Robust Deep Learning for Stroke Detection in Clinical Neuroimaging
PhD thesis, University College London.
Abstract. This thesis develops robust automated segmentation of stroke lesions in clinical neuroimaging, addressing the diversity of human MRI and CT data in contrasts, resolutions, and artefacts. The work connects architectural advances, data-centric training, and physics-guided synthesis to deliver reliable lesion maps for downstream clinical decision-making.
Chapter 1: I introduce the clinical and imaging background for stroke, outline evaluation metrics for lesion segmentation, and frame the domain-shift challenges that motivate the subsequent methods.
Chapter 2: I present a segmentation model that uses a convolutional variant of the transformer to achieve a larger receptive field than standard convolutional neural networks (CNNs). The enhanced shape awareness and inductive biases yield more robust feature learning and generalisation to out-of-distribution clinical scanners.
Chapter 3: I explore hypernetworks—networks that generate the weights of another network—to adapt dynamically to imaging protocols. Conditioning on discrete domains (e.g., CT vs. MRI) or continuous MRI sequence parameters specialises the model on-the-fly, enabling broad coverage of clinical imaging conditions without retraining.
Chapter 4: I refine stroke-focused synthetic data generation. Improved tissue priors, a lesion-pasting strategy for heterogeneous stroke appearances, and domain adaptation methods expand the anatomical and contrast diversity encountered during training, strengthening robustness to unseen domains.
Chapter 5: I incorporate quantitative MRI (qMRI) parameter estimation to enforce physical plausibility in synthetic image generation. By simulating images from qMRI-driven intensity priors and forward models of MRI signal formation, I produce realistic synthetic data that support domain-agnostic augmentation. Collectively, these architectural innovations (Ch. 2, 3), synthetic training pipelines (Ch. 4), and qMRI-based augmentation strategies (Ch. 5) deliver more dependable stroke lesion segmentation across heterogeneous clinical imaging settings.
Year 2026Kind thesisVenue University College London
Materials
UCL Discovery provides the integrated account of convolutional attention, hypernetwork adaptation, synthetic stroke training and qMRI synthesis across five chapters.
@phdthesis{chalcroft2026robust,
author = {Chalcroft, Liam Ford},
title = {{Robust Deep Learning for Stroke Detection in Clinical Neuroimaging}},
school = {University College London},
year = {2026},
url = {https://discovery.ucl.ac.uk/id/eprint/10220048/}
}