Large-scale analyses of functional interactions in the human brain

Abstract

A grand challenge in neuroscience is to understand how a human brain functions. Much previous research has focused on studying the activities of brain regions. However, the functional interactions between brain regions are not well understood. This dissertation focuses on computational methods to analyze data from functional magnetic resonance imaging (fMRI) scanners to study the functional interactions in the human brain.

This dissertation proposes, designs and implements efficient systems to conduct full correlation matrix analysis (FCMA) as an unbiased way to explore functional interactions in the human brain. Since a straightforward way to study FCMA would take years to complete one run with a typical neuroscience study dataset on a modern compute server, no previous attempt has been made in the past. This dissertation makes several contributions. First, we proposed and implemented parallel algorithms and optimizations on a multi-processor cluster, which improved FCMA computation by three orders of magnitude.

Second, we demonstrated that FCMA can effectively show functional interactions in the human brain by conducting a neuroscience study on an fMRI dataset. Our study successfully identified functional interactions between certain brain regions.

Third, we proposed and implemented optimization methods for FCMA for emerging many-core processors such as the IntelXeon Phi coprocessors and improved the performance of computing FCMA by another order of magnitude. On a 96-node Xeon Phi cluster, our system can finish an FCMA study with a typical dataset in minutes.

Finally, we proposed, designed and implemented a service for real-time, closed loop neuroscience studies. Our real-time FCMA can process and analyze brain volumes from multiple fMRI experiments on a 40-node compute cluster simultaneously and send the neurofeedback to each fMRI scanner over the Internet within 1.5 seconds. This system uses a novel method to improve the performance and utilization of compute nodes while meeting the real-time requirements in the presence of node failures.