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

Background

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.

Aims

• Develop a world model — a learned simulator — that predicts tumour appearance at the next treatment fraction given the current tumour state and a proposed treatment action (dose, fractionation).
• Enable counterfactual reasoning over competing adaptive plans without needing to deliver treatment.
• Train a reinforcement learning agent to discover optimal fraction-by-fraction adaptive treatment strategies.

Methods

The world model is trained on the LUMIERE dataset (91 GBM patients, 638 longitudinal MRI studies). Tumour state is represented by volume, segmentation mask, diffusion signal, and MGMT methylation status. The model takes the current observed state and a proposed treatment action as input and predicts the tumour’s appearance at the next timepoint.

Outcomes

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.