Learning Spatio-Temporal Dynamics from Neuroimaging Datasets

Noninvasive brain recordings utilizing Magnetic Resonance Imaging (MRI) have significantly increased the availability of neuroimaging data across a large number of subjects. Neuroimaging studies often include multiple samples of data captured at varying temporal scales—ranging from months in the case of structural MRI to seconds for functional MRI. This breadth of data offers a unique opportunity to analyze the spatio-temporal dynamics of the human brain through data-driven methods. Recent advancements in deep learning have facilitated the learning of subject-specific dynamics directly from these data sets. However, several challenges persist: (1) neuroimaging data is inherently noisy and heterogeneous, (2) most neuroimaging studies provide only a limited number of observations per subject, which hampers the ability to learn robust spatio-temporal dynamics, (3) the high dimensionality of neuroimaging data complicates the analysis, and (4) many deep learning approaches lack interpretability, which is critical in computational neuroscience community for understanding novel brain mechanisms. Structural neuroimaging data is characterized by slow temporal dynamics, often constrained by the biological constraints inherent in brain structure. In this thesis, I explore how these biological constraints can be leveraged to denoise the data and train parameter-efficient neural network models that are capable of learning from limited longitudinal data sets. Conversely, functional neuroimaging data often exhibit rapid transitions. To address this, I propose a dynamical systems perspective of the brain, demonstrating how the representational learning capabilities of deep learning can be integrated with theories from the dynamical systems community. This integration facilitates the learning of robust spatio-temporal patterns from high-dimensional functional neuroimaging data, despite the numerous challenges previously outlined. These methodologies have been implemented across diverse neuroimaging datasets to explore a range of applications within neuroscience. Specifically, my research has focused on tracing the progression of Alzheimer's disease through the brain’s structural connections. Additionally, I have analyzed spatio-temporal patterns associated with fundamental brain functions, including cognition, memory, emotion, general intelligence, and other cognitive tasks, as well as the brain's resting state.