Reinforcement learning applied to MLC tracking

Sammanfattning: Radiotherapy has become an ever more successful treatment option for cancer.Advances in imaging protocols combined with precise therapy devices suchas linear accelerators contribute towards millimeter precision of treatmentdelivery with far fewer side effects. The ultimate goal of radiotherapy is tomaximize tumor control while minimizing adverse effects to healthy tissues,more importantly organs at risk surrounding the tumor. External beamradiotherapy is currently on the brink of breaking a new frontier: MagneticResonance Imaging (MRI) guided tumor tracking. Here, a combined linearaccelerator and MRI system can be used to treat and follow the tumor duringirradiation, called Real-time Adaptive Radiotherapy (ART). Tailoring of thebeam shape, by means of the Multi-leaf Collimator (MLC) on the fly has thepotential to complete a fully automated radiotherapy process. Recent advances in Reinforcement Learning (RL), a sub field of artificialintelligence has pushed the frontiers further in sequential decision making processesfurther in various fields. In a MLC tracking scenario, we hypothesizethat an RL agent trained on real-time tumor delineations and dose informationcould fulfill a specified dosimetric criteria on the fly over the moving target.To investigate the feasibility of RL for MLC tracking further: we designeda simulator, devised an appropriate RL framework and interfaced them to aDeep Q-Network (DQN) algorithm. Our results demonstrate the feasibility of employing RL for MLC trackingalong with numerous design choices that need to be considered while developingsuch a system. We believe to have taken the first step to bridge MLCtracking and RL by proposing a closed loop solution using dose information.