Toward Reliable and Interpretable AI for 3D Medical Image Computing: Registration, Cross Modal Synthesis, and Resource Efficient Segmentation

Abstract

Medical imaging is essential for diagnosing and tracking neurological disease, but anatomy, molecular pathology, and tissue mechanics are typically measured with separate modalities that vary in availability, burden, and cost. Structural MRI is widely available, whereas PET and MRE yield valuable molecular and mechanical biomarkers that are harder to obtain and scale. Learning-based models can fuse these complementary data to extract richer information, but struggle when anatomy is highly altered, behavior is hard to interpret, resources are limited, or key signals are rare and extreme. This dissertation develops structural and information-driven methods to address these challenges in three-dimensional medical image computing. Across four studies, I investigate postoperative MRI registration via multi-contrast fusion, source-factorized latent representation learning for MRI-to-tau-PET synthesis, partial information decomposition for MRI input selection in brain tumor segmentation, and synthesis of MRE-derived stiffness maps from structural and diffusion MRI. Together, these studies propose methods to improve the robustness, interpretability, and resource efficiency of multi-contrast medical imaging models