A Whole-Cortex Probabilistic Diffusion Tractography Connectome

Abstract

The WU-Minn Human Connectome Project (HCP) is a publicly-available dataset containing state-of-the-art structural magnetic resonance imaging (MRI), functional MRI (fMRI), and diffusion MRI (dMRI) for over a thousand healthy subjects. While the planned scope of the HCP included an anatomic connectome, resting-state fMRI (rs-fMRI) forms the bulk of the HCP's current connectomic output. We address this by presenting a full-cortex connectome derived from probabilistic diffusion tractography and organized into the HCP-MMP1.0 atlas. Probabilistic methods and large sample sizes are preferable for whole-connectome mapping as they increase the fidelity of traced low-probability connections. We find that overall, connection strengths are lognormally distributed and decay exponentially with tract length, that connectivity reasonably matches macaque histologic tracing in homologous areas, that contralateral homologs and left-lateralized language areas are hyperconnected, and that hierarchical similarity influences connectivity. We compare the dMRI connectome to existing rs-fMRI and cortico-cortico-evoked potential connectivity matrices and find that it is more similar to the latter. This work helps fulfill the promise of the HCP and will make possible comparisons between the underlying structural connectome and functional connectomes of various modalities, brain states, and clinical conditions.

Keywords: Human Connectome Project; diffusion MRI; human; structural connectome; tractography.

Figure 1.

Probabilistic diffusion tractography structural connectome of the human cortex. A, Group average (N = 1065) structural connectivity matrix consisting of the 360 HCP-MMPS1.0 atlas parcels organized into 10 functional networks. Raw streamline counts are fractionally scaled yielding the log probability F_pt. The white arrows highlight the diagonal which contains contralateral homologs. B, The first row of the connectivity matrix, showing connection probabilities from left V1 to all other parcels, projected onto the fsaverage template cortex. C, Single subject (100307) volume ray casting visualization of left V1-originating streamline probabilities within the skull-stripped T1-weighted structural MR volume. D, Ten functional networks, adapted from Ji et al. (2019), within HCP-MMPS1.0 atlas. These are indicated by red boxes in panel A.

Figure 2.

Connectivity strength exponential decays with fiber tract length. A, B, Connections within the right and left hemispheres, respectively. C, Connections between the right and left hemisphere. D, All connections. Each marker represents a pair of parcels. Red traces show the least-squares exponential fit; inset are the length constant λ and r² of this fit. Note that F_pt is log-transformed making these axes effectively semi-log.

Figure 3.

Interindividual variability. Shown are (A) the matrix of connectivity coefficients of variation (CV) across subjects (B) pairwise CV versus fiber tract length, (C) the distribution of CV across all connections, (D) the F_pt versus fiber tract length for the connections in the highest quintile of interindividual consistency, and (E) the F_pt of right hemisphere V1–V2 connection in all subjects versus left hemisphere V1–V2 connection. In panels B, D, each marker represents a sample statistic for a connection between two parcels. E, Each marker represents an individual subject. D, The red trace show the least-squares exponential fit, and inset is the length constant λ and r² of this fit. Note that F_pt is log-transformed making this panel’s axes effectively semi-log. In panel E, the r² of the least-squares linear fit is reported.

Figure 4.

Comparison of human diffusion tractography and macaque retrograde tracing connectomes. Subset of homologous parcels in the human HCP-MMPS1.0 and macaque fv91 atlas. A, Macaque group-average retrograde tracer derived structural connectome, gray indicates missing data. B, Human probabilistic diffusion tractography connectome. C, Pairwise correlation between macaque and human structural connectivity, r = 0.35, p = 0.0013.

Figure 5.

Interhemispheric connectivity. Differential connectivity between ipsilateral and contralateral connectivity. Greater ipsilateral connectivity dominates and is indicated in red. Parcel-pairs with greater contralateral connectivity than ipsilateral are blue. The green cortical patches show anatomic extent of parcel groups of notable contrast.

Figure 6.

Contralateral homologs. Differential connectivity between contralateral homologous parcels versus the mean of all other contralateral parcels. Red indicates contralateral homologous connectivity greater than mean contralateral connectivity. Note that many language-implicated regions have relatively weak connectivity with their contralateral homologs.

Figure 7.

Language/auditory network hyperconnectivity and left-lateralization. A, Distance-binned connectivity within the language and auditory networks compared with connectivity between the language and auditory networks and other networks, separately for the left and right hemispheres. B, Differential trace for the within-connectivity and between-connectivity in both hemispheres. In both panels, gray patches show Bonferroni-corrected bootstrapped 95% confidence intervals across subjects.

Figure 8.

Connectivity is influenced by the cortical hierarchy. A, B, Connectivity is strongly predicted by hierarchical similarity in some networks and modestly predicted overall. A, All connectivity versus myelination difference, including within-network and across-network connections, for the left, right, and callosal connections. For both panels, each marker represents a parcel pair. B, Within-network connectivity versus myelination difference for 10 functional networks. Linear fits and correlation coefficients computed independently for the left and right hemisphere. A negative correlation indicates that parcels at similar hierarchical levels tend to be more connected. C, D, Higher order prefrontal areas are better connected. C, Histogram of correlation coefficients between areal myelination and F_pt connectivity to each parcel. Only significant coefficients after Bonferroni correction are shown. Most coefficients are negative indicating high connectivity to low-myelination (i.e., higher-order) areas. D, Significant negative coefficients (red) map onto bilateral prefrontal cortex. Only the bilateral DVT and V6A are show positive significant correlations (blue).

Figure 9.

Probabilistic dMRI more closely resembles CCEPs than rs-fMRI. A, Connectivity matrices for probabilistic dMRI tractography, CCEP, and rs-fMRI. For CCEPs missing data has been colored gray and pre-log zero-strength connections black. B, Correlations among the three modalities. The least-squares linear fit is shown in red. C, Non-zero pairwise connection strength distributions. Note that rs-fMRI connectivity values, which are not log-transformed, display two modes, separated at 0.0014. D, Cortical parcels displaying lower (left) and higher (right) modes of rs-fMRI connectivity.

Figure 10.

Network theoretic differences between the connectivity modalities. Binarized network metrics after thresholding by edge weight (connectivity strength).