Using deep learning to ensure the safety of patients who cannot remain still during ultra-high field MRI

Blanter, Katherine 2025. Using deep learning to ensure the safety of patients who cannot remain still during ultra-high field MRI. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Download (45MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (128kB)

Abstract

Ultra-high field (UHF) MRI may enable improved image resolution and contrast but has the potential to introduce tissue-heating-related patient safety concerns, approximated by specific absorption rate calculations (SAR). In an effort to prioritize patient safety, the full potential of UHF MRI is impeded, limiting clinical utility. This becomes more pressing when patients cannot remain still. Rigid motion during imaging alters the location of local SAR in an unpredictable way. Traditional safety models are not responsive to or reflective of these dynamics. This thesis investigates the use of deep learning to predict the effect of motion on SAR distributions during parallel radio frequency transmission (pTx) MRI at 7 Tesla. The work begins by establishing the foundational physics of MRI and the mechanisms underlying radio frequency power deposition. The limitations of conventional SAR simulation methods are discussed. Deep learning is introduced as a method capable of mapping non-linear spatial relationships between field data and anatomy. The first study chapter provides background to the pipeline development process. The objective here is to ensure producibility and caution against practices in deep-learning-based research which may hinder project reproducibility. Later chapters describe two deep learning models, Conditional Generative Adversarial Networks (cGANs) (Chapter 4) and U-Nets (Chapter 7), trained to estimate motion-induced changes in local SAR and SAR-related datatypes. Based on the initial cGAN implementations in Chapter 4, it is hypothesized that the U-Net would yield more accurate results. Several data formats and reprocessing strategies are assessed to determine which pipeline produces the most reliable estimates (Chapters 4 and 5 ). The work examines how neural network structure and pre and post-processing decisions affect prediction accuracy. As a result, smoother datatype representations and simpler pipelines are expected to yield better estimations in subsequent investigations (Chapters 6 and 7). The overarching research objective was to design a proof-of-concept pipeline for deep-learning-driven SAR prediction during head movement during UHF MRI. The cumulative results offer guidance for deep learning pipeline development and suggest that such models may support safer scanning of patients who move during imaging. The findings support the potential of deep learning to extend safe imaging capabilities, reduce conservative safety constraints, and improve scan efficiency without compromising patient safety.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Psychology
Date of First Compliant Deposit:	23 March 2026
Last Modified:	23 Mar 2026 14:23
URI:	https://orca.cardiff.ac.uk/id/eprint/185745

Actions (repository staff only)

Edit Item