Landman, Bennett A.2024-08-162024-08-162024-082024-07-16August 202http://hdl.handle.net/1803/19262Abdominal computed tomography (CT) offers high-resolution tissue maps that facilitate the quantification of body composition. The existing longitudinal CT dataset from the Baltimore Longitudinal Study of Aging (BLSA) enables us to explore the relationship between body composition and cognitive decline. To minimize radiation exposure, 2D axial scans are used instead of 3D volumetric scans, presenting several challenges. Notably, 2D scans lack the contextual information inherent in 3D data, complicating body composition segmentation. Furthermore, difficulties in consistently positioning cross-sectional scans mean that 2D slices may not always be taken at the same vertebral level, resulting in variability in the abdominal slices obtained from patients over different years. In this dissertation, we utilize existing 3D volumetric data for representation learning to mitigate the anatomical limitations present in single-slice data. We apply advanced deep learning techniques for segmentation and generation to address these challenges. Ultimately, we identify a potential underlying phenotype associated with cognitive decline using the derived body composition metrics, providing new insights into the connection between the body and the brain.application/pdfen3D-2D, Single-Slice CT, Body CompositionThesis2024-08-16