3D Neuroimaging Data Analysis and Processing with Deep Learning

Abstract

Diffusion magnetic resonance imaging (dMRI) and tractography enable non-invasive mapping of the brain’s white matter pathways, providing a basis for studying structural connectivity and its relation to cognition. The geometric morphology of tractography fiber clusters and tracts remains comparatively underexplored. This thesis addresses this gap by systematically investigating tractography-derived shape measures as novel descriptors of white matter structure, and by developing deep learning algorithms that integrate geometric multimodal representations for large-scale cognitive prediction and neuroimaging analysis applications.

The first section of this thesis introduces the conceptual background of diffusion MRI, tractography, and tract quantification, highlighting current methodological trends and limitations. I then introduce a set of shape descriptors to capture the geometry of white matter fiber clusters, and evaluate their predictive value for cognition. The second section of this thesis develops a set of geometric deep learning algorithms designed for multi-shape prediction across large-scale tractography datasets. This framework provides efficient, robust, and generalizable shape representations, enabling accurate prediction while maintaining computational scalability. The final section presents a geometry-guided score-based diffusion model for tractography, enabling microstructure imputation.

In summary, this thesis establishes tractography shape analysis as a powerful, interpretable, and scalable representation of white matter analysis. By bridging diffusion MRI, deep learning, and brain connection measures, the work advances our understanding of the geometric foundations of brain structure and their association with cognition, while providing methodological innovations that can support future neuroimaging research and translational applications.