Resting-state functional magnetic resonance imaging (rs-fMRI) is a non-invasive imaging technique widely used in neuroscience to understand the functional connectivity of the human brain. Modern network neuroscience is based on imaging techniques and mathematical graph theory for modeling the rs-fMRI data to discover the functions of the human brain, as well as for the detection of brain disorders. As our contributions to network neuroscience, we designed and developed GPU-based high-performance algorithms to compute the Sparse Fast Fourier Transform (SFFT) of k-sparse signals, the implementation of the breadth-first search algorithm, as well as the computation of the widely used betweenness centrality graph metric – all of which can be used for the analysis of large brain graphs. Further, we also designed and conducted a study on a multi-factorial assessment of functional autistic brain network analysis.
Initial fMRI studies based in data collected in an imaging single site, usually had limited statistical power, due to the difficulties to obtain large amounts of data. To overcome these limitations, multi-site neuroimaging data have been extensively used in network neuroscience research, because they provide a larger sample size of rs-MRI data resulting in higher statistical power compared to data obtained for a single site. One main challenge for the use of multi-site databases is the existence of confounding effects, associated to variables resulting from imaging and population heterogeneity among different sites, which affect the performance of the machine learning analysis of this data. As a solution to these problems, we propose a comprehensive approach to maximize the classification scores of the machine learning analysis of multi-site rs-fMRI data by i) identifying the population and imaging variables producing the confounding effects, and ii) controlling these confounding effects. We propose novel techniques to produce homogeneous sub-samples of the subjects in the sites, as well as the implementation of statistical models and methods to quantify and control the confounding effects produced by the population and imaging variables of the multi-site data.
Oswaldo Artiles is a Ph.D. candidate at the Knight Foundation School of Computing & Information Sciences (KFSCIS) at Florida International University (FIU). His research focuses on the implementation of GPU-based high performance graph algorithms, as well as on the optimization of machine learning analysis of functional MRI multi-site data. He completed his PhD degree in Physics in 2017 from Florida International University. Oswaldo is the first author of four peer-reviewed publications that have appeared in IEEE Big Data, IEEE IPDPS, International Conference on Parallel Processing, and IEEE BIBM conferences proceedings.