Optimization of Radiotherapy Treatment Plans Based on Monte Carlo Dose Computations

Sammanfattning: Treatment planning plays a vital role in providing good treatment to cancer patients. In order to reach an adequate treatment plan, the algorithm used for simulating the dose in the patient must model the reality well. The most accurate algorithm for this is the Monte Carlo method, which in the context of radiotherapy treatment most often is used for pre-computing spot doses and optimizing the intensity for each spot. Due to storage limitations, it would be preferred to do the mentioned computations simultaneously as the optimization. This, however, makes the optimization problem non-convex due to the statistical noise introduced by Monte Carlo. This thesis investigates the feasibility of using first-order optimization methods for treatment planning based on Monte Carlo simulations and addresses the challenges posed by the noise. A simplified proton Monte Carlo dose engine was implemented together with a matching analytical such, in order to assess the effect of the noise during optimization. The results demonstrate that despite the noise, an adequate treatment plan can be achieved. Convergence is found to be dependent on how simulations are used within an iteration. Techniques such as accumulating total dose and computing the gradient and Hessian separately show promise for improving convergence and time efficiency, respectively. The impact of noise on error computation and the need for appropriate comparisons are highlighted. This work provides insights for advancing Monte Carlo treatment planning and its integration into clinical settings. The findings are applicable not only to proton treatment plans but also to other ions and perhaps even to photons.