Computational Methods for Neuroscience Discovery with Neuroimaging Data
Open Access
- Author:
- Liu, Yikang
- Graduate Program:
- Biomedical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- December 11, 2020
- Committee Members:
- Nanyin Zhang, Dissertation Advisor/Co-Advisor
Nanyin Zhang, Committee Chair/Co-Chair
Xiao Liu, Committee Member
William Evan Higgins, Committee Member
Kevin Douglas Alloway, Outside Member
Daniel J Hayes, Program Head/Chair - Keywords:
- rsfMRI
rat
computational methods
data analysis
deep learning
high-dimensional data mining
image segmentation
systems neuroscience
rsfMRI denoising
spatiotemporal dynamics - Abstract:
- Resting-state functional magnetic resonance imaging (rsfMRI) is a non-invasive method to study brain function and organization. The last decades have seen a dramatic growth of human rsfMRI studies, public datasets, and data analysis software, which advanced our understanding of neuroscience and brain diseases. However, studies and resources of rodent rsfMRI are still lacking, despite its essential role in translational research. In this dissertation, we first present an open database of rsfMRI data collected from 90 awake rats with a well-established awake imaging paradigm that avoids anesthesia interference, together with a preprocessing pipeline optimized for rat data. Based on this dataset, we propose two methods termed SHERM and fastClean towards automated preprocessing of rodent rsfMRI data. First, SHERM targets rodent brain extraction, which is an essential step to aid with rsfMRI image registration. Current methods usually require manual adjustment of input parameters due to widely different image qualities and/or contrasts. SHERM, however, only requires a brain template mask as the input and is shown to automatically and reliably extract the brain tissue in both rat and mouse MRI images. fastClean is an unsupervised deep learning method that removes rsfMRI artifacts induced by the scanner, head motion, and non-neural physiological noise. Existing machine learning methods either perform unsatisfactorily in low-dimensional rodent data or suffer from long online training. With an efficient network architecture and meta-learning techniques, fastClean generates equivalently clean or cleaner data in minutes on both rodent and human datasets. Finally, we systematically investigated the spatiotemporal dynamics of spontaneous brain activity in the awake rat brain with a graph-based data mining method. We found that brain activity traverse among multiple resting-state functional connectivity patterns with nonrandom and reproducible sequential orders and time delays, revealed a network structure of these transition paths, and showed prominent brain regions involved and their temporal evolutions during the propagation of spontaneous brain activity. Taken together, this dissertation presents multiple computational methods for rsfMRI studies, demonstrates their contribution to automatic data preprocessing, data cleaning, and spatiotemporal neural pattern discovery, and advances our understanding of network organization and dynamics of the awake rat brain.