Analyzing Heterogeneity in Neuroimaging with Probabilistic Multivariate Clustering Approaches

Automated quantitative neuroimaging analysis methods have been crucial in elucidating normal and pathological brain structure and function, and in building in vivo markers of disease and its progression. Commonly used methods can identify and precisely quantify subtle and spatially complex imaging patterns of brain change associated with brain diseases. However, the overarching premise of these methods is that the disease group is a homogeneous entity resulting from a single, unifying pathophysiological process that has a single imaging signature. This assumption ignores ample evidence for the heterogeneous nature of neurodegenerative diseases and neuropsychiatric disorders, resulting in incomplete or misleading descriptions. Accurate characterization of heterogeneity is important for deepening our understanding of neurobiological processes, thus leading to improved disease diagnosis and prognosis. In this thesis, we leveraged machine learning techniques to develop novel tools that can analyze the heterogeneity in both cross-sectional and longitudinal neuroimaging studies. Specifically, we developed a semi-supervised clustering method for characterizing heterogeneity in cross-sectional group comparison studies, where normal and patient populations are modeled as high-dimensional point distributions, and heterogeneous disease effects are captured by estimating multiple transformations that align the two distributions, while accounting for the effect of nuisance covariates. Moreover, toward dissecting the heterogeneity in longitudinal cohorts, we proposed a method which simultaneously fits multiple population longitudinal multivariate trajectories and clusters subjects into subgroups. Longitudinal trajectories are modeled using spatiotemporally regularized cubic splines, while clustering is performed by assigning subjects to the subgroup whose population trajectory best fits their data. The proposed tools were extensively validated using synthetic data. Importantly, they were applied to study the heterogeneity in large clinical neuroimaging cohorts. We identified four disease subtypes with distinct imaging signatures using data from Alzheimer’s Disease Neuroimaging Initiative, and revealed two subgroups with different longitudinal patterns using data from Baltimore Longitudinal Study on Aging. Critically, we were able to further characterize the subgroups in each of the studies by performing statistical analyses evaluating subgroup differences with additional information such as neurocognitive data. Our results demonstrate the strength of the developed methods, and may pave the road for a broader understanding of the complexity of brain aging and Alzheimer’s disease