Deep reinforcement learning for isocenter placement in Gamma Knife® radiosurgery

Abstract: Disposition of isocenters inside the brain tumor is of paramount importance for the quality of the Leksell Gamma Knife® radiosurgery. This work presents a novel approach based on deep reinforcement learning aimed at optimizing isocenter locations. In this proof-of-concept, tumors are treated in two dimensions, and the number of isocenters is fixed at a user-defined level. A great emphasis is placed on experimentation, including regularization, reduction of the convex optimization problem size, a comprehensive analysis of the reward function, learning in different environments, and tuning of hyper-parameters. Extensive use of visualization offers intuitive insights into the problem and proposed solutions. Deep Q-network agents coupled with a reinforcement learning environment of our design demonstrate efficient learning and achieve presumably optimal results on training targets. An attempt at training agents capable of positioning isocenters in previously unseen targets yields sub-optimal results and offers informed suggestions for improvements based on empirical results.