Department of Radiation Oncology, Stanford University, Stanford, CA 94305-5847, United States of America.
Phys Med Biol. 2018 Jul 2;63(13):135014. doi: 10.1088/1361-6560/aaca17.
An important yet challenging problem in LINAC-based rotational arc radiation therapy is the design of beam trajectory, which requires simultaneous consideration of delivery efficiency and final dose distribution. In this work, we propose a novel trajectory selection strategy by developing a Monte Carlo tree search (MCTS) algorithm during the beam trajectory selection process. To search through the vast number of possible trajectories, the MCTS algorithm was implemented. In this approach, a candidate trajectory is explored by starting from a leaf node and sequentially examining the next level of linked nodes with consideration of geometric and physical constraints. The maximum Upper Confidence Bounds for Trees, which is a function of average objective function value and the number of times the node under testing has been visited, was employed to intelligently select the trajectory. For each candidate trajectory, we run an inverse fluence map optimization with an infinity norm regularization. The ranking of the plan as measured by the corresponding objective function value was then fed back to update the statistics of the nodes on the trajectory. The method was evaluated with a chest wall and a brain case, and the results were compared with the coplanar and noncoplanar 4pi beam configurations. For both clinical cases, the MCTS method found effective and easy-to-deliver trajectories within an hour. As compared with the coplanar plans, it offers much better sparing of the OARs while maintaining the PTV coverage. The quality of the MCTS-generated plan is found to be comparable to the 4pi plans. Artificial intelligence based on MCTS is valuable to facilitate the design of beam trajectory and paves the way for future clinical use of non-coplanar treatment delivery.
在基于 LINAC 的旋转弧形放射治疗中，一个重要但具有挑战性的问题是束流轨迹的设计，这需要同时考虑输送效率和最终剂量分布。在这项工作中，我们通过在束流轨迹选择过程中开发蒙特卡罗树搜索（MCTS）算法，提出了一种新的轨迹选择策略。为了搜索大量可能的轨迹，实现了 MCTS 算法。在这种方法中，从叶节点开始，通过考虑几何和物理约束，依次检查下一级链接节点，探索候选轨迹。最大上置信界限树（UCB）是平均目标函数值和测试节点被访问次数的函数，用于智能地选择轨迹。对于每个候选轨迹，我们运行具有无穷范数正则化的逆通量图优化。然后，根据相应的目标函数值对计划进行排名，并将反馈信息用于更新轨迹上节点的统计信息。该方法在胸壁和脑病例中进行了评估，并与共面和非共面 4π 射束配置进行了比较。对于这两种临床情况，MCTS 方法在一个小时内找到了有效且易于输送的轨迹。与共面计划相比，它在保持 PTV 覆盖的同时，更好地保护了 OAR。发现 MCTS 生成的计划的质量与 4π 计划相当。基于 MCTS 的人工智能对于促进束流轨迹的设计很有价值，并为未来非共面治疗的临床应用铺平了道路。
