Year: 2026
Type: dissertation
Abstract: Diffusion magnetic resonance imaging provides valuable information about tissue microstructure but remains challenging at low magnetic field strengths because of reduced signal-to-noise ratio, limited spatial resolution and long acquisition times. These limitations are particularly important when imaging populations who may be unable to remain still, such as neonates, and when developing accessible imaging systems for settings where conventional high-field MRI is unavailable. This thesis investigates computational methods for improving the efficiency and quality of diffusion Imaging under these constraints. First, a four-direction diffusion tensor imaging framework was developed to reduce the number of diffusion-weighted images required for tensor estimation. Several estimation approaches were investigated, including principal component analysis, constrained nonlinear optimisation and deep learning. The results demonstrated that useful diffusion metrics could be estimated from a substantially reduced acquisition. Second, digital diffusion phantoms were developed to provide known ground-truth tissue properties for evaluating the effects of denoising on diffusion tensor estimation. These phantoms were used to compare several established denoising methods across different signal-to-noise ratios, numbers of diffusion directions and tissue configurations. Third, a wavelet-informed deep-learning framework was developed for denoising complex MRI images. The model combined image-domain information with wavelet scattering representations and was quantitatively evaluated using pixelwise errors and errors in derived diffusion tensor metrics. Finally, a diffusion-informed super-resolution framework was developed to reconstruct higher-resolution diffusion-weighted images. Fibre orientation distribution functions were incorporated as structural priors. Quantitative validation on matched 3 T acquisitions showed improved image and diffusion-metric accuracy compared with standard cubic interpolation, while application to low-field data demonstrated the feasibility of the framework under realistic acquisition conditions. Together, these contributions show how carefully designed computational techniques can be used to recover useful information from limited DWI measurements. The findings provide a foundation for faster and higher-quality diffusion imaging at low field. All models and code developed in this study will be released as open source, with the code already hosted on GitHub. All datasets generated, including the digital phantom datasets, will be made available through publicly accessible repositories.
Author: Ametepe, Joshua
Language: English
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7225369105
