Date of Graduation

12-2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Engineering (PhD)

Degree Level

Graduate

Department

Computer Science & Computer Engineering

Advisor/Mentor

Nakarmi, Ukash

Committee Member

Gauch, Susan E.

Second Committee Member

Panda, Brajendra N.

Third Committee Member

Pan, Yanjun

Fourth Committee Member

Wang, Dongyi

Keywords

Deep Learning; Magnetic Resonance Imaging; Oversampling; Reconstruction; Segmentation; Unsupervised

Abstract

Magnetic Resonance Imaging (MRI) is typically a slow process because of its sequential data acquisition. To speed up this process, MR acquisition is often accelerated by undersampling k-space signals and solving an ill-posed problem through a constrained optimization process. Image reconstruction from under-sampled data is posed as an inverse problem in traditional model-based learning paradigms. While traditional methods use image priors as constraints, modern deep learning methods use supervised learning with ground truth images to learn image features and priors. However, in some cases, ground truth images are not available, making supervised learning impractical. Recent data-centric learning frameworks such as deep learning (DL) frameworks are data hungry, and demand a large, labeled training data sets. To address the lack of large training datasets, in MRI reconstructions, researchers approach the problem in two ways: (1) using a method that efficiently use the limited dataset for training purpose. (2) unsupervised method where the model is trained without the presence of fully sampled data. In this dissertation, we first systematically investigate advantages and limitations of current oversampling methods. Then, we propose a novel oversampling method and a DL framework that systematically exploits the oversampling technique in the learning process as well as also increase the size of training data set. For (2) we propose an unsupervised deep learning framework for accelerated MRI that does not require ground truth images for training. Our framework combines a system prior derived from the MR acquisition model with generic image priors to build a more effective unsupervised deep learning framework. The system prior enforces data consistency while the generic image priors regulate the neural network parameters. Our experimental results demonstrate that our proposed unsupervised method outperforms state-of-the-art unsupervised methods and achieves performance comparable to that of supervised methods that require ground truth images for training.

Share

COinS