Development of Advanced Generative Priors for MRI Reconstruction

by Guanxiong Luo
Doctoral thesis
Date of Examination:2023-11-10
Date of issue:2024-05-02
Advisor:Prof. Dr. Martin Uecker
Referee:Prof. Dr. Martin Uecker
Referee:Prof. Dr. Philipp Wieder
Persistent Address: http://dx.doi.org/10.53846/goediss-10451

 

 

Files in this item

Name:thesis_twoside.pdf
Size:25.9Mb
Format:PDF

This file will be freely accessible after 2024-11-10.

Abstract

English

This thesis investigates generative models for parallel compressed magnetic resonance imaging (MRI) reconstruction. The challenges of compressed MRI reconstruction using deep learning are identified, including model interpretability, data availability, and generalizability. The thesis proposes three approaches to address these challenges. First, a Bayesian imaging framework using diffusion models is proposed to address the uncertainty introduced by missing data. This framework allows for the computation of not only the most likely image but also uncertainty maps. Second, a workflow to construct generic and robust generative image priors from magnitude-only images is presented. These priors can be used to improve the quality of reconstructed images. The importance of incorporating phase information and using large datasets for training the priors is highlighted. Third, a Python package named "spreco" is introduced to facilitate the development life cycle of generative priors, from experimentation to deployment.
Keywords: Generative AI; Inverse Problem; Magnetic Resonance Imaging; Image Reconstruction