Reinforcement learning for real-time adaptive radiotherapy

State-of-the-art radiotherapy machines with integrated magnetic resonance (MR) imaging, known as MR-Linacs, provide the capability to track tumors in real time. This capability aids delivery of precise irradiation in the presence of patient motion, such as breathing, by adjusting the radiation beam. However, current solutions rely solely on geometric tracking without closing the loop by considering the actual endpoint-the delivered radiation (here called dose). Real-time dose-based adaptation within a single session remains highly challenging due to the immense dimensionality of the problem. To overcome this, we have developed a radiotherapy simulator and propose a novel reinforcement learning (RL)-based approach for real-time adaptive radiotherapy using 2D fluence, as a surrogate to 3D dose. To our knowledge, this is the first application of RL in real-time adaptive radiotherapy. Our in-silico experiments showed the feasibility of using RL to close the feedback loop, dynamically adapting to patient motion and minimizing discrepancies between delivered and intended dose in clinical cases. Our approach introduces a new treatment delivery paradigm, enabling delivery based on a reference fluence and motion without predefined machine settings.