The success of radiotherapy relies on precisely targeting cancerous cells while minimising harm to healthy tissue. Real-time adaptive radiotherapy addresses the challenge of rapidly changing cancer positions, dynamically adjusting the radiation beam during treatment with the help of MR-linacs—integrating Magnetic Resonance Imaging (MRI) scanners with linear accelerators. However, this approach faces significant computing challenges. The process involves converting raw scanner signals into MRI images, transforming them into a patient model, reconstructing the delivered dose, and optimising treatment parameters. These computationally intensive steps must be executed quickly to ensure the effectiveness of real-time adaptive radiotherapy.

The first contribution of the thesis introduces a novel parallelisation scheme tailored for Four Dimensional Magnetic Resonance Imaging (4D-MRI) data, effectively minimising multithreading overhead and concealing data transfer latency. This results in an efficient architecture for algorithms based on the golden angle stack-of-stars MRI reconstruction. One such algorithm has been implemented and accelerated using Graphics Processing Units (GPUs) to facilitate clinical applications.

The second contribution aims to provide a methodology for predicting the performance of accelerators in the context of Monte Carlo (MC) simulations. The methodology is derived from GPU and Field Programmable Gate Array (FPGA) computational capabilities through benchmarks. It is then validated with Coulomb Scattering, assessing the accuracy of the forecasts. The results demonstrate that the model’s forecasted speedup assists in deciding the optimal accelerator, if any, to achieve the best performance.

The third contribution of this thesis is Optimised Dose Planning Method (ODPM), an MC simulation for radiotherapy dose reconstruction designed to meet real-time adaptive radiotherapy time requirements. The use of accelerators for ODPM is analysed using performance modelling. The forecasted speedup is then used to design an architecture capable of meeting real-time adaptive radiotherapy time requirements. This contribution also introduces multiple mathematical, algorithmic, and architectural optimisations applicable to MC simulations and integrates them into ODPM.