Enabling Large-scale Analysis of Neurodegeneration through Diffusion Imaging Harmonization

Abstract

Diffusion-weighted imaging (DWI) plays a critical role in characterizing brain network structure and connectivity, yet its broader application is limited by methodological variability and site-specific biases that compromise reproducibility and scalability. This dissertation addresses these challenges by proposing harmonization strategies and machine learning methods to enable robust, large-scale analysis of white matter connectivity, particularly in the context of aging and Alzheimer’s disease (AD). First, we evaluate the reproducibility and comparability of tractography pipelines, a cornerstone of DWI analysis. Second, we tackle the problem of cross-site variability by assessing and developing harmonization techniques for DWI-based connectivity and bundle analysis. We organized QuantConn, a community challenge focused on harmonizing tractography and bundle metrics across protocols. Third, we present a machine learning framework to disentangle and remove site effects from connectome features while preserving biological variability. Leveraging a conditional variational autoencoder, we harmonize nine network connectivity measures across 38 unique DWI acquisition protocols from 6,956 individuals, including those with mild cognitive impairment and AD. Collectively, this work establishes a comprehensive pipeline for tractography characterization, data harmonization, and site bias removal in network connectivity analysis.