Debra Dawson
Published: 2021
Total Pages:
Get eBook
"Resting-State functional MRI (RS-fMRI) is often utilized to characterize the functional relationships between brain regions, these relationships being termed functional connectivity (FC). While whole-brain and network-specific FC have been widely explored, there is still much to be discovered, particularly with respect to FC on a fine scale. In the first part of this thesis, the fine-scaled RS-FC of the human visual cortex was explored in detail. The RS-FC was found to be retinotopically organized and partial correlation-based FC was found to reflect more than simply the distance between regions. In fact, the connectivity described by partial correlation in the visual cortex appeared to roughly follow that of the direct anatomical connectivity known to exist based on the thoroughly studied monkey visual cortex.It has been theorized that RS-FC may be driven by direct anatomical connections. There are, however, a number of computational approaches one can use to describe FC and these do not all result in the same pattern of predicted connectivity. Thus, in the second part of the thesis, 16 network models were tested in terms of their abilities to predict direct connectivity in the densely connected early visual cortex, in the context of no prior knowledge of the system of study. The evaluation of the model predictions was based on previous non-human primate tracer-injection studies. Multivariate models, partial correlation and Inverse Covariance, were successfully able to predict direct connectivity. In a novel approach, a “global signal” of the local network (Local Principal Component Regression) was regressed from network time-courses prior to computing the FC between regions. With this regression, bivariate models, correlation and Kappa, were also highly successful in predicting direct connectivity. Importantly, with a larger sample size, network predictions converged toward the inferred ground truth using common statistical significance thresholding, whereas they did not when using connection density-based thresholding. Finally, having identified computational methods that appeared promising for extracting direct connectivity from RS-fMRI data with relatively fine-scaled brain regions, we sought to test the usefulness of this approach in a more challenging context: the classification of individuals with mild cognitive impairment (MCI) using a whole-brain, data-driven approach. MCI is often a precursor state for dementia, thus, if subjects who will progress can be identified while they are only mildly impaired, more timely and targeted dementia treatment development and implementation can be pursued. A novel positive predictive value (PPV)-optimized multi-classifier machine learning pipeline was developed that could classify a subset of the subjects who did not progress to dementia with high PPV and high specificity. Three subtypes of MCI individuals who progressed to dementia and two subtypes of individuals who did not progress were characterized. MRI-based features (gray matter volume, cortical thickness, and correlation-based FC) involving occipital regions, sensorimotor regions, and medial parietal regions were most often implicated in the differences between MCI subtypes.We conclude that network modeling methods applied to RS-fMRI in the context of fine-scaled brain regions are useful to characterize retinotopically organized connectivity in the human visual cortex, can extract direct connectivity reliably in a densely connected local network, and contribute to the precise classification of a subset of stable MCI subjects"--