Integrative Bayesian modeling of imaging and genetic data
My Ph.D. dissertation research considered two inherently multidisciplinary topics pertaining to the development of principled statistical methods of inference for data integration in the context of medical image studies. ^ 1. Integrative Bayesian analysis of neuroimaging-genetic data with application to cocaine dependence^ This multidisciplinary research pertains to integrative Bayesian analysis of neuroimaging–genetic data with application to the study of cocaine dependence. This area of research necessitates an understanding of genomics, biological and clinical contexts, image processing, spatial statistical methods that are appropriate for conducting inference of volumetric imaging phenotypes, and efficient computational algorithms that facilitate the implementation of the analysis. We examine the link between neuroimaging and genetic factors through (i) voxel-wise residual error modeling and (ii) and component-wise inference through dimension reduction. In voxel-wise analysis, a general statistical framework for integrative Bayesian analysis of neuroimaging-genetic (iBANG) data was established. Statistical inference necessitated the integration of spatially dependent voxel-level measurements with various patient-level genetic and demographic characteristics under an appropriate probability model to account for the multiple inherent sources of variation. In component-wise inference, a general framework was introduced to perform advanced principal component analysis to account for the spatial correlation of voxels on each region of interest (ROI) and conduct Bayesian analysis to predict the impact of genetic and demographic features on the white matter of the brain. Both study results suggested that cocaine consumption was associated with fractional anisotropy (FA) reduction in most white matter regions of interest in the brain. Additionally, gene polymorphisms associated with GABAergic, serotonergic and dopaminergic neurotransmitters and receptors were associated with FA. Due to the presence of some discordance in the results effectuated by the voxel-wise and component-wise approaches, our research also demonstrates that high-dimensional imaging studies can be sensitive to statistical method and level of inference.^ 2. Bayesian models for integration of multivariate perfusion maps with implications for evaluation of metastatic sites^ Imaging technologies (such as perfusion CT, DCE-MRI, DWI-MRI, FDG-PET) provide numerous avenues for extracting biomarkers from imaging features that can be leveraged to improve cancer detection. Yet, due to a lack of understanding of complex models in general and multivariate analysis in particular, radiologists often attempt to construct biomarkers by imposing thresholds on single-number region of interest (ROI) summaries, thereby discarding much of the information in the data. To fulfill cancer evaluation on the basis of imaging features, we have defined two general statistical objectives: (i) enhance signal extraction from spatiotemporal maps that involve multivariate voxel-level data for assessing the extent of disease at multiple interdependent sites within an organ, and (ii) establish statistical methods for building posterior probability maps of organs of the basis of perfusion characteristics to provide tools for cancer detection, prognostication, and treatment monitoring. The applied objective of this project is to develop probabilistic-based methods of inference that can be used to integrate multiple perfusion characteristics for detection, prognostication, and monitoring of metastatic sites in the liver.^
Azadeh, Shabnam, "Integrative Bayesian modeling of imaging and genetic data" (2015). Texas Medical Center Dissertations (via ProQuest). AAI10027842.