Conserved and variable architecture of human white matter connectivity

Abstract

Whole-brain network analysis of diffusion imaging tractography data is an important new tool for quantification of differential connectivity patterns across individuals and between groups. Here we investigate both the conservation of network architectural properties across methodological variation and the reproducibility of individual architecture across multiple scanning sessions. Diffusion spectrum imaging (DSI) and diffusion tensor imaging (DTI) data were both acquired in triplicate from a cohort of healthy young adults. Deterministic tractography was performed on each dataset and inter-regional connectivity matrices were then derived by applying each of three widely used whole-brain parcellation schemes over a range of spatial resolutions. Across acquisitions and preprocessing streams, anatomical brain networks were found to be sparsely connected, hierarchical, and assortative. They also displayed signatures of topo-physical interdependence such as Rentian scaling. Basic connectivity properties and several graph metrics consistently displayed high reproducibility and low variability in both DSI and DTI networks. The relative increased sensitivity of DSI to complex fiber configurations was evident in increased tract counts and network density compared with DTI. In combination, this pattern of results shows that network analysis of human white matter connectivity provides sensitive and temporally stable topological and physical estimates of individual cortical structure across multiple spatial scales.

Introduction

Recent advances in diffusion-based magnetic resonance imaging (MRI) techniques and complementary white matter tractography have made it possible to estimate the locations of anatomical highways spanning the cortex in the healthy and diseased human brain (Basser et al., 2000, Lazar et al., 2003, Behrens et al., 2003, Hagmann et al., 2003, Parker & Alexander, 2003). Collectively, these large-scale pathways form a network architecture which, far from being random or haphazard, suggests that the cortex has precise internal structure. Studying this measured architecture may therefore provide insight into how macroscopic white matter structure both constrains and facilitates healthy cognitive function (Johansen-Berg, 2009, Bandettini, 2009, Assaf & Pasternak, 2008). For example, large-scale anatomical brain networks constructed from diffusion imaging data display a clustered or community structure, where groups of brain regions are more highly connected to each other than to regions in other groups (Iturria-Medina et al., 2007, Hagmann et al., 2008, Gong et al., 2009, Zalesky et al., 2010). These clusters of highly interconnected regions are then bridged by a few important tracts, forming a topological structure which is thought to be theoretically conducive to both segregation (through clusters or modules) and integration (through connecting paths) of information processing (Sporns et al., 2000, Sporns et al., 2004). Through a complex structure–function relationship (Honey et al., 2007, Honey & Sporns, 2008), this delicate balance of segregation and integration in anatomical architecture may form the backdrop of modular cognitive function (Simon, 1962). Despite this confluence of results, it is not yet clear how similar our measured network structure is to true cortical organization, an uncertainty which is complicated by the availability of multiple different measurement streams including diffusion tensor imaging (DTI) (Basser et al., 1994, Pierpaoli et al., 1996), diffusion spectrum imaging (DSI) (Wedeen et al., 2005) and a variety of high angular resolution diffusion imaging (HARDI) based methods such as Q-ball imaging, spherical deconvolution, and PAS-MRI. In the present work, we restrict ourselves to a comparison of classical DTI and the more recently developed DSI, which has been shown to better resolve diffusion directions where white matter fibers cross (Wedeen et al., 2008).

While large-scale network descriptions have enabled an understanding of the global structure of anatomical connectivity, contemporaneous studies of single tracts or small portions of white matter have begun to elucidate the underlying forces constraining individual connectivity (Bosnell et al., 2008). In particular, it has been suggested that anatomical connectivity is not completely hard-wired for an individual's lifetime, but instead changes appreciably with age (over long time scales) (Giorgio et al., 2010, Wozniak & Lim, 2006), during development (over intermediate time scales) (Bava et al., 2010, Cascio et al., 2007), and with rehabilitation or training (over short time scales) (Scholz et al., 2009, Bosnell et al., 2008). In addition to their sensitivity to time-dependent changes, measures of white matter integrity are also modulated by disease and cognitive ability (White et al., 2008, Sexton et al., 2009, Madden et al., 2009).

Together, these two important strands of global and local interrogation suggest that it may be possible to characterize the large-scale anatomical connectivity of an individual and to map connectivity properties to individual cognitive ability, traits, and the effects of training. Theoretically, the utility of network theory in the study of individual variation lies in its increased power to capture large-scale alterations in structure as a combined result of many small-scale changes. However, before embarking on a study of behavioural or cognitive correlates of individual network properties, it is important to assess the sensitivity of network analysis to individual variation as well as its robustness to iterative measurement. Therefore, in this study we sought to answer two distinct but complementary questions: 1) What network properties are conserved across individuals and robust to changes in image acquisition and analysis methods? and 2) Do network properties have the ability to characterize individual differences that are both accurate and reproducible in large-scale cortical structure across multiple scanning sessions?

In order to address these questions, we acquired diffusion imaging scans from healthy individuals and systematically varied our processing stream to ascertain the effects of imaging acquisition (DSI/DTI), atlas, and spatial resolution on measured cortical architecture over the whole population as well as on its reproducibility in a single individual. More specifically, both diffusion tensor imaging (DTI) and diffusion spectrum imaging (DSI) scans were each acquired in triplicate from a group of healthy young adult individuals. Deterministic tractography was used to construct subject-specific networks and a whole-brain atlas was then applied to the tractography data in order to attain an inter-regional connectivity matrix; see Fig. 1 for a schematic of brain network construction. Atlases used in this work included the Automated Anatomical Labeling Atlas (AAL), the Harvard–Oxford Atlas (HO) and LONI Probabilistic Brain Atlas (LPBA40). In order to probe the effects of spatial resolution, each atlas was iteratively upsampled into twice as many regions of interest (ROIs) (‘Sub 1’), four times as many ROIs (‘Sub 2’), and eight times as many ROIs (‘Sub 3’); see Fig. 1 for a schematic of the upsampling scheme and Materials and methods for details.

Structural properties of cortical connectivity were measured on both raw connectivity matrices and binarized brain graphs. We hypothesized that while exact values of network metrics might vary from individual to individual, qualitative architectural properties would be conserved across the population and robust to effects of individual variability, diffusion scanning technique, and methodological variation. To complement topological analysis, we also tested for the existence of relationships between graph measures of connectivity and physical measures of connectivity, such as the length of connections in stereotactic space. Finally, in order to determine whether network properties can accurately measure individual differences, we used the intra-class correlation coefficient (ICC) to quantify the reproducibility of network measures across the 3 scanning sessions and its dependence on type of acquisitions (DTI and DSI), anatomical parcellation and spatial resolution. Specifically, by using the ICC we were able to test whether the variance between subjects in a given network property was larger than the variability within a subject over scanning sessions. By these means both the conserved and individually varying properties of healthy brain architecture, robust to imaging modality, atlas, and spatial scale, could be identified.