| dc.description.abstract | Diffusion weighted MRI (DW-MRI) is known for mapping white matter fibers of the brain and serves as the only available technique to probe tissue structure at a microscopic level in-vivo. This has opened new investigations into cognitive neuroscience and brain dysfunction in aging, addiction, mental health disorders, and neurological disease. However, statistical analysis of DW-MRI is held back by bias introduced by many factors. Variability in DW-MRI measurements can result from a difference in the number of head coils used, the sensitivity of the coils, the imaging gradient non-linearity, the magnetic field homogeneity, the differences in the algorithms used to reconstruct the data, as well as changes made during software upgrades. Harmonization approaches and methods aim to increase reproducibility and reduce error caused by variance and bias introduced by hardware and site effects. This thesis explores multiple facets of DW-MRI harmonization using machine learning techniques. First, we propose empirical harmonization methods which can correct for spatially varying signal drift and gradient nonlinearities in DW-MRI systems and show the resulting DW-MRI metrics have reduced reproducibility error. Second, we develop a population based white matter atlas and an automated white matter bundle segmentation method which can make accurate predictions given T1 contrast information and compare their performance in multiple datasets. Third, we formalize the theory of a semi-supervised contrastive learning method called Nullspace Tuning and show how it can leverage paired data in many tasks as well as in harmonization to achieve reduced error. Last, we utilize Nullspace Tuning as well as T1 derived information to develop a semi-supervised framework to explore the harmonization of multiple DW-MRI datasets and acquisition parameters and show that the resulting harmonized DW-MRI metrics have lower reproducibility error than the baseline methods. | |
