Bridging the sim-to-real gap in deep learning quantification of GABA in MEGA-PRESS MRS

Ma, Zien 2026. Bridging the sim-to-real gap in deep learning quantification of GABA in MEGA-PRESS MRS. PhD Thesis, Cardiff University. Item availability restricted.

	PDF (Thesis) - Accepted Post-Print Version Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (15MB)
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (604kB)

Abstract

Proton magnetic resonance spectroscopy (1H-MRS) enables non-invasive estimation of brain metabolite concentrations, but reliable quantification remains difficult because resonances overlap and weak targets are sensitive to acquisition variability. This challenge is particularly acute for y-aminobutyric acid (GABA) in edited Mescher–Garwood Point RESolved Spectroscopy (MEGA-PRESS) spectra, where subtraction-based editing improves specificity but increases susceptibility to instability and structured background effects. Consequently, strong performance on idealised benchmarks does not necessarily translate to reliable estimates under experimental conditions. This thesis evaluates learning-based MEGA-PRESS quantification within a controlled cross-domain validation framework, using simulated spectra as the source domain and experimentally acquired phantom spectra as the target domain. Models are trained on large, physics-based simulated datasets with known concentrations and are selected under a unified optimisation protocol. Robustness is then tested on an experimental phantom benchmark with known prepared concentrations, providing ground-truth validation without relying on inter-algorithm agreement. To address the simulation-to-phantom discrepancy and improve phantom robustness, the thesis investigates data-centric refinements to training data generation while holding network architectures and optimisation settings fixed, including phantom-derived non-Gaussian noise injection, dictionary diversity via basis-set variability, and linewidth augmentation. On idealised simulated validation data, optimised models achieve near-perfect quantification performance (e.g. GABA mean absolute error (MAE) ≈ 0.011–0.013 with coefficient of determination r^2 ≥ 0.99). However, performance degrades substantially on phantoms: pooled GABA MAE increases to 0.161 (Y-shaped autoencoder) and 0.203 for the convolutional neural network (CNN), with 0.220 for LCModel fitted to the edit-OFF spectrum (LCModel-OFF). Data-centric refinements measurably improve phantom robustness. Replacing the time-domain Gaussian noise model withphantom-derived Generalised Gaussian (GG) noise reduces pooled overall phantom MAE from 0.147 to 0.115 for the CNN, and combining GG noise with diverse simulated basis sets achieves the lowest pooled overall MAE of 0.110 for the CNN. Linewidth augmentation reduces pooled phantom GABA MAE to 0.151 for the autoencoder, while errors remain above simulated levels, indicating that linewidth and noise alone do not account for the remaining performance loss on phantoms. Overall, the results show that near-perfect simulated validation is not a reliable indicator of experimental robustness, and that discrepancy-guided, data-centric refinements can reduce, but not eliminate, the sim-to-real performance gap on this phantom benchmark

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Computational & Mathematical Sciences Schools > Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA76 Computer software Q Science > QC Physics
Uncontrolled Keywords:	MEGA-PRESS, Sim-to-Real Gap, GABA Quantification, Deep Learning, Magnetic Resonance Spectroscopy.
Date of First Compliant Deposit:	9 October 2026
Date of Acceptance:	8 October 2026
Last Modified:	09 Oct 2026 14:13
URI:	https://orca.cardiff.ac.uk/id/eprint/190109

Actions (repository staff only)

Edit Item