Predictive analysis and deep learning of functional MRI in Alzheimer's disease

BME Ph.D. Defense: Michelle Karker

WHERE: Virtual

WHEN: September 26, 2022 10:00 am-11:00 amADD TO CALENDAR

Predictive analysis and deep learning of functional MRI in Alzheimer's disease: BME Ph.D. Defense: Michelle Karker

Alzheimer's disease (AD) and dementia pose a significant burden to individuals and public health. AD is expected to grow in prevalence in the coming decades due to the aging population. Brain atrophy is a major component of AD pathology and can occur before symptoms of cognitive impairment. However, pathological brain atrophy and symptoms of cognitive impairment may be a result of many years of disease impacts. Evidence supports the need for early detection of impacted neurocircuitry to foresee future progression to advanced stages of AD and develop treatments. This dissertation examines predictive modeling and deep learning methods to identify brain-behavior relationships and learn low-dimensional representations of brain activity from MR imaging data. The dissertation and methods are separated into four parts.  

Part one of this work examines multivariate analysis approaches applied to functional connectivity from subjects with an early clinical phenotype of AD, mild cognitive impairment (MCI). A regression framework using partial least squares and feature selection demonstrated significant brain-behavior relationships with measures of cognition and memory. The results also confirm other findings that ecologically relevant task-based connectivity serves as a ``stress-test" for memory-related deficits such as those observed in MCI. This approach elucidated brain regions that may be implicated in MCI and warrant future study (superior temporal gyrus, inferior parietal lobule, and superior frontal gyrus). Part two extends the multivariate analyses studied in part one to an additional brain imaging modality, arterial spin labeling (ASL). Cerebral blood flow (CBF) as measured by ASL demonstrated brain-behavior relationships with composite measures of memory and learning in a cohort along the spectrum of AD, demonstrating that CBF data warrant further investigation as a predictor in this application.

Parts three and four utilize a variational autoencoder (VAE) model, a deep learning approach to encode latent representations that aim to disentangle sources of fMRI signal. A surface-based VAE trained on only healthy controls is shown to be generalizable to patients with known AD pathology. The results maintained individual separation and high input/decoder output spatial reconstruction correlation of r=0.8 across all three groups. Part four extended the surface-based model used in part three to a volumetric fMRI approach. Similarly to the surface-based model, high reconstruction accuracy (NRMSE=0.68) and temporal correlation (r=0.8) between input and decoder output are demonstrated. This approach is more readily applicable to 3D fMRI data as compared to the surface-based model. 

In summary, this work has proposed and developed multivariate and deep learning analysis techniques for brain imaging data in the context of AD with the ultimate goal of improving detection and intervention for early pathological changes in the brain.

Zoom Link:
*Registration Required

Committee Chair(s):
Dr. Scott J. Peltier and Dr. Douglas C. Noll