Robust Reinforcement Learning optimization of radiotherapy dose fractioning

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien, à finalité spécialisée

Abstract

Radiation therapy is an integral treatment option for cancer, its success hinging on the precise dosage delivery to the target and the preservation of surrounding healthy tissues. This master thesis examines the robustness of reinforcement learning (RL) optimization, a potential tool for automating radiation therapy dose fractionation, in response to varying parameters within cellular models. Specifically, we use the combined cellular model proposed by A. Jalalimanesh and al. and O'Neil and al., and manipulate key parameters such as nutrients consumption rate of cancer cells, radiation sensitivity, and cell cycle duration to observe their impact on optimization results. Employing RL algorithms like Q-learning, Sarsa, and Expected Sarsa, we optimize the radiotherapy dose fractionation schedule and compare their effectiveness under different parameter conditions. The goal of this study is to generate superior treatment plans concerning Tumor Control Probability (TCP), number of radiation fractions, total radiation dose, treatment duration, and healthy cell survival rate. We extensively test our RL agents' ability to navigate a wide range of scenarios. In benchmark tests, our agents outperformed the baseline treatment by an average of 51.1% across these metrics. Moreover, we incorporate observations of the patient during treatment for a more realistic approach. We've developed an ML-based adaptative decision-making framework that autonomously modifies treatment plans based on individual tumor images. Leveraging machine learning algorithms, we automated the decision-making process for treatment adaptation. This framework, in the benchmark tests, showed a 54.9% average improvement compared to the baseline treatment plan. Remarkably, this automated system demonstrated its prowess in crafting adapted treatment plans even in extraordinarily aggressive environments where standalone reinforcement learning algorithms failed to find any solutions. These advancements present opportunities for more personalized and effective cancer treatments, reducing patient morbidity and enhancing overall treatment efficacy.