Towards Robust Clinical Segmentation of Paediatric Brain Tumours in Magnetic Resonance Images Using Weakly-supervised Deep Neural Networks
- Publication Type:
- Thesis
- Issue Date:
- 2026
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
The delineation of paediatric brain tumours in magnetic resonance imaging (MRI) remains a major clinical challenge due to complex tumour presentations, and the heightened sensitivity of developing brains to treatment. Manual annotation is time-consuming and subjective, with inconsistencies arising from differences in imaging quality and tumour morphology. Automated approaches show promise in addressing these issues, yet state-of-the-art supervised deep learning (DL) methods depend on extensive, pixel-level annotations that are costly and scarce in paediatric populations. To overcome these limitations, this thesis investigates weakly-supervised anomaly detection based on denoising diffusion probabilistic models (DDPMs) as an alternative for delineating paediatric brain tumours in MRI.
The first contribution introduces a 3D-latent diffusion model (LDM) with a novel patch-based training strategy that enables efficient learning on volumetric data while reducing computational demand. This strategy also facilitates the extraction of pseudo-healthy anatomy from diseased individuals, mitigating data collection requirements. The applicability of a novel encoding mechanism, adapted from natural images, is further assessed for medical imaging. The approach surpasses existing weakly-supervised baselines across several benchmarks. However, the generation of artefacts raises important questions regarding performance on small lesion detection.
To address these limitations, the second contribution exploits the generative capacity of LDMs to synthesise datasets with precisely controlled lesion sizes. Various conditioning strategies are systematically compared to balance fidelity and dataset consistency. Building on this foundation, the third contribution explores spatial resolution enhancement via super-resolution (SR) using a conditional LDM to improve the detection of small lesions. The results demonstrate clear gains in lesion sensitivity and resolution-aware segmentation.
The fourth contribution assesses the generalisability of the proposed framework to paediatric tumours, an underexplored domain limited by scarce annotations. Experimental results show that LDMs trained solely on adult data generalise effectively to paediatric cases, while fine-tuning yields negligible gains. Additional evaluation on a private multi-institutional cohort encompassing diverse tumour types and acquisition conditions further supports the framework’s robustness. These findings demonstrate that anomaly detection can extend beyond its original training domain and underscore the framework’s relevance in low-annotation regimes. Together, these contributions advance the use of LDMs for weakly-supervised anomaly detection in medical imaging, unifying lesion detection, spatial resolution enhancement, and synthetic data generation. The framework reduces dependence on large annotated datasets and demonstrates robustness across adult and paediatric cohorts. As a result, this thesis outlines a pathway towards scalable and clinically applicable paediatric tumour segmentation beyond conventional supervised paradigms.
Please use this identifier to cite or link to this item: