Development of Novel Deep Learning Techniques for Accelerated Diffusion MRI

Abstract

Diffusion Magnetic Resonance Imaging (dMRI) is an imaging modality of MRI that features a non-invasive technique for qualitatively and quantitatively characterizing microstructural characteristics in tissue. This is achieved by dMRI being sensitive to the Brownian motion of water molecules in tissue measured over an applied magnetic field gradient. The directional orientations of motion of water molecules are estimated over an isotropic gaussian propagator. This capability enables dMRI to achieve resolution of microstructural features of up to 2-3 orders of magnitude below the resolution limit of conventional MRI. However, delineation of resolving features and characteristics with high degrees of architectural specificities requires making underlying assumptions of the underlying diffusion signals and invoking appropriate mathematical and computational solutions to achieve this. One computational technique of interest for dMRI is Diffusion Tensor Imaging (DTI). DTI enables anisotropic measurements to be acquired by fitting multiple diffusion-weighted images (DWIs) to a tensor, enabling 3D reconstructions of microstructural features of the brain, including the ability to reconstruct trajectories of white-matter tracts. Some of the challenges of performing DTI in routine clinical and research studies include long data acquisition times required to obtain sufficiently large number of DWIs for outputting robust tensor estimates. This involves long scan times and typically introduces undesired image distortions and artifacts that deteriorate image quality. In addition, DTI has limitations in accurately delineating more complex microstructural features in white-matter tracts. Models in Diffusion Kurtosis Imaging (DKI) and Constrained Spherical Deconvolution (CSD) offer improvements over DTI. However, these models require considerably more data than DTI and are generally more complex to implement. Although some prior deep learning (DL) techniques have addressed limitations of conventional diffusion models, these techniques typically involve supervised-learning frameworks that require large amounts of clean training data to successfully produce robust estimates of metrics and are typically constrained to diffusion-specific models. For this research, we propose DL techniques that address some of these challenges. The contributions made by this dissertation involve a Self-Supervised with Fine-Tuning DL pipeline that can produce robust DTI metrics for an accelerated acquisition of DWIs and reduces the need of a large volume of clean training data. We also demonstrate a Generative Diffusion Deep Learning model that can effectively leverage uncertainty to generalize well to the underlying distribution of tensor model metrics in DTI and DKI and bypasses the need for diffusion-model fits. We also present a DL pipeline, AcceleraTed deep-LeArning for model-free and multi-Shell (ATLAS) DWI, that can predict a full acquisition of DWIs across multiple shells, given an accelerated acquisition in one shell or multiple shells. This enables for the potential for robust DTI tensor estimates, overcoming the requirement for large amounts of clean training labels, and eliminates the constraint of diffusion-specific models, which introduces the exciting potential to obtain diffusion metrics that more accurately delineate white-matter tracts.