Optimizing radiation therapy dose planification using deep reinforcement learning

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique

Abstract

Introduction and goal: Radiation therapy treatments are delivered in fractions to slowly destroy the tumour while avoiding severe side effects in surrounding healthy tissues. While conventional treatments deliver fractions of equal dose each day, recent studies have claimed that there could be an advantage in varying the fraction size during the treatment. The goal of this project is to discover such unconventionnal fractionation schedules by using deep reinforcement learning algorithms agents that will aim to maximise the probability of destroying tumours while minimising the probability of causing severe side effects to surrounding tissues. These agents will be trained in conjunction with a simulation of tumoural development in healthy tissue to find successful treatments. Materials and methods: A simulation of tumoural development in healthy tissue was implemented, based on a 50x50 grid on which healthy cells and cancer cells are modeled individually. The effect of radiation on these cells is simulated by using the LQ model that gives each cell a probability of surviving a certain dose. Deep reinforcement learning agents will receive observations of the simulation in the form of images that show densities of healthy cells and cancer cells on the grid, and will have to choose treatment parameters such as the dose with which to irradiate the tumour or the time between two fractions. Their will use the DQN and DDPG algorithms to learn to optimise the sum of the rewards that they receive after each action that they choose. Reward functions were designed to either find treatments that maximise the number of cancer cells killed and minimise the number of healthy cells killed, or to minimise the total dose used in the treatment. Results: The trained agents suggested treatments that achieved much better results on the simulation than a baseline based on a conventional radiation therapy schedule. However, the simulation was implemented to be fast and several effects and phenomenons that are important in radiobiology have been omitted from the model. It is therefore likely that the simulation underestimates the damage that could be caused by the treatments suggested by the agents on healthy tissues. Conclusion: The results are promising and show that the approach is worth pursuing. To further this work, a more complex simulation should be used. It should be able to more accurately represent the side effects caused by a treatment on healthy tissues and should be validated using clinical data.

Citations

• APA

Moreau, G. (2020). Optimizing radiation therapy dose planification using deep reinforcement learning. https://dial-mem.test.bib.ucl.ac.be/handle/2078.2/17556