Date of Award

2025-08-01

Degree Name

Doctor of Philosophy

Department

Electrical and Computer Engineering

Advisor(s)

Sergio D. Cabrera

Second Advisor

Ali Bilgin

Abstract

The proliferation of data-driven tools for solving problems in every possible domain, coupled with rapid advances in computing technology, has led to an arms race of AI development and application research in industry and academia. One field of research that stands to gain immeasurably from this revolution is medical imaging. It is a critical part of modern diagnostics, and advancements in this area can directly benefit the average person by making healthcare more accessible, accurate, and affordable. Breakthroughs in mainstream image processing and computer vision have long fueled development in medical imaging, and it is now common to see cutting edge machine vision models developed less than a year ago being explored to solve problems in this field. Convolutional models are already being tested in commercial scanners for a variety of tasks such as pathology detection, organ segmentation, and image denoising, with Vision Transformers lining up to be the next big breakthrough.

This Ph.D. research focuses on the effective use of the notoriously large and unwieldy Transformer models to accelerate Magnetic Resonance Imaging (MRI) scans. To this end, it explores the use of deep neural networks for undersampled multi-channel reconstruction of complex MRI data in a k-space data consistent fashion with a focus on Shifted-Window (Swin) Transformers. Specifically, cascaded reconstruction pipelines were designed using customized Swin Transformer blocks featuring overlapped window attention, and their performance tested against the state-of-the-art methods.

The results show that cleverly designed small Transformer architectures working together can outperform larger monolithic structures. Additionally, an exploration of transfer learning in this area revealed that it is possible to leverage the abundance of widely available magnitude MR images to pre-train transformer blocks and adapt them into cascaded reconstruction architectures to be fine-tuned for multi-channel complex reconstruction tasks.

Language

en

Provenance

Received from ProQuest

File Size

100 p.

File Format

application/pdf

Rights Holder

Tahsin Rahman

Download

Share

COinS