Deep Learning-based 3D Tractography Data Analysis

Abstract

The understanding of brain white matter connections is fundamentally important in cognitive neuroscience and neurological disease studies. Diffusion magnetic resonance imaging uniquely enables in vivo mapping of brain’s white matter connections at macro scale via tractography. The generated tractography data provides unique information about brain’s structural connectivity and brain tissue “microstructure” relating to the health of neural tissue. Therefore, tractography data analysis plays an important role in studies of brain’s connections in health and disease, as well as in translational research on computational neuroscience. Tractography data is composed of 3D trajectories of white matter fiber pathways estimated through tractography, known as fibers. The fact that tractography is a unique, high-dimensional, unstructured geometric data poses great challenges to the development of approaches for tractography data analysis. In recent years, deep learning techniques have demonstrated success in tractographyfocused applications. However, key challenges remain in each specific task. In this thesis, we study three tractography-focused applications that span important tasks in computer vision. The three application tasks are the parcellation of tractography, the classification of individuals based on tractography, and the prediction of cognitive performance scores based on tractography. For each task, we designed novel geometric deep learning frameworks for processing tractography data. The frameworks are able to incorporate novel anatomical features that have not been previously employed, including information about the white matter trajectory and gray matter connectivity. As a result, performance is improved across multiple tractography-focused tasks by proposing novel geometric deep networks incorporating additional brain anatomical information. Our studies provide effective tools for tractography analysis and tractography-based neuroscientific studies.