Daryl Wilding-McBride

Research Projects

• Automated Longitudinal Segmentation of Glioblastoma on the MR-Linac active

  Glioblastoma is the most aggressive primary brain tumour in adults, with a median survival of around 15 months despite surgery, radiotherapy, and chemotherapy. Tumours are diffuse, morphologically heterogeneous, and can shift substantially during the treatment course, yet standard radiotherapy planning treats them as static. The MR-Linac addresses this by acquiring MRI at every treatment fraction, creating a longitudinal imaging record unique among radiotherapy modalities. In current clinical practice this data is largely underused, because manual tumour delineation is too time-consuming to perform daily. This project addresses that gap with an automated segmentation model based on the nnU-Net framework, trained on the UPENN-GBM dataset (630 de novo GBM patients) using T2 and FLAIR sequences, and evaluated against expert manual delineations on MR-Linac patient data. Results across two patients show Dice similarity coefficients of 0.84–0.89 maintained across fractions, with mean surface distances of 1.6–4.5 mm. A key finding is tumour migration of up to 19 mm between fractions — a shift large enough to compromise dose coverage — which automated tracking makes routinely detectable where manual delineation cannot. The longer-term aim is a clinical alert system that flags volumetric or positional changes exceeding predefined thresholds, supporting the clinical team in deciding whether plan adaptation is warranted.

• A World Model for Adaptive Radiotherapy of Glioblastoma on the MR-Linac active

  Glioblastoma's poor prognosis reflects in part the limitations of fixed treatment plans applied to a tumour that evolves throughout the radiotherapy course. The MR-Linac makes visible what was previously hidden — fraction-by-fraction changes in tumour volume, morphology, and position — but current planning tools remain reactive, working only from today's anatomy with no capacity to anticipate how the tumour will respond to alternative treatment decisions. This project develops a world model — a learned simulator trained on longitudinal GBM MRI data — that takes the current tumour state observed on the MR-Linac and a proposed treatment action (dose, fractionation) as input, and predicts tumour appearance at the next fraction without delivering the treatment. This enables counterfactual reasoning over competing adaptive plans and, eventually, training of a reinforcement learning agent to discover optimal fraction-by-fraction strategies. The model is trained on the LUMIERE dataset (91 GBM patients, 638 longitudinal studies), with tumour state represented by volume, segmentation mask, diffusion signal, and MGMT methylation status. A first prototype demonstrates plausible prediction of the contrast-enhancing tumour region at the next timepoint. Key open questions include data volume requirements for generalisation, validation standards for clinical utility, and how to formalise treatment outcome as an optimisation objective.