Explainable Signatures of Diabetes and Pancreas Aging From Clinical Imaging and AI

Abstract

Large-scale clinical medical imaging presents a powerful opportunity for data-driven discovery of anatomical patterns in diabetes. However, the high variability of clinical imaging necessitates the extraction of robust, explainable representations to enable meaningful characterization of disease signatures. This thesis addresses two primary aims: (1) to characterize explainable anatomical signatures of diabetes in the abdomen, and (2) to characterize the structure of the aging pancreas. The work is organized around four conceptual pillars: navigating heterogeneous clinical imaging, understanding the limitations of current measurement approaches, developing explainable representations of abdominal anatomy, and characterizing anatomical signatures of diabetes and pancreas aging. Together, these contributions form a system for pattern discovery in highly variable imaging data, enabling validation of known diabetes patterns and supporting the search for novel phenotypes.