Developing non-invasive neuroimaging biomarkers for Alzheimer's disease using machine learning

Abstract

With the FDA’s recent approval of the anti-amyloid antibodies lecanemab and aducanumab, the prospect of disease-modifying therapies for Alzheimer’s disease has become a clinical reality. However, these treatments are only effective in the early stages of the disease which is why there is a need for accurate AD early detection methods. PET imaging has the ability to detect key biomarkers associated with AD up to 20 years before symptoms occur. However, PET is largely unavailable in many countries and even in the U.S. is limited and costly. Here, we present a deep learning framework for synthesizing Aβ, Tau, and FDG-PET from MR images, which can reduce reliance on standard PET imaging by reconstructing the same pathology from learned representations in the MR inputs. We find that our UNet is able to synthesize structurally similar and clinically relevant Aβ, Tau, and FDG-PET images, potentially introducing MRI as a more accessible biomarker detection modality. Early AD is also characterized by network-wide functional connectivity changes, which have been observed in the DMN and other regions including the salience and dorsal attention regions. Currently, functional connectivity evaluation methods do not incorporate spatial information into their analysis nor do they evaluate connectivity dynamically, which neglects important interactions that occur in the brain. In the second portion of this thesis, a previously developed Graph Diffusion Autoregressive (GDAR) Model is applied to fMRI data to analyze dynamic functional changes in physiologically distinct brain networks. We found notable differences in connectivity among AD versus control subjects, informing future analysis to develop features that can separate AD from control subjects with the ultimate goal of developing AD early detection functional biomarkers.