Automated Object Contouring Methods & Software for Radiotherapy Planning

摘要 2015年，美国估计将发生1，658，370例新癌症病例，其中近三分之二将受到辐射 治疗（RT）。考虑到美国有超过2,300个RT中心，以及目前的器官轮廓系统， 风险（OAR）主要依赖于人工方法，因此生产软件有很大的商业机会 能够以高度自动化在医学图像中描绘OAR轮廓的系统 RT计划的实践。受我们在胸部和头颈部（H&N）身体方面的强大I期结果的鼓舞 与当前的行业系统相比，我们寻求的是准确性、效率和临床接受度。 我们的软件产品输出的轮廓显著超过现有系统。我们的总体目标是 第二阶段是将第一阶段开发的算法和原型软件推进到领先的商业应用中。 软件产品，并在全国多个医疗中心展示其功效， 人口。第二阶段的具体目标有三个方面：（1）进一步推进自动解剖识别 第一阶段的算法使用了先进的深度学习技术。(2)开发基于云的软件自动化 轮廓服务。(3)对新软件在H&N和胸部病例上进行临床评价。 目标1将分三个阶段实现：（a）自动化定义给定的身体区域的过程， 患者CT研究，目前在我们的系统中手动完成，通过虚拟地标的新概念， 深度学习技术(b)从当前的2个体素提高对象识别/定位精度， 通过使用虚拟仪器，将质量“好”的数据集从1体素转换为4-5体素，将质量“差”的数据集从2-3体素转换为2-3体素。 地标来学习对象关系。(c)通过组合对象定位方法来改进对象描绘 将深度学习技术应用于局部对象的附近， 在1 voxel内。目标2将通过开发基于云的软件即服务模型来实现 整合了算法的软件为了实现目标3，进行了一项涉及四名学术人员的评价研究， RT中心将负责评估 新软件为了评估有效性，将比较当前临床过程所用的塑形时间与 新的软件方法所花费的时间加上所需的任何手动调整。将通过以下方式评估准确度： 将软件输出与精心准备的地面真实轮廓进行比较。可接受性将由 进行盲态阅片员研究，其中放射肿瘤学家对软件给出可接受性评分（1-5） 生成的轮廓、地面真实轮廓和正常临床过程生成的轮廓，并比较 这些分数。 预期临床结局是显著改善塑形的临床效率/可接受性， 目前的做法。

Abstract In 2015, 1,658,370 new cancer cases are estimated to occur in the US, where nearly two-thirds will have radiation therapy (RT). Given that there are over 2,300 RT centers in the US, and current systems for contouring organs at risk (OARs) rely mostly on manual methods, there is a strong commercial opportunity for producing a software system that can contour OARs in medical images at a high degree of automation and for impacting current practice of RT planning. Encouraged by our strong Phase I results in thoracic and head and neck (H&N) body regions compared to current industry systems, we seek the accuracy, efficiency, and clinical acceptance of the contours output by our software product to significantly exceed those of existing systems. Our overall aim for Phase II is to advance the algorithms and prototype software developed in Phase I into a leading commercial software product, and demonstrate its efficacy in multiple medical centers across the country with diverse populations. Phase II specific aims are three-fold: (1) Further advance the automatic anatomy recognition algorithms from Phase I using advanced deep learning techniques. (2) Develop a cloud-based software auto contouring service. (3) Perform clinical evaluation of the new software on H&N and thoracic cases. Aim 1 will be accomplished in three stages: (a) Automating the process of defining the body region on given patient CT studies, which is currently done manually in our system, via a new concept of virtual landmarks using deep learning techniques. (b) Improving object recognition/ localization accuracy from the current 2 voxels for “good” quality data sets to 1 voxel and from 4-5 voxels for “poor” quality data sets to 2-3 voxels by using virtual landmarks to learn object relationships. (c) Improving object delineation by combining object localization methods with deep learning techniques applied to the vicinity of the localized objects to bring boundary distance accuracy within 1 voxel. Aim 2 will be achieved by developing a cloud-based Software-as-a-Service model to implement the software that incorporates the algorithms. To accomplish Aim 3, an evaluation study involving four academic RT centers will be undertaken to assess the efficiency, accuracy, and acceptability of the contours output by the new software. To assess efficiency, contouring time taken by the current clinical process will be compared to the time taken by the new software method plus any manual adjustment needed. Accuracy will be assessed by comparing software output to carefully prepared ground truth contours. Acceptability will be determined by conducting a blinded reader study, where an acceptability score (1-5) is given by radiation oncologists to software produced contours, ground truth contours, and contours produced by the normal clinical process, and comparing these scores. Expected clinical outcomes are significantly improved clinical efficiency/ acceptability of contouring compared to current practice.