DEEP LEARNING FOR ANATOMICAL BRAIN MRI HARMONIZATION

Abstract

Mega-analyses have significantly propelled neuroimaging research by amalgamating extensive data from multiple sites. Harmonization is vital to mitigate undesired site differences during data amalgamation, yet existing MRI harmonization algorithms often overlook downstream predictive applications, potentially yielding suboptimal results. Furthermore, many deep learning harmonization approaches neglect covariate inclusion, eliminating differences across sites rather than reducing undesired differences. In this thesis, we focused on addressing the corresponding research gaps by proposing novel deep learning models. Using three large-scale datasets with a total of 2787 participants and 10085 anatomical T1 scans, we demonstrated that our approaches outperformed existing approaches in preserving relevant biological information while removing site differences. Overall, this thesis contributes novel and superior neuroimaging harmonization methodologies, emphasizes the importance of appropriately framing neuroimaging problems under the deep learning paradigm and highlights the importance of adequately modeling based on specific research problems future harmonization studies.