Computer vision for 3D neuroimaging data analysis

Abstract

Understanding brain white matter connections is critical for cognitive neuroscience and neurological disease studies. Diffusion magnetic resonance imaging (dMRI) stands out as the only technique that can non-invasively map the brain’s white matter connections at a macro scale by tractography, offering insights into the brain’s structural connectivity and the health of neural tissues. Recently, deep learning methods have been used for anaylsis dMRI data and its derived data (e.g., Tractography). These advanced techniques have opened new possibilities for understanding and interpreting complex neuroimaging data. However, this field continues to face significant challenges. This thesis presents a series of advanced deep learning frameworks for 3D neuroimaging analysis, particularly focusing on dMRI and its derived data. In a collaborative effort between the BMIT Research Group at the University of Sydney and the O’Donnell Laboratory at Harvard Medical School, this research leverages computer vision and deep learning innovations to enhance the understanding of the brain’s structural connections. This thesis focuses on four unique and challenging tasks in dMRI analysis: registration-free tractography segmentation, superficial white matter segmentation, neurocognitive score prediction, and dMRI signal modeling parameter estimation. Our proposed frameworks and algorithms have significantly improved precision and efficiency across these four applications and can potentially extend to other 3D neuroimaging and general computer vision. Also, the enhanced performance is consistently observed across populations (e.g., babies, children, young and older adults, patients, etc).