Doctor of Philosophy (PhD)


Physics & Astronomy

Document Type



Effective connectivity, based on functional magnetic resonance imaging (fMRI) time series signals, is the quantification of how strongly brain activity in a certain source brain region contributes to brain activity in a target brain region, independent of the contributions of other source regions. Current methods to solve this problem have several limitations. They are either unable to model nonlinear relationships between source and target signals, unable to efficiently quantify time lags in source-target relationships, unable to identify time-varying relationships, or fail to account for variability in the hemodynamic response function that converts neuron activity to a measurable signal. In this dissertation we have proposed a series of deep learning methods to solve the above limitations. In Chapter 2, we have proposed a deep stacking network (DSN) architecture that characterizes conditional nonlinear effective connectivity among multiple time series while efficiently estimating time lags in those relationships. In Chapter 3, we extended the DSN architecture with adaptive convolutional kernels (ACK) to characterize time-varying nonlinear conditional effective connectivity with time-varying time lags. In Chapter 4, we applied the proposed DSN-ACK architecture to real-world task-based and resting-state fMRI from Bogalusa Heart Study participants to capture relevant aspects of brain health in a large epidemiological cohort. To overcome the key limitation that hemodynamic responses to neural events are not accounted for in effective connectivity analyses, in Chapter 5, we proposed a deep learning architecture to jointly estimate neural events and HRFs from task fMRI. The proposed deep learning methods can capture more information about brain connectivity than previous methods and the information it provides is relevant to brain health.



Committee Chair

Carmichael, Owen



Available for download on Sunday, March 08, 2026