Reinforcement Learning for Automatic Dose Control in Radio Therapy

Proton therapy is a treatment method for various cancer types. One of the challenges found while performing proton therapy is motion management. Especially the breathing motion of the patient poses a challenge. We propose an artificial intelligence (AI) trained with reinforcement learning (RL) to control the delivery of proton therapy and actively adapting to the breathing motion of the patient. The real-life environment of treatment delivery was simplified into a virtual 2D environment, and in a first step a static target was used to represent the tumour. We used an advantage actor-critic (A2C) model in combination with different neural networks (NN) to train on the virtual environment. The network which combined a simple convolutional neural network (CNN) with a long-short term memory cell (LSTMcell) performed better that the networks that only composed of a simple CNN. We were able to show that with an adequate network the agent was able to hit the target and that it was able to improve performance over time.