU data and HCP test set. Given that intra-subject and inter-subject network variability were different across networks, we were interested in evaluating whether intra-subject network reproducibility and inter-subject network similarity were also different across networks.
Individual-specific MS-HBM parcellations were independently inferred using the first two runs and the last two runs of the HCP test set. Therefore, there were two individual-specific parcellations for each subject based on data from two independent sets of rs-fMRI data. MS-HBM parcellations were also independently inferred using sessions 1-5 and sessions 6-10 of the CoRR-HNU dataset. Therefore, there were two individual-specific parcellations for each subject based on data from two independent sets of five sessions.
To evaluate the reproducibility of individual-specific parcellations, the Dice coefficient was computed for each network from the two parcellations of each subject. The Dice coefficients were then averaged across all networks and all subjects to provide an overall measure of intra-subject parcellation reproducibility. To evaluate inter-subject parcellation similarity, for each pair of subjects, the Dice coefficient was computed for each network. Since there were two parcellations for each subject, there were a total of four Dice coefficients for each network, which were then averaged. The Dice coefficients were then averaged across all networks and all pairs of subjects to provide an overall measure of intersubject parcellation similarity.
Behavioral relevance of individual-specific MS-HBM parcellations
Given that individual-specific functional networks exhibited unique topological features not observed in group-level networks, we further investigated whether the spatial configuration of individual-specific cortical parcellations was behaviorally meaningful. Since the HCP dataset has a rich repertoire of behavior data, we selected 58 behavioral phenotypes measuring cognition, personality and emotion (Table S1). Individual-specific MS-HBM parcellations were estimated for each HCP test subject (N = 596) using all four rs-fMRI runs. 17 subjects were excluded from further analyses because they did not have all behavioral phenotypes, resulting in a final set of 579 subjects.
Kernel regression (Murphy et al., 2012) was utilized to predict each behavioral phenotype in individual subjects. Suppose y is the behavioral measure (e.g., fluid intelligence) and l is the individual-specific parcellation of a test subject. In addition, suppose yi is the behavioral measure (e.g., fluid intelligence) and li is the individual-specific parcellation of the i-th training subject. Then kernel regression would predict the behavior of the test subject as the linear combination of the behaviors of the training subjects: y κ ∝i∊training set yi Similarity(li, l). Here, Similarity(li, I) is set to be the Dice coefficient for each network, averaged across 17 networks. Therefore, kernel regression makes the appealing assumption that subjects with more similar parcellations have similar behavioral measures.
In practice, we included a regularization term (i.e., kernel ridge regression) estimated via an inner-loop cross-validation procedure (Murphy et al., 2012). More specifically, we performed 20-fold cross-validation for each behavioral phenotype. Care was taken so that family members were not split between folds. For each test fold, inner-loop cross-validation was applied to the remaining 19 folds to determine the best regularization parameter. The optimal regularization parameter from the 19 folds was then used to predict the behavioral phenotype in the test fold. Accuracy was measured by correlating the predicted and actual behavior across all subjects within the test fold (Finn et al., 2015), resulting in 20 correlation accuracies for each behavior. To test whether the predictions were statistically better than chance, the accuracies were averaged across all behaviors and a corrected two-sided resampled t-test (dof = 19) was performed (Nadeau and Bengio, 2000; Bouckart and Frank, 2004).
Finally, we should mention that certain behavioral measures are known to correlate with motion (Siegel et al., 2016). Therefore, age, sex and motion were regressed from the behavioral data before kernel ridge regression. To prevent information from the training data to leak to the test data, for each test fold, the nuisance regression was performed on the training folds and the regression coefficients were applied to the test fold.
Code availability
Code for this work is freely available at the github repository maintained by the Computational Brain Imaging Group (https://github.com/ThomasYeoLab/CBIG). More specifically, the GSP and CoRR-HNU datasets were preprocessed using an in-house pipeline (https://github.com/ThomasYeoLab/CBIG/tree/master/stableprojects/preprocessing/CBIGfMRIPreproc2016). The group-level parcellation code (Yeo et al., 2011) are available here (https://github.com/ThomasYeoLab/CBIG/tree/master/stableprojects/brainparcellation/Yeo2011fcMRIclustering). Finally, the individual-specific parcellation code is also available (GITHUB_LINK_TO_BE_ADDED).
Results
Overview
The MS-HBM (Figure 1) was applied to three multi-session rs-fMRI datasets to ensure that inter-subject and intra-subject variability can be reliably estimated. Given that inter-subject and intra-subject variability are different across functional brain networks, a parcellation strategy might benefit from distinguishing between the two types of variability. We then tested whether the MS-HBM can produce better individual-specific parcellations than other approaches. Finally, having established that the MS-HBM produced better parcellations, we then characterized the parcellations’ reproducibility, inter-subject differences, and behavioral relevance.
Sensory-motor networks exhibit lower inter-subject, but higher intra-subject, functional connectivity variability than association networks
Figure 2A shows the 17-network population-level parcellation estimated from the HCP training set. The 17 networks were divided into eight groups (Visual, Somatomotor, Auditory, Dorsal Attention, Salience/Ventral Attention, Control, Default and TempPar), which broadly corresponded to major networks discussed in the literature. The 17 networks are referred to as “Default A”, “Default B” and so on (Figure 2A).
The HCP population-level parcellation was replicated in the GSP (Figure S1A) and CoRR-HNU (Figure S2A) datasets, although there were some interesting differences, likely due to acquisition differences. For example, the Limbic networks (A and B) from the GSP population-level parcellation (Figure S1A) were absorbed into the Default networks (A and B) in the HCP population-level parcellation (Figure 2A). Instead, there were two additional networks in the HCP population-level parcellation: Visual C and Auditory networks. The Visual C network (Figure 2A) might correspond to the foveal representation within the primary visual cortex, while the Auditory network (Figure 2A) appeared to have split off from the Somatomotor B network in the GSP population-level parcellation (Figure S1A). The higher resolution HCP data might allow the separation of the auditory and Somatomotor network B, which are in close spatial proximity.
Figure 2B shows the inter-subject functional connectivity variability map estimated from the HCP training set. Sensory-motor networks exhibited lower inter-subject functional connectivity variability than association networks. More specifically, Somatomotor (A and B) and Visual (A and B) networks were the least variable, while Salience/Ventral Attention Network B was the most variable. The results were largely consistent in the GSP (Figure S1B) and CoRR-HNU (Figure S2B) datasets, although there were some notable differences. For example, the Somatomotor B network exhibited low variability in both the GSP and HCP datasets, but intermediate variability in the CoRR-HNU dataset.
Figure 2C shows the intra-subject functional connectivity variability map estimated from the HCP training set. In general, association networks exhibited lower intra-subject functional connectivity variability than sensory-motor networks. More specifically, Default networks (A and B) were the least variable, while Somatomotor (A and B), Auditory and Visual C networks were the most variable. The results were largely consistent in the GSP (Figure S1C) and CoRR-HNU (Figure S2C) datasets, although there were some interesting differences. Of particular note is that the Visual Network B exhibited high intra-subject functional connectivity variability in the GSP dataset, but low or intermediate functional connectivity variability in the CoRR-HNU and HCP datasets.
It is worth noting that in the model (Figure 1), higher values of ∊l and σl indicate lower variability. The values in Figure 2C are much larger than Figure 2B, suggesting that intra-subject functional connectivity variability is much lower than inter-subject functional connectivity variability. These results are replicated in the GSP (Figure S1) and CoRR-HNU (Figure S2) datasets.
Sensory-motor networks are less spatially variable than association networks across subjects
The MS-HBM model differentiated between inter-subject functional connectivity and network spatial variability. Like inter-subject functional connectivity variability, the sensory-motor networks were found to be less spatially variable than association networks across subjects. For example, Figure S3 shows the inter-subject spatial variability maps of four representative networks from the HCP training set. Yellow color at a spatial location indicates that across subjects, there is a high probability of the network appearing at that spatial location, suggesting low inter-subject spatial variability. The Somatomotor network A and Visual network B showed higher probabilities (more yellow color) than the Dorsal Attention networks, suggesting that Somatomotor network A and Visual network B exhibited lower inter-subject spatial variability than Dorsal Attention networks. These results were consistent in the GSP (Figure S4) and CoRR-HNU (Figure S5) datasets.
Individual-specific networks generated by MS-HBM exhibit higher resting-state homogeneity than other approaches
Individual-specific parcellations were estimated using one rs-fMRI session from the CoRR-HNU dataset and HCP test set. The resting-state homogeneity of the parcellations were evaluated in the leave-out sessions (Figure 3A). Across both CoRR-HNU and HCP datasets, the group-level parcellation (Yeo2011) achieved the worst resting-state homogeneity, while MS-HBM performed the best. In the CoRR-HNU dataset, compared with Yeo2011, YeoBackProject and Gordon2017, the MS-HBM achieved a homogeneity improvement of 16.6% (p = 3.23e-21), 5.32% (p = 4.47e-18) and 6.88% (p = 1.23e-17) respectively. In the HCP dataset, compared with Yeo2011, YeoBackProject and Gordon2017, the MS-HBM achieved an improvement of 9.8% (p < 5e-324), 9.54% (p < 5e-324) and 5.74% (p < 5e-324) respectively. All significant p-values (i.e., p < 0.05) survived false discovery rate (q < 0.05) correction.
Individual-specific parcellations were estimated with increasing number of rs-fMRI sessions using the CoRR-HNU dataset. The resting-state homogeneity of the parcellations were evaluated in the leave-out sessions (Figure 3B). Not surprisingly, performance of the Yeo2011 group-level parcellation remained constant regardless of the amount of data. The remaining three approaches (YeoBackProject, Gordon2017 and MS-HBM) exhibited higher homogeneity with increased number of sessions. Critically, the improvement of our model over the other approaches grew with the inclusion of additional fMRI sessions. For example, as the number of sessions was increased from two to three to four to five, our approach achieved improvement of 5.44%, 5.9%, 6.13% and 6.38% respectively over Gordon2017. Interestingly, the improvement of our approach over Gordon2017 was largest when only one rs-fMRI session was utilized (6.88%). Furthermore, using just one fMRI sessions (10 min), our algorithm was able to match the homogeneity achieved with the Gordon2017 approach that used five fMRI sessions (50 min).
Individual-specific networks generated by the MS-HBM exhibit lower task functional inhomogeneity than other approaches
Individual-specific parcellations were estimated using one rs-fMRI session (15 min) from the HCP test set. Figure S6 shows the task inhomogeneity of the different approaches. Compared with Yeo2011, YeoBackProject and Gordon2017, our approach achieved a modest average improvement of 0.54% (p = 0.9 for social, p = 0.578 for motor, p < 5e-324 for other 5 domains), 1.93% (p < 5e-324 for all domains) and 0.94% (p < 5e-324 for all domains) respectively. All significant p-values (i.e., p < 0.05) survived false discovery rate (q < 0.05) correction. Interestingly, the Yeo2011 group-level parcellation performed as well as (or even better than) YeoBackProject and Gordon2017.
Individual-specific MS-HBM parcellations exhibit high intra-subject reproducibility and low inter-subject similarity
To assess intra-subject reproducibility and inter-subject similarity, our model (Figure 1) was tuned on the HCP training and validation sets, and then applied to the HCP test set. Individual-specific parcellations were generated by using the first two runs and last two runs separately for each subject. Figures 4 and S7 show the parcellations of four representative subjects. The 17 networks were present in all individual-specific parcellations, but the shapes, sizes and topologies were varied across subjects.
For example, the Default A (yellow) network exhibited a posterior temporal component for certain subjects (black arrows in Figure 4), but was missing in other subjects. As another example, the two lateral prefrontal components of the Control A (orange) network (Figure 2A) were fused into a single component in certain subjects (green arrows in Figure 4). These features were mostly replicated across sessions. Examples from the CoRR-HNU dataset are shown in Figures S8 and S9.
Figure 5A shows the across-subject spatial similarity (Dice coefficient) of individual-specific parcellations. A higher value (hot color) indicates greater inter-subject agreement. Figure 5B shows the within-subject reproducibility (Dice coefficient) of individual-specific parcellations. A higher value (hot color) indicates greater inter-session agreement within subjects. Further quantification is shown in Figure 5C, where the Dice coefficients were averaged across sub-networks.
Across all networks, intra-subject reproducibility was greater than inter-subject similarity. Compared with association networks, the Somatomotor networks (A and B) and Visual networks (A and B) were more spatially similar across subjects, but also exhibited greater within subject inter-session reproducibility. Overall, the MS-HBM parcellation model achieved 77.9% intra-subject reproducibility and 65.4% inter-subject similarity.
The results are similar in the CoRR-HNU dataset (Figure S10), although intra-subject reproducibility was higher (81.6%) and inter-subject similarity was lower (59.4%). The improvement might be the result of longer scan duration in the CoRR-HNU dataset (50 min versus 30 min).
Individual differences in cortical network parcellations can predict cognition, personality and emotion
Across all 58 behavioral measures, average prediction accuracy was r = 0.084 (p < 4e- 10). While the accuracy might seem modest, they were comparable to (if not better than) other studies using functional connectivity for behavioral prediction (HCP MegaTrawl; https://db.humanconnectome.org/megatrawl/; Noble et al., 2017; Dubois et al., biorxiv). For example, of the 58 behavioral measures, 49 of them were also utilized in the HCP MegaTrawl. For the 300-dimensional group-ICA results, HCP MegaTrawl achieved an average accuracy of r = 0.059 (original data space), while kernel regression yielded an average accuracy of r = 0.091.
Figure 6 shows the prediction accuracy for 13 cognitive measures from the NIH toolbox. Average prediction accuracy was r = 0.15 (p = 1.7e-8). The prediction accuracies for the remaining cognitive, emotion and personality measures are found in Figures S11 and S12. In the case of the NEO-5 personality scores (Figure S11), average predication accuracy was r = 0.10 (p = 0.0018). Interestingly, the prediction of emotional recognition (Figure S12) was poor with an average prediction accuracy of r = −0.036 (p = 0.21). In the case of the emotional measures (all items in Figure S12 except for emotional recognition), the average prediction accuracy was r = 0.10 (p = 5.9e-4). All significant p-values (i.e., p < 0.05) survived false discovery rate (q < 0.05) correction.
Discussion
We proposed a multi-session hierarchical Bayesian model (MS-HBM) that took into account inter-subject and intra-subject network variability. Across three multi-session datasets, we found that compared to association networks, sensory-motor networks exhibited lower inter-subject, but higher intra-subject network variability. Furthermore, in both rs-fMRI and task-fMRI data, the MS-HBM individual-specific parcellations were more homogeneous than parcellations derived with three alternative approaches. Finally, we showed that individual-specific parcellations were reproducible within individuals, while reflecting individual differences. Importantly, individual differences in the spatial arrangement of cortical networks could be used to predict individuals’ cognition, emotion and personality.
Association networks exhibit more inter-subject variability than sensory-motor networks
Over the course of primate evolution, the human association cortex underwent marked expansion, while the size of primary sensory cortices largely stayed constant (Hill et al., 2010; Preuss 2011). This rapid expansion might result in massive organizational differences between association and sensory cortices (Buckner & Krienen, 2013). Furthermore, the association cortex matures late during neurodevelopment (Hill et al., 2010; Buckner & Krienen, 2013). The prolonged exposure to environmental factors during a time of high neuroplasticity (Petanjek et al., 2011) might lead to greater individual differences in association cortical anatomy, function and connectivity. Indeed, anatomical studies have shown that early sensory-motor cortical areas (e.g., Area 17) exhibit less inter-subject spatial variability than association areas (e.g., Areas 44 and 45) after accounting for cortical folding patterns (Amunts et al., 1999; Amunts et al., 2000; Fischl et al., 2008; Yeo et al., 2010a).
The hypothesis that association regions exhibit greater inter-subject functional connectivity variability than sensory-motor regions is strongly supported by recent rs-fMRI studies (Mueller et al., 2013; Chen et al., 2015; Laumann et al., 2015). One important methodological consideration is that previous studies assumed functional correspondence across subjects after macro-anatomical alignment (Mueller et al., 2013; Chen et al., 2015; Laumann et al., 2015). However, it is well-known that macro-anatomical alignment (or even functional alignment) is not sufficient to achieve perfect functional correspondence across subjects (Fischl et al 1999b; Yeo et al., 2010b; Robinson et al., 2014; Harrison et al., 2015; Langs et al., 2016; Glasser et al., 2016). Therefore, a portion of the inter-subject functional connectivity variability observed in previous studies might be the result of functional network misalignment across subjects (also see Bijsterbosch et al., biorxiv).
By contrast, we explicitly differentiated between inter-subject spatial variability and inter-subject functional connectivity variability, allowing the possibility that for certain networks, inter-subject variability might be attributed to spatial variability, rather than functional connectivity variability. Neverthless, our results were largely in agreement with previous studies. Across three datasets, association networks exhibited higher inter-subject functional connectivity (Figures 2, S1, S2) and spatial (Figures S3 to S5) variability than sensory-motor networks. Among the association networks, the Salience/Ventral Attention network B was especially variable. Furthermore, networks with higher inter-subject functional connectivity variability also exhibited greater inter-subject spatial variability.
Sensory-motor networks exhibit more intra-subject variability than association networks
While there have been many rs-fMRI test-retest studies (Meindl et al., 2010; Wang et al., 2011; Guo et al., 2012; Zuo and Xing 2014), there are few studies focusing on the spatial topography of intra-subject functional connectivity variability (Mueller et al., 2013; Chen et al., 2015; Laumann et al., 2015). Laumann and colleagues found that sensory-motor (visual, somatosensory, motor) regions exhibited high intra-subject functional connectivity variability, while association regions exhibited low intra-subject functional connectivity variability. On the other hand, Mueller and colleagues (2013) found that low signal-to-noise regions (orbital frontal and temporal pole) exhibited high intra-subject variability, while portions of the default network exhibited low intra-subject variability. Therefore, there were agreements and discrepancies between the two studies. Like before, it is worth noting that Mueller et al. (2013) assumed functional correspondence after macro-anatomical registration, while Laumann et al. (2015) utilized a subject-specific parcellation.
By contrast, our model differentiated between intra-subject and inter-subject functional connectivity variability, as well as inter-subject network spatial variability. Our results largely agreed with Laumann et al. (2015) in that sensory-motor networks exhibited high intra-subject variability, while association networks exhibited low intra-subject variability. Default networks (A and B) were the least variable, consistent with Mueller et al. (2013). These results were replicated across three datasets, although a particularly interesting difference is that Visual B network showed high intra-subject variability in the GSP dataset, but low or intermediate intra-subject variability in the CoRR-HNU and HCP datasets. This difference might be due to the fact that subjects were told to fixate on a cross in the CoRR-HNU and HCP datasets, while subjects were told to keep their eyes open (with no fixation cross) in the GSP dataset.
An important criterion for a good biomarker is high test-retest reliability, which requires inter-subject differences to dominate intra-subject variability. One approach to reduce intra-subject variability is to increase the acquisition time (Van Dijk et al., 2010; Xu et al., 2016). However, intra-subject functional connectivity variability is detectable even when concatenating many sessions of data (~100 minutes; Anderson et al., 2011; Laumann et al., 2015; Gordon et al., 2017c). Since intra-subject variability cannot be completely removed, a better parcellation strategy might be achieved by taking into account intra-subject variability.
Individual-specific MS-HBM parcellations are more homogeneous than other approaches during resting and task states
If an individual-specific parcellation is capturing the unique network organization of a subject’s cerebral cortex, then regions within the same network should have similar resting-state time series, as well as similar activation amplitude for any given task contrast (Gordon et al., 2017c; Schaefer et al., in press). Across the CoRR-HNU and HCP datasets, individual-specific MS-HBM parcellations exhibited greater resting-state functional connectivity homogeneity than parcellations from three other approaches (Figure 3), suggesting that MS-HBM parcellations better capture the “intrinsic” organization of individuals’ cerebral cortex. Importantly, model parameters (e.g., inter-subject and intra-subject variability) estimated from the GSP dataset could improve the estimation of individual-specific parcellations in the CoRR-HNU dataset (Figure 3A). This demonstration is important because estimates of intersubject and intra-subject functional connectivity variability were similar, but not the same across datasets (Figures 2, S1,S2). Therefore, our results suggest that the MS-HBM approach can be used to parcellate individuals from new datasets (using the same preprocessing pipeline), without having to re-estimate the model parameters (e.g., inter-subject and intrasubject functional connectivity variability).
In the HCP dataset, individual-specific MS-HBM parcellations also exhibited greater task functional homogeneity than parcellations from three other approaches (Figure S6), suggesting that MS-HBM parcellations better capture the “extrinsic” organization of individuals’ cerebral cortex. Given the strong link between task fMRI and resting-state fMRI (Smith et al., 2009; Mennes et al., 2010; Cole et al., 2014; Krienen et al., 2014; Bertolero et al., 2015; Yeo et al., 2015a; Tavor et al., 2016), this might not seem surprising. However, it is worth pointing out that the group-level parcellation performed as well as, if not better than the two other individual-specific parcellation approaches (Figure S6). Furthermore, the MS-HBM approach only demonstrated (modest) improvements over the group-level parcellation in five of seven functional domains, while there was no statistical difference in the two remaining two functional domains. One explanation is that the resting-state parcellations might be too coarse to capture the finer details of task activation. For example, the right-hand motor task preferentially activates the hand region of the left somatomotor cortex. However, Somatomotor network A is bilateral and covers the hand, foot and body regions of bilateral somatomotor cortex. As such, even if individual-specific Somatomotor network A was highly accurate, the resulting task inhomogeneity might still be relatively high.
MS-HBM approach works well with single-session rs-fMRI data
As discussed in a previous section, increasing the scan duration of resting-state fMRI can improve the reliability of functional connectivity measures (Van Dijk et al., 2010; Xu et al., 2016). While earlier studies have suggested that 5 to 12 minutes of resting-state scan might be sufficient to provide reliable measurements (Van Dijk et al., 2010; Birn et al., 2013), more recent studies have suggested the need for 25 to 30 minutes of data (Anderson et al., 2011; Laumann et al., 2015; Gordon et al., 2017c). However, it is important note that the amount of data necessary for reliable measurements depends on the functional connectivity measures being computed (Gordon et al., 2017c).
Consistent with previous studies, our experiments showed that the quality of the individual-specific parcellations improved with more rs-fMRI data, although the improvements plateaued after around 30 to 40 minutes of data (Figure 3B). Importantly, even though the MS-HBM was developed for multi-session rs-fMRI, the algorithm performed well even with single-session data. For example, the individual-specific MS-HBM parcellations estimated with one rs-fMRI session (10 minutes) exhibited comparable resting-state connectional homogeneity with parcellations estimated using a recent prominent approach with five times the amount of data (Gordon et al., 2017a, 2017b).
Spatial configuration of individual-specific cortical parcellations is behaviorally meaningful
Given that inter-subject and intra-subject functional connectivity variability are different across functional brain networks, it is important for a parcellation strategy to distinguish between the two types of variability. For example, Somatomotor networks (A and B) exhibited low inter-subject, but high intra-subject, functional connectivity variability. A naïve algorithm might wrongly attribute differences in somatomotor connectivity between two subjects to inter-subject differences, rather than just within-subject (inter-session) noise.
The individual-specific parcellation approach in this paper modeled both inter-subject and intra-subject variability, allowing the identification of individual-specific functional networks that were highly reproducible within each subject, while also capturing variations across subjects (Figures 4, S8). Although all networks showed higher intra-subject reproducibility than inter-subject similarity, there were also differences across networks, with sensory-motor networks showing higher intra-subject reproducibility and higher inter-subject similarity than association networks (Figures 5, S10).
Recent work has suggested that individual-specific functional networks exhibit unique topological features not observed in group-level networks (Harrison et al., 2015; Laumann et al., 2015; Glasser et al., 2016; Langs et al., 2016; Braga & Buckner, 2017; Gordon et al., 2017a; 2017b; 2017c). This is also clearly the case with individual-specific MS-HBM parcellations (Figures 4, S8). While we have pointed out two examples (Default A and Control A networks), it is also obvious that many of these individual-specific parcellation features are replicable across sessions.
A major unanswered question in the literature is whether individual differences in cortical parcellations are actually behaviorally meaningful. Here, kernel regression was utilized to demonstrate that the spatial arrangement of individual-specific cortical networks can be used to predict behavior in individual subjects (Figures 6, S11, S12). More specifically, kernel regression models the possibility that subjects with more similar parcellations exhibited similar behavior. Successful prediction suggests that inter-subject variation in the spatial configuration of cortical networks are strongly related to inter-subject variation in behavior.
Previous works have suggested that inter-region functional connectivity can be utilized as an effective fingerprint of human intelligence (Finn et al., 2015) and a positivenegative axis of human behavior (Smith et al., 2015). Here, we showed that the spatial topography of individual-specific networks can be used to predict a wide range of behavioral measures covering cognition, personality and emotion. It would be worthwhile to investigate whether inter-subject network spatial variability and inter-subject functional connectivity variability can be combined to improve the prediction of individuals’ behavior.
Methodological considerations and future work
Although the MS-HBM approach did not account for inter-site variability, we demonstrated that model parameters estimated from one site can generalize to another site with a different acquisition protocol (Figures 3, S8 to S10). Given the increasing availability of multi-session rs-fMRI (Zuo et al., 2014; Holmes et al., 2015; Poldrack et al., 2015; Filevich et al., 2017; Gordon et al., 2017c), it might be possible to add another layer to the hierarchical model to account for inter-site variability, in addition to intra-subject and inter-subject variability. Furthermore, our experiments did not differentiate between rs-fMRI runs collected within the same session versus rs-fMRI runs collected from different sessions. Another layer could again be inserted into the model to differentiate between within-subject intra-session and within-subject inter-session variability. However, we suspect diminishing returns.
By assuming individual-specific parcellations to be the same across sessions (Figure 1), the MS-HBM essentially treats inter-session differences as noise. The implication is that the individual-specific MS-HBM parcellations seek to capture stable, trait-like network organization in individuals. However, it is well-known that certain factors (e.g., caffeine intake, sleepiness, attention) result in different brain states and thus functional network organization (Tagliazucchi and Laufs, 2014; Laumann et al., 2015; Poldrack et al., 2015; Yeo et al., 2015b; Wang et al., 2016; Shine et al., 2016). Moreover, in longitudinal studies of certain populations, e.g., Alzheimer’s Disease dementia, the goal is to detect neurological changes between consecutive sessions that are relatively far apart in time (Misra et al., 2009; Raj et al., 2015; Risacher et al., 2010; Zhang et al., 2016; Lindemer et al., 2017). To capture transient session-specific or longitudinal changes in brain network organization, the model could be modified to allow for spatial differences in individual-specific parcellations across sessions.
Here, we focused on parcellating the cerebral cortex into a small number of (less than twenty) networks. Each spatial (e.g., parietal) component of a network likely spans multiple cytoarchitectonically, functionally and connectionally distinct cortical areas (Kaas 1987; Felleman and Van Essen 1991; Amunts and Zilles 2015; Eickhoff et al., in press). It would be interesting to extend the MS-HBM to estimate a finer division of the cerebral cortex that might approximate classically defined cortical areas. The main challenge is that because of strong long-range functionally connectivity (Sepulcre et al., 2010), the MS-HBM will always result in spatially distributed networks even when estimating large number (e.g., hundreds) of networks. We are working on an additional spatial prior to ensure parcels are spatially localized, but not necessarily spatially connected (Glasser et al., 2016).
Conclusions
We developed a multi-session hierarchical Bayesian model (MS-HBM) that differentiated between inter-subject and intra-subject variability when estimating individual-specific cortical network parcellations. Across three datasets, sensory-motor networks exhibited lower inter-subject, but higher intra-subject functional connectivity variability than association networks. Sensory-motor networks were also more spatially variable across subjects than association networks. Using a single rs-fMRI session (10 min), our approach yielded parcellations comparable to those estimated by a recent template matching algorithm using five rs-fMRI sessions (50 min). Furthermore, individual-specific MS-HBM parcellations were highly reproducible within individuals, while capturing network variations across subjects. Finally, inter-subject variation in the spatial configuration of cortical networks are strongly related to inter-subject variation in behavior, suggesting their potential utility as fingerprints of human behavior.
Acknowledgement
This work was supported by Singapore MOE Tier 2 (MOE2014-T2-2-016), NUS Strategic Research (DPRT/944/09/14), NUS SOM Aspiration Fund (R185000271720), Singapore NMRC (CBRG/0088/2015), NUS YIA, the Singapore National Research Foundation (NRF) Fellowship (Class of 2017). AJH was supported by the National Institute of Mental Health (573 K01MH099232). XNZ was supported by the National Basic Research (973) Program (2015CB351702), the Natural Science Foundation of China (81471740, 81220108014), and the Beijing Municipal Science and Tech Commission (Z161100002616023, Z171100000117012). Our research also utilized resources provided by the Center for Functional Neuroimaging Technologies, P41EB015896 and instruments supported by 1S10RR023401, 1S10RR019307, and 1S10RR023043 from the Athinoula A. Martinos Center for Biomedical Imaging at the Massachusetts General Hospital. Data were also provided by the Brain Genomics Superstruct Project of Harvard University and the Massachusetts General Hospital (Principal Investigators: Randy Buckner, Joshua Roffman, and Jordan Smoller), with support from the Center for Brain Science Neuroinformatics Research Group, the Athinoula A. Martinos Center for Biomedical Imaging, and the Center for Human Genetic Research. Twenty individual investigators at Harvard and MGH generously contributed data to the overall project. Data were also provided by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University.