Human-like automated radiotherapy treatment planning via imitation learning

项目总结 放射治疗是癌症治疗的主要方法之一。治疗计划，这一过程 为每个患者设计最优的治疗方案，是最关键的步骤之一。如果治疗效果不佳 无论其他治疗步骤的质量如何，都无法达到令人满意的结果。 现代放射治疗中的治疗计划被表述为一个数学优化问题，其定义如下 一组超参数。虽然有几个可量化的指标来量化计划质量并指导 计划过程中，这些都是简化的表述，不能完全描述医生的意图。此外, 这些指标仅从基于人口的角度衡量计划质量，不能指导治疗 计划实现针对患者的最佳治疗计划。因此，医生首选的最佳解决方案通常 处于灰色地带，只有通过广泛的试错超参数调整过程才能实现，并且 规划者和医生之间的互动。因此，复杂的计划可能需要长达一周的时间 如果计划员缺乏经验和/或在严格的时间限制下，案例和计划质量可能会很差。这些 后果大大恶化了治疗结果，临床上已经清楚地证明了这一点 学习。最近，人工智能(AI)的进步，特别是模仿学习的进步，使人类-- 喜欢通过观察人类专家的行为并在内部建立自己的决策来做出决策 系统。作为对PAR-18-530的响应，该项目的目标是开发和翻译一个模仿 人类专家的行为来生成高质量的计划。人工智能规划者不会取代人类规划者。相反， 人工智能计划将作为当前规划过程的起点，以提高规划质量和规划 效率。人工计划员对进一步计划改进的行动可以通过以下方式反馈给AI计划员 为其不断进化而不断学习。我们将以前列腺癌作为试验床来追求这一目标 通过学术和产业合作，将德克萨斯大学西南分校强大的研究和临床专业知识结合在一起 在瓦里安医疗系统公司拥有丰富商业产品开发经验的医疗中心 确定了以下具体目标。目标1：模型和算法开发。我们将收集专家的行为 常规治疗计划中的数据和培训人工智能计划员。目标2：系统验证和翻译。我们会 将人工智能计划器集成到瓦里安日蚀治疗计划系统中，并在临床上验证了该系统 逼真的环境。这些创新包括使用最先进的人工智能模仿学习算法来解决 临床上重要的问题，由开发的系统实现的新技术能力，以及 连贯的翻译活动，为最终用户提供新功能。交付能力由广泛的 初步研究和整合互补专门知识和资源的伙伴关系。临床翻译 AI规划师的出现将通过提供高质量和高效的治疗来给放射治疗带来实质性的影响 规划造福患者，特别是资源有限地区的患者。

PROJECT SUMMARY Radiation therapy is one of the major approaches for cancer treatment. Treatment planning, the process of designing the optimal treatment plan for each patient, is one of the most critical steps. If a treatment is poorly designed, a satisfactory outcome cannot be achieved, regardless of the quality of other treatment steps. Treatment planning in modern radiotherapy is formulated as a mathematical optimization problem defined by a set of hyperparameters. While there exists several quantifiable metrics to quantify plan quality and guide the planning process, these are simplified representations that cannot fully describe the physician’s intent. In addition, these metrics only measure plan quality from a population-based perspective, and cannot guide treatment planning to achieve the patient-specific best treatment plans. Hence, the best physician-preferred solution often sits in a gray area, only achievable by an extensive trial-and-error hyperparameter tuning process and interactions between the planner and physician. Consequently, planning time can take up to a week for complex cases and plan quality may be poor, if the planner is inexperienced and/or under heavy time constraints. These consequences substantially deteriorate treatment outcomes, as having been clearly demonstrated in clinical studies. Recently, the advancement in artificial intelligence (AI), particularly in imitation learning allows human- like decision making by observing a human expert’s actions and internally building its own decision-making system. In response to PAR-18-530, the goal of this project is to develop and translate an AI planner that mimics human experts’ behavior to generate a high quality plan. The AI planner will not replace human planners. Instead, the AI plan will be used as a starting point in the current planning process to improve plan quality and planning efficiency. The human planner’s actions on further plan improvement can feed back to the AI planner through continuous learning for its continuous evolution. We will pursue this goal using prostate cancer as the test bed through an academic-industrial partnership, jointing strong research and clinical expertise at UT Southwestern Medical Center with extensive commercial product development experience at Varian Medical Systems Inc. The following specific aims are defined. Aim 1: Model and algorithm development. We will collect experts’ behavior data in routine treatment planning and train the AI planner. Aim 2: System validation and translation. We will integrate the AI planner into Varian Eclipse treatment planning system and validate the system in a clinically realistic setting. The innovations include the use of a state-of-the-art AI imitation learning algorithm to solve a clinically important problem, the novel technological capabilities enabled by the developed system, as well as coherent translation activities to deliver new capabilities to end users. Deliverability is ensured by extensive preliminary studies and the partnership integrating complementary expertise and resources. Clinical translation of the AI planner will bring substantial impacts to radiotherapy by providing high-quality and efficient treatment planning to benefit patients, especially those in resource-limited regions.