Jones, DK, Chao, Y (2013 - 2015) Redefining brain network analyses: From macro to micro and hours to minutes. Royal Society. £11,827.
The brain operates as a distributed network, with different cortical areas interconnected by a dense complex network of white matter (WM) fibres. For centuries, the WM was regarded as an inert substance playing largely a structural role. However, it is now realised, that (to use an oft-used phrase), 'Nothing defines the function of a neuron better than its connections' (Mesulam). Thus, a complete understanding of the brain requires not just a study of its function (using, for example, functional MRI, EEG, MEG, PET), but also of its structure, and its networks in particular.
Diffusion MRI provides a non-invasive method for characterising tissue microstructure (most usually done using metrics derived from diffusion tensor imaging (DTI) and, equally importantly, for reconstructing WM pathways (termed 'tractography'). The latter has been exploited to trace out brain connections noninvasively, with a view to mapping the whole brain 'connectome. Recently, there has been increased activity in the application of Graph Theoretical Analyses (GTA) to the study of such networks. A graph comprises 'nodes' (cortical / subcortical areas) and 'edges' (connections between them), and GTA provides a number of metrics to characterize the local and global efficiency of the network. GTA has already shown that the brain network exhibits 'small word' topology, suggesting that the WM is optimally organized not only to support both specialized and integrated information processing, but also to maximise the efficiency of information transfer at a relatively low wiring cost. Important applications are emerging, where GTA has been used to highlight the impact of risk alleles on brain networks, the relationship between intelligence and network efficiency, and the relationship between cognitive decline and white matter connections in ageing. Thus, GTA of WM networks is a powerful tool. However, we identify two main issues:
1. Due to computational time and memory constraints, the number of nodes has been kept relatively small (e.g., 90 cortical parcellations). Increasing the number of nodes (e.g. 20k-100k) should provide more accurate localization of important hubs.
2. While GTA can operated on an undirected binary graph (all edges have equal weight), weighting each graph edge according to parameters that are likely to influence conduction along that edge, should provide a more complete characterization. However, to date, only 'blunt' weights have been utilised, namely the fractional anisotropy (obtained from DTI) and the probabilistic tractography score, to weight the edges. We have previously argued that neither represents a particularly useful weight, as both blend interesting and relevant microstructural information with uninteresting, irrelevant geometrical / topological sources of variance (including intra-axonal orienational dispersion, tract-length, tract-branching and tract curvature).
Prof Chao has addressed Point 1 by developing GPU/ CUDA-based implementations of graph-theory analyses, that permit networks comprising much larger numbers of nodes to be analysed with GTA than previously possible (primarily due to quantum reductions in time required to sort path lengths in the graph). Prof Jones has addressed Point 2 by developing 'tractometry' - a method for mapping meaningful biological information (such as axon density/diameter and putative markers for myelin) along specific white matter tracts. This facilitates more meaningful weighting of graph-edges. Moreover, his group has developed fully automated clustering of whole brain pathways to identify corresponding pathways across different brains.
This proposal brings the two teams together to exchange ideas and to work on the development of a low-cost, efficient tool to provide the most complete and detailed graph theoretical analysis of white matter brain networks to date.
For centuries it was thought that particular brain functions were served by distinct, tightly circumscribed regions of the cortex. However, this 'localizationist' model is now considered ersatz and a 'connectionist' model, which considers the brain as a distributed network of interconnected areas, has become widely adopted. In analogy to a road network, if the grey matter acts as the service stations, then the white matter forms the motorways connecting them.
A complete understanding of brain thus involves studying not just the function of grey matter with methods like fMRI, but also the white matter networks. Diffusion MRI tractography is a method that allows white matter pathways to be reconstructed non-invasively. Recently, tractography results have been subjected to graph theory analysis (GTA) - a mathematical technique which allows attributes of the brain networks (such as its 'efficiency') to be quantified. Studies have found differences exist in these metrics in a range of disorders and that they correlate with measures of normal cognition such as intelligence. While GTA is already providing new insights into the brain, we firmly believe that there are vast improvements to be made. Current GTA approaches largely treat all connections between networks nodes equally ('binary', either there is a connection = 1 or no connection = 0), or use naïve weightings to define a 'strong' or 'weak' connection. This is the first severe limitation and ignores recent developments in quantitative MRI techniques. Over the last 15 years, we have developed methods to quantify specific aspects of the make-up of the white matter, such as the diameter and density of the axonal bundles, and the amount of insulation, called myelin.
We wish to explore how combining these more tissue-specific metrics into GTA will enhance our understanding of how the brain is wired-up. However, there is huge overhead in terms of computational requirements, which means that on a standard PC, processing times are prohibitive – and the number of questions that researchers can ask of the data is necessarily constrained
We will address this second limitation by capitalizing on rapid developments in computer hardware (largely driven by the computer gaming industry), to accelerate GTA. Rather than operate on the central processing unit (CPU), we will exploit the enhanced power of the graphical processing unit (GPU), to significantly accelerate processing time and facilitate network analyses on a much larger scale than previously possible
Why have we chosen to work in this area? The brain is the last frontier in neuroscience and understanding its function, what goes wrong in disease, and even what underpins individual differences in our abilities to perform different tasks, has to be one of the most fascinating puzzles of all time. The proposed work is particularly exciting as it has the potential to open new vistas on the brain – and to explore, quite literally, how we are 'wired up'
Brain disorders present a huge burden to society and our proposal will add a vital tool in the war on mental health. We can study the influence of genes on brain networks and perhaps gain insight into the heritability of certain psychological/psychiatric disorders. Our method should enable more complete characterization of the impact of interventions (e.g. pharmacological / physiotherapy/ rehabilitation strategies), and in turn, may help to refine them to improve patient treatment.