Using deep reinforcement learning to facilitate automated adaptive radiation therapy planning: an initial study

Photon radiotherapy is widely used for the treatment of a huge range of cancers, and adaptive radiotherapy has the potential to improve treatment effectiveness while reducing radiation exposure to organs at risk (OARs). However, current adaptive radiotherapy workflows rely heavily on repeated manual replanning by radiation oncologists, which is inefficient and subject to inter-operator variability. This study aims to develop an AI-based automatic planning framework for radiotherapy treatment planning. To this end, we propose a convolutional deep Q-network (DQN) that iteratively refines treatment plans based on the current multileaf collimator (MLC) configuration and the corresponding estimated dose distributions to the target and OARs. At each step, DQN predicts an action that updates the MLC settings, resulting in incremental dose corrections toward the desired distribution. Through sequential decision-making, the proposed approach enables automated plan refinement with reduced reliance on manual intervention. Retrospectively collected clinical head and neck cancer patient data were used for model optimization and performance evaluation. The results demonstrate that the deep reinforcement learning-optimized (DRL-optimized) tool can effectively generate a treatment plan. Quantitative evaluation demonstrates that the DRL-optimized plan substantially improved target coverage and homogeneity, with increased V95% and median target dose. Meanwhile, peak and near-maximum spinal cord doses were reduced, indicating improved organ-at-risk sparing under the same geometry and beam model. Collectively, our results demonstrate that the proposed approach enables reliable automated plan refinement under standardized clinical conditions and establishes a scalable framework for future extensions to multi-field geometries, multi-organ constraints, and large-scale clinical validation.