Lamb, Alistair Charles; (2025) Deep Learning for Advancing Whole-Body MRI as a Robust Clinical Tool in Metastatic Cancer. Doctoral thesis (Ph.D), UCL (University College London).

Text Lamb_10211680_Thesis.pdf Access restricted to UCL open access staff until 1 August 2026. Download (68MB)

Abstract

Despite a long publication history of whole-body (WB) quantitative magnetic resonance imaging (qMRI) as an imaging biomarker in oncology, it has not become widely used in clinical practice. In this thesis, some main obstacles to its adoption are investigated, such as variability of measurements between scans, and the presence of artefacts. The application of deep learning (DL) techniques to address some of these issues is also explored. In the first contribution, the within-subject variability of apparent diffusion coefficient (ADC) estimates in whole-body diffusion-weighted imaging (WB-DWI), both within- and between-scanners, is investigated to identify the largest source of variance in multicentre trials. This is done with a dataset of ten healthy volunteers tested and retested within- and between-scanners from different vendors with minimal differences in acquisition protocol and post-acquisition analysis. The next part of this thesis starts with a detailed overview of the appearance and causes of common artefacts present in diffusion-weighted echo-planar imaging (DW-EPI). WB-DWI data from a multi-centre trial is then explored in order to determine the occurrence of these artefacts in a clinical setting. A supervised DL approach is then proposed for detecting the presence of Nyquist ghosting in WB-DWI data. A transfer learning-based model is first trained and tested on data from a single site of the multi-centre trial. The model is then tested on data from a second site to evaluate its robustness to differences in scanner and protocol. Finally, the feasibility of WB variable flip angle (VFA) quantitative T1 (qT1) mapping using only two flip angles (FAs) is investigated, as this could allow data to be collected in a clinically feasible time frame. First, the FA pair which best reproduces results using eight FAs was determined for conventional linear least squares (LLS) fitting. The use of DL to improve fitting with two FAs was then explored.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item