Fast And Quantitative Imaging Methods Of Brain Microstructure Based On Diffusion Magnetic Resonance Imaging

Diffusion magnetic resonance imaging(d MRI)utilizes its sensitivity for the Brownian motion of water molecules to probe brain microstructure at the cellular level.By modelling connections between signals and tissue microstructure,the existing methods have extracted various biomarkers from d MRI signals,making d MRI an important tool for neuroscience and clinical research.However,several major issues still limit the widespread uptake of these microstructure imaging methods,including long acquisition time,diffusion model fitting susceptible to imaging artefacts,and insufficient specificity in microstructural measures.In this thesis,we propose several methods to address these specific issues,in order to improve feasibility of fast and quantitative brain microstructure imaging.In Chapter 3 and 4,two fast d MRI methods using convolutional neural network(CNN)algorithm are introduced,achieving over 10 times acceleration of acquisition time for diffusion kurtosis imaging(DKI)and fiber orientation distribution function(f ODF)imaging respectively.By tailoring the network design for specific diffusion model properties,our proposed methods provide improved estimation performance compared to existing methods.Specifically,we design a partially shared hierarchical network structure taking into consideration the complexity of DKI and diffusion tensor measures,and making full use of the intrinsic correlations of all model measures.For fast f ODF imaging,we use spherical harmonics representation of DWI signals to incorporate the diffusion gradient directions and strengths as inputs,since they are vital for orientation estimation.Furthermore,we use small convolution kernels to exploit local spatial correlation for both methods,which substantially improves the estimation especially when higher acceleration is used.In Chapter 5,we develop a deep-learning-based technique to minimize residual motion effects for quantitative diffusion studies,in order to address the challenge of motion level dependent bias in diffusion model-derived measures,introduced by conventional motion correction techniques.The technique combines a hierarchical CNN method we developed with a module of motion assessment and corrupted image volume rejection,to overcome limitations of existing methods in which the quality of the parameter estimation depends strongly on the proportion of the data discarded.Results suggest great potential of our method for reducing residual motion effects in motioncorrupted DWI data,bringing benefits that include reduced bias in derived diffusion metrics in individual scans and reduced motion-level dependent bias in population studies employing d MRI.In Chapter 6,we introduce MTE-NODDI(multi echo time neurite orientation dispersion and density imaging),a combined diffusion and T2 relaxation microstructure modelling technique,which provides non-T2-weighted intra-neurite and free water fractions.It solves the TE-dependence of conventional compartment fractions;meanwhile,it also provides the intra-and extra-neurite T2 relaxation time that relaxation methods alone cannot distinguish.We show theoretically that simple linear fittings can estimate the specific model parameters,and experimentally we verify that even data from 2-3 TE can achieve robust estimation.This technique improves the specificity and quantitative property of micro-parameters,which help improve the interpretability of future neuroimaging studies,especially those in brain development,maturation and aging.In summary,the four methods improve diffusion magnetic resonance imaging from a fast and quantitative perspective,which can assist the clinical application and scientific research of microstructure imaging in the future.