The current analyses were conducted using the same dataset reported in our previous article (Figley et al., 2015). Briefly, the sample included 32 neurologically healthy volunteers (16 males and 16 females; age = 29.9 ± 10.7 years), with no self-reported history of neurological injury/disease, psychiatric illness, or substance abuse. The Johns Hopkins University Institutional Review Board approved the study; and all participants, who were financially compensated for their participation, provided written informed consent prior to enrollment.

All data acquisition and analysis methods have been thoroughly described in our previous article (Figley et al., 2015), and thus are only briefly outlined here. All MRI data were acquired on a 3T Philips Achieva system equipped with a 32-channel SENSE head coil (Philips Healthcare, Best, Netherlands). T₁-weighted anatomical images were obtained using a 3D MP-RAGE pulse sequence (TR = 7.93 ms; TE = 3.66 ms; Flip Angle = 8°; SENSE Factor = 2.4; FOV = 212 mm × 150 mm × 172 mm; Spatial Resolution = 1.00 mm × 1.00 mm × 1.00 mm). Both T₂-weighted (TR = 4162 ms; TE = 80 ms; Flip Angle = 90°; SENSE Factor = 2; FOV = 212 mm × 154 mm × 212 mm; Spatial Resolution = 1.10 mm × 1.10 mm × 2.20 mm) and T₂-weighted Fluid Attenuated Inversion Recovery (TR = 11,000 ms; TI = 2800 ms; TE = 120 ms; Refocusing Angle = 120°; SENSE Factor = 1.75; FOV = 230 mm × 149 mm × 184 mm; Spatial Resolution = 1.00 mm × 1.20 mm × 5.00 mm) images were also acquired and assessed by a board-certified radiologist to confirm that none of the participants had structural brain abnormalities or pathologies. Finally, a spin-echo echo-planar imaging sequence was used to acquire diffusion-weighted data (number of diffusion-encoding gradients = 30; b-value = 700 s/mm²; number of reference images without diffusion-weighting = 5 (b-value = 0 s/mm²); TR = 6904 ms; TE = 69 ms; Flip Angle = 90°; SENSE Factor = 2.5; FOV = 212 mm × 212 mm; Acquired Matrix Dimensions = 96 × 96; Reconstructed Matrix Dimensions = 256 × 256; Number of Transverse Slices = 70 (no gap); Slice Thickness = 2.2 mm).

Images for each participant were processed using a multi-stage analysis pipeline (see Figure 1 from Figley et al., 2015) to: (1) coregister the diffusion-weighted and mean b = 0 s/mm² images; (2) correct for motion and eddy current distortions; (3) reorient the gradient direction for each diffusion-weighted image; (4) generate the six tensor images (Landman et al., 2007); (5) skull-strip the coregistered mean b = 0 s/mm² image (and apply the mask to the six tensor images); (6) resample all of the skull-stripped images to 1.0 mm³ resolution; (7) normalize the data to the “JHU_MNI_SS_b0_ss” template (Mori et al., 2008) in Montreal Neurological Institute (MNI) space (Mazziotta et al., 1995) using high-dimensional, nonlinear warping (Beg et al., 2005) with cascading degrees of nonlinearity (Ceritoglu et al., 2009); and (8) perform whole-brain deterministic tractography in DTIStudio (Jiang et al., 2006) using the Fiber Association by Continuous Tracking (FACT) algorithm (FA > 0.15 and Angle < 50°) and an exhaustive search approach (Mori et al., 1999; Xue et al., 1999). Using a previously reported functional connectivity atlas (Shirer et al., 2012)³, a multi-ROI approach was used to constrain the whole-brain tractography data by isolating the streamlines between each pair of functionally-defined nodes (i.e., all possible connections) within the AN, BGN, LN, PN, SMN, PVN, HVN and VSN (Figure 1 and Supplementary Videos 1–8).

Since the AN consists of three nodes (3 ROI-to-ROI combinations), the BGN consists of five nodes (10 ROI-to-ROI combinations), the LN consists of seven nodes (21 ROI-to-ROI combinations), the PN consists of four nodes (6 ROI-to-ROI combinations), the SMN consists of six nodes (15 ROI-to-ROI combinations), the PVN consists of two nodes (1 ROI-to-ROI combination), the HVN consists of two nodes (1 ROI-to-ROI combination) and the VSN consists of 11 nodes (55 ROI-to-ROI combinations), a total of 112 ROI-to-ROI combinations were assessed for each of the 32 participants—yielding a total of 3584 tractography analyses. For each of these, data were visually inspected to identify participants for whom streamlines were present, and all resulting streamlines were saved as binary maps in MNI space. Group probability maps (aka, “probabilistic connectomes”) were then generated for each of the 112 functionally-defined tracts for which streamlines were identified in at least 8/32 participants. This was achieved by adding together the binary maps for each participant (i.e., for a given ROI-to-ROI connection) and dividing by the number of participants. Image intensities for each of the probabilistic connectomes therefore range between 0 and 1 (i.e., where no participants or all 32 participants had streamlines, respectively).

The volume of white matter associated with each resting state network was then calculated by adding all of the group probability maps together for each tract and then multiplying the number of voxels with intensity >0 by the voxel size (i.e., 1 mm³). After creating binary masks of each overall connectome (i.e., a combination of all the functionally-defined tracts within each network), we then calculated the amount of spatial overlap between the white matter regions assigned to each network and report these in terms of both actual volumes and normalized ratios (relative to the size of each network), as previously reported (Figley et al., 2015). These results therefore indicate the amount of spatial overlap between a given network and each of the other networks.

Finally, 3D renderings of both the nodes within each network and the resulting white matter group probability maps were overlaid on the JHU_MNI_SS template (Mori et al., 2008) using the Volume and Volume Rendering tools within 3D Slicer (Brigham and Women’s Hospital, Boston, MA, USA; Fedorov et al., 2012)⁴, as previously reported (Figley et al., 2015).

Although several anatomically-defined white matter atlases, such as the JHU “Adam” and “Eve” atlases (Oishi et al., 2008, 2009), are already freely available, our group has taken a different approach by creating functionally-defined white matter atlases for various resting-state brain networks. Our previous article (Figley et al., 2015) systematically mapped white matter regions underlying the dorsal and ventral Default Mode Networks (dDMN and vDMN), left and right Executive Control Networks (lECN and rECN) and anterior and posterior Salience Networks (aSN and pSN). The current study now expands on that effort by including comprehensive white matter maps of the Basal Ganglia Network (BGN), Language Network (LN), Precuneus Network (PN), Sensorimotor Network (SMN), Higher Visual Network (HVN) and Visuospatial Network (VSN).

Based on our current understanding of the brain and how it is organized into distributed functional networks, we anticipate that these atlases will prove to be useful tools—in concert with quantitative white matter imaging methods like diffusion, magnetization transfer, and/or myelin water imaging—for future studies examining how structural connectivity differences between individuals or groups relate to task performance, clinical outcomes, etc. Our previous article (Figley et al., 2015) demonstrated how these functionally-defined white matter atlases can be used for voxel-wise and/or ROI-based analyses to examine relationships between structural measures throughout functionally-defined tracts or networks, and the initial white matter atlases have already proven useful for examining network-specific structural differences related to body composition (Figley et al., 2016) and Multiple Sclerosis (Ma et al., 2017). However, it is our hope that the addition of these new networks—related to different domains (e.g., language, vision, etc.)—will enhance the overall utility of the UManitoba-JHU Functionally-Defined Human White Matter Atlas.

Although the functionally-defined tracts identified in the current study will correspond to varying degrees with anatomically-defined white matter structures, an exhaustive comparison between anatomically- and functionally-defined tracts goes beyond the scope of the current manuscript. However, in order to illustrate how these functionally-defined atlases differ, we compared the language network LN1_LN4 connection (4068 mm³) from the current study to the left arcuate fasciculus (13,997 mm³), which was previously identified using the same data and analysis methods, but with anatomically-defined cortical ROIs (Figley et al., 2015). In addition to occupying only 29% of the total white matter volume of the anatomically-defined tract, we found that only 52% of voxels in the LN1_LN4 atlas overlapped with the arcuate fasciculus. Therefore, while approximately half of the voxels in the LN1_LN4 connection appear to be a subset of the nominal arcuate fasciculus, the other half are anatomically distinct.

Given that the PVN and AN represent two of the primary physical senses, and might therefore be assumed to have robust white matter connections, it seems surprising at first glance that no reproducible connections were identified. However, in the case of the PVN at least, the answer is actually quite simple. It turned out that although the resting state functional connectivity atlas for the PVN (Shirer et al., 2012)⁶ supposedly contained two nodes, one of the masks was actually an empty set (i.e., contained no voxels with a value >0), rendering it impossible for any diffusion streamlines to run between the two ROIs.

Unfortunately, there is not such a straightforward explanation for the lack of reproducible connections identified within the AN. Indeed, another recent study was able to identify white matter connections between bilateral Heschl’s gyri (Andoh et al., 2015) using slightly more sophisticated diffusion imaging parameters and probabilistic tractography. Therefore, the most likely explanation has to do with methodological differences, and each method has certain advantages and disadvantages regarding sensitivity vs. specificity tradeoffs (please see Study Limitations below for a more detailed discussion).

Finding that the BGN and SMN show high degrees of bilateral symmetry (Figure 2) is perhaps not surprising given that the cortical nodes for both networks were bilaterally distributed and both networks have prominent roles in motor function and coordination—which include balanced control of both hemispheres/sides of the body. Our finding of a strongly left-lateralized language white matter connectome is also not surprising given that language is traditionally thought of as a left-lateralized function (Knecht et al., 2000), along with the notion that left-hemisphere regions tend to interact predominantly within the same hemisphere (Gotts et al., 2013). On the other hand, the predominantly left-lateralized VSN connectome is somewhat surprising, given that visuospatial processing has previously been thought to be mostly right-lateralized (Thiebaut de Schotten et al., 2011), although this can perhaps be partially explained by the theory that right-hemisphere regions and functions tend to be less connected within hemisphere due to strong interactions between hemispheres (Gotts et al., 2013).

One of the main observations from the network overlap analysis (Figure 4) was that there was substantial overlap between certain pairs of connectomes. Nonetheless, although certain white matter regions have been ascribed to multiple networks, many of these are consistent with their expected topologies. For example, since the Precuneus is one of the central nodes within the DMN, it is perhaps not surprising that the PN overlapped substantially (33% and 43%) with the dDMN and vDMN, respectively. Given that both the LN and lECN were highly left lateralized, it is also not surprising that a large proportion (26%) of the LN overlapped with the lECN. Finally, the BGN connectome—which overlapped with 44% of the pSN, 31% of the rECN, 31% of the SMN, 27% of the aSN, 22% of the PN and 20% of the dDMN—turned out to be the third largest by volume (Figure 3), and occupied substantial portions of central white matter structures. Therefore, it is not surprising that the BGN consistently overlapped with other networks, since any long-range projections through the same central white matter regions in those networks would quite likely lead to overlap (especially when examined at relatively poor spatial resolutions, such as those that are achievable with diffusion-based MRI approaches). Moreover, the BGN may have overlapping functional roles with many of the other networks, as the basal ganglia are known to be involved in motor function (SMN), executive function (lECN and rECN) and emotional regulation (aSN and pSN; Lanciego et al., 2012).

A point that was highlighted in our previous article (Figley et al., 2015), and one that bears repeating here, is that any future applications of these atlases for ROI or along-tract analyses will require either: (1) participant data (i.e., individual quantitative images such as FA maps, etc.) to be accurately warped and coregistered to the same template as our atlases (i.e., the SPM or MRIStudio MNI templates, which are distributed with our atlases for convenience); or (2) our functionally-defined white matter atlases to be accurately warped and coregistered to each individual participant’s native space. Importantly, previous analyses have shown that linear normalization is not sufficient to accurately align subcortical regions, including central white matter structures (Figley et al., 2015); therefore, in order for images to be “accurately warped”, high-dimensional, non-linear spatial normalization—e.g., using Advanced Normalization Tools (ANTs)⁷, Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL)⁸, fMRIB’s Nonlinear Image Registration Tool (FNIRT)⁹, Large Deformation Diffeomorphic Metric Mapping (LDDMM)¹⁰, etc. (see Klein et al., 2009)—is absolutely necessary.

A very detailed discussion of the pros, cons and limitations of the general methodology employed here was published in our original article (Figley et al., 2015), which is freely available (and in fact part of the same Frontiers Research Topic). Therefore, because the current manuscript used the same dataset and methods, we will not replicate that here in full. Instead, we will briefly highlight some of the main study limitations, and extrapolate on a few points based on new literature.