Brain tumour segmentation from multimodal MRI: Efficient models and structured robustness assessment

Alwadee, Ebtihal 2026. Brain tumour segmentation from multimodal MRI: Efficient models and structured robustness assessment. PhD Thesis, Cardiff University. Item availability restricted.

	PDF - Accepted Post-Print Version Restricted to Repository staff only until 2 November 2027 due to copyright restrictions. Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (16MB) | Request a copy
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (463kB) | Request a copy

Abstract

Brain tumour segmentation from multimodal Magnetic Resonance Imaging (MRI) supports diagnosis, treatment planning, and response assessment, but it is not a generic medical-image segmentation problem. Glioma subregions are heterogeneous and partly overlapping in appearance, and the relevant evidence is distributed across complementary MRI contrasts: FLAIR emphasises oedema and whole-tumour extent, T1ce highlights enhancing tumour, and T2 contributes additional structural information. This modality dependence means that acquisition artefacts can affect tumour regions differently, so clean benchmark accuracy alone is insufficient for deployment-oriented evaluation. This thesis addresses two related challenges: the computational cost of 3D segmentation models and the need for structured robustness assessment under MRI acquisition variability. First, it introduces LATUP-Net, a lightweight 3D U-Net-family architecture using parallel convolutions, squeeze-and-excitation attention, and a class-imbalance-aware weighted Dice loss. Evaluated on BraTS 2020 and BraTS 2021, LATUP-Net achieves competitive segmentation performance with a compact parameter count and computational budget. Second, the thesis develops a structured robustness evaluation framework for BraTSstyle multimodal glioma segmentation. The framework applies simulated, clinically motivated perturbations per modality and across ordinal severity levels, then analyses degradation using Dice-severity slopes, error distributions, Wasserstein distance, and non-parametric statistical comparisons. The results support the interpretation that clean performance and robustness are distinct: models with strong baseline accuracy may degrade substantially under artefacts, while stable models may remain less accurate overall. Within the LATUP family, affinity-guided robustness training improves the accuracy-robustness trade-off without increasing inference-time cost, although it adds training-time overhead. All results are based on preprocessed BraTS data and simulated perturbations, not raw clinical deployment. External validation, scanner-specific severity calibration, and native clinical workflow testing remain necessary future work

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software
Funders:	Jazan University, Saudi Arabian Cultural Mission
Date of First Compliant Deposit:	2 November 2026
Date of Acceptance:	16 October 2026
Last Modified:	03 Nov 2026 09:47
URI:	https://orca.cardiff.ac.uk/id/eprint/190591

Actions (repository staff only)

Edit Item