Studies of functional neuroimaging data have provided compelling evidence that brains function as highly organized networks. This has prompted the need for computational and statistical tools to reliably construct networks from imaging data, and subsequently to discover patterns and differences between individuals or groups, for instance between cognitively normal subjects and others exhibiting particular pathologies. The common approach to defining these networks is to compute temporal correlations between signals measured at distinct locations (voxels) in the brain, with stronger correlations corresponding to network connections. Due to temporal trends and noise from physiological and other sources that can contaminate the measured signals, the seemingly simple task of quantifying functional connectivity between brain regions in fact requires careful statistical modeling and efficient computational tools in order to draw reliable inferences related to brain networks. The primary aim of this project is to develop flexible statistical models of fMRI data that build on conventional correlation-based network construction to provide a more robust and complete picture of connectivity in the brain. The project will develop a graphical user interface for computing and visualizing connectivity properties. And the project will provide trainees with extensive international collaborative experience.
The project consists of three parts. In part 1, fMRI signals are modeled as a spatio-temporal process, where voxels within the same brain region share a common stochastic structure. In contrast to conventional methods, the model does not assume stationarity of the process over time or perform a preliminary averaging of signals from voxels in the same region. The removal of the stationarity assumption adds robustness to the method since such a property is unlikely to hold in experimental conditions. Methods from functional data analysis allow for estimation using all voxel-wise data. In part 2, a novel definition of functional connectivity is given that is parameter- and model-free. For any two brain regions, the distribution of temporal correlations across all pairs of voxels within these regions constitutes their connectivity profile, and is termed the correlation density. Methods for analyzing distributional data, including exploratory, clustering, and regression analyses, can be used to extract information from this rich representation. Additionally, network analyses can still be performed as in conventional studies by evaluating specific quantiles of the correlation density, such as the median. In part 3 of the project, validation of network construction through reliability and classification scores will be carried out on real data sets. These will include established data banks such as the Human Connectome Project, data gathered from small animals that have been anesthetized, and lesioned brains of individuals with consciousness disorders.
A companion project is being funded by the French National Research Agency (ANR).
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
