Machine Learning Assisted Decision Support Platform for Radiation Treatment Assessment

放射治疗癌症是一个复杂的过程，需要临床和技术专长。虽然放射治疗是安全有效的，但在治疗的设计和实施方面存在可能导致患者不良结局的错误。放射肿瘤学目前的质量保证过程需要大量的资源，利用放射医学团队的集体经验，以确保质量和安全，同时最大限度地减少错误的可能性。我的实验室以前开发了一种新的概念验证机器学习（ML）框架，用于自动化治疗质量审查过程，该框架将专家放射医学团队的共享知识编入法典，以根据成像，解剖，技术和剂量特征优先考虑复杂病例。ML提供了增强人类任务的能力，并将临床决策集成到临床过程中，这构成了拟议研究计划的基础。使用ML来预测放射治疗的质量并了解具有潜在异常特征的放射治疗的能力可以用于1）加速同行评审过程，2）通过标记更可能需要注意或错误的放射治疗来确保更高质量的治疗，以及3）通知需要实时临床决策的复杂过程的临床管理。目的：拟议研究计划的总体假设是，ML辅助的治疗质量见解将是放射肿瘤学过程的重要组成部分，以确保患者接受高质量和及时的放射治疗。拟议的研究计划将开发一个基于新型ML方法的高质量放射肿瘤学平台，其目标如下：i）构建和部署一个专用的基于ML的质量平台，以根据治疗复杂性对放射治疗进行优先排序，供放射医学专家团队进行审查，并在治疗批准之前突出需要注意的潜在错误的治疗。 ii）开发用于自适应放射治疗的实时质量平台，该平台在治疗过程中结合时间序列成像数据和解剖信息，以确保根据患者不断变化的解剖结构对治疗进行适当的更新。结果和意义：拟议研究计划的新技术发展包括：i）生成具有人类可理解输出的新离群值检测模型，以在同行评审中为治疗评审提供ML辅助决策支持，并结合时间序列数据为自适应放射治疗提供实时决策支持。拟议的研究对于改善放射肿瘤学质量流程的工作流程具有直接的适用性和意义，除了使患者能够更及时地获得复杂的治疗外，还能够改善癌症护理。

The delivery of radiation for the treatment of cancer is a complicated process that requires both clinical and technical expertise. Although radiation treatments are safe and effective, there is a potential for errors in how treatments are designed and delivered that could lead to adverse patient outcomes. The current quality assurance process in radiation oncology requires substantial resources that leverages the collective experience of the radiation medicine team to ensure quality and safely, while minimizing the likelihood of errors. My lab has previously developed a novel proof-of-concept machine learning (ML) framework for automating the treatment quality review process that codifies the shared knowledge of the expert radiation medicine team to prioritize complex cases based on imaging, anatomical, technical, and dosimetric features. ML provides the ability to augment human tasks as well as integrate clinical decision making into the clinical process which forms the basis of the proposed research program. The ability to use ML to predict the quality of radiation treatments and understand radiation treatments which have potentially anomalous features can be used to 1) expedite the peer review process, 2) ensure higher quality treatments by flagging radiation treatments that are more likely to require attention or be erroneous, and 3) inform clinical management for complex processes that require real-time clinical decisions. Objectives: The overarching hypothesis of the proposed research program is that ML-assisted insights of treatment quality will be essential components of the radiation oncology process to ensure patients receive high quality and timely radiation treatments. The proposed research program will develop a quality radiation oncology platform built on novel ML methods with the following objectives: i) Build and deploy a dedicated ML-based quality platform to prioritize radiation treatments for review by the expert radiation medicine team based on treatment complexity and highlight treatments with potential errors requiring attention prior to treatment approvals. ii) Develop a real-time quality platform for adaptive radiation treatments that incorporates time-series imaging data and anatomical information over the course of treatment to ensure updates to treatments are appropriate based on the patients' changing anatomy. Outcomes and Significance: The novel technical developments from the proposed research program include: i) generating new outlier detection models with human understandable output to enable ML-assisted decision support for treatment review in peer review rounds and incorporating time-series data to provide real-time decision support for adaptive radiation treatments. The proposed research has direct applicability and significance for improving the workflow of quality processes in radiation oncology and enabling improved cancer care in addition to enabling patients more timely access to complex treatments.