Dose-distribution radiomics to predict morbidity risk in radiotherapy

 描述(申请人提供)：肺癌仍然是美国癌症死亡的主要原因，对于当地晚期、无法手术的病例，最终放射治疗是治疗的标准。放射治疗也是高度灵活的，可以适应肿瘤体积和位置的变化。然而，缺乏决策支持模型来优化治疗计划，避免在逐个患者的基础上发生发病率。最近，随机第三阶段试验RTOG 0617对非小细胞肺癌(NSCLC)的高剂量(74GY)治疗和标准(60GY)治疗进行了比较，结果显示高剂量组的死亡率出人意料地高。我们推测，这主要是由于心脏照射和肺照射因素共同作用造成的。该项目的目标是确定和验证胸部RT术后发病率的预测模型，以便在局部控制和发病率之间实现更安全和更优化的权衡。我们将使用创新的图像配准、机器学习和剂量特征(基于剂量的放射组学)联合分析三个大型、高质量的数据集，以开发可用于预测和避免放射治疗计划中关键发病率终点的数学模型。所有终点将使用剂量分布比较方法来检查剂量和发病率之间的关系作为解剖位置的函数，从而为正确识别放射治疗毒性的来源提供更大的保证。在特定目标1(SA1)下，我们将使用创新的可变形图像配准方法将所有患者的剂量分布映射到分段的参考病例解剖(包括心脏、肺、支气管炎和血管)。将使用一种创新的非参数估计方法仔细绘制变形的准确性地图。在SA2和SA3下，我们将提取与发病率高度相关的组织区域用于机器学习。在SA2下，我们将应用创新的机器学习，目标是将SA1下提取的剂量因素与治疗相关的死亡风险联系起来。根据SA3，我们将对严重肺炎的风险采用相同的方法。我们假设与治疗相关的死亡危险和放射性肺炎形成了一个单一的风险增加的可预测连续体，我们将尝试将SA2和SA3的结果结合在一个单一的预测模型中。创新的互斥预测因素的建模后分析将用于使机器学习模型具有可解释性。这笔赠款的总体目标是：(1)促进个人优化放射肿瘤学领域的科学状况，以及(2)提供决策支持工具，以指导肺癌放射治疗中局部控制和毒性之间的权衡。

 DESCRIPTION (provided by applicant): Lung cancer remains the leading cause of cancer deaths in the U.S., and definitive radiotherapy is the standard of care for locally advanced, inoperable cases. Radiation therapy is also highly flexible, and can be adapted for variations in tumor volume and location. However, decision support models to optimize treatment planning and avoid morbidity on a patient-by-patient basis are lacking. Recently, the randomized phase III trial, RTOG 0617, compared high dose (74 Gy) treatments to standard (60 Gy) treatments for non-small cell lung cancer (NSCLC), and demonstrated an unexpectedly higher rate of mortality in the high-dose arm. We hypothesize that this was primarily due to heart irradiation, in combination with lung irradiation factors. The goal of this project is to determine and validate predictive models of morbidity following thoracic RT, in order to achieve safer and more optimal tradeoffs between local control and morbidity. We will jointly analyze three large, high-quality datasets using innovative image-registration, machine learning, and dose characteristics (dose-based radiomics), to develop mathematical models that could be used to predict and avoid key morbidity endpoints in radiotherapy treatment planning. All endpoints will use dose-distribution comparison methods to examine the correlation between dose and morbidity as a function of anatomic location, thus providing greater assurance that the source of radiotherapy toxicity is correctly identified. Under Specific Aim 1 (SA1), we will use innovative deformable image registration methods to map all patient dose distributions to a segmented reference case anatomy (including heart, lungs, bronchii, and blood vessels). The accuracy of the deformations will be mapped carefully using an innovative non-parametric estimation methodology. Tissue regions with high correlations to morbidity will be extracted for machine learning under SAs 2 and 3. Under SA2, we will apply innovative machine learning with the goal of relating dose factors extracted under SA1 with treatment-related death hazard. Under SA3, we will apply the same methodology to the risk of severe pneumonitis. We hypothesize that treatment related death hazard and radiation pneumonitis form a single predictable continuum of increasing risk, and we will attempt to bring together the results of SA2 and SA3 in a single predictive model. Innovative post-modeling analyses of mutually exclusive predictors will be used to make the machine learning models interpretable. The overall goals of this grant are: (1) to advance the state of science in the field of personally-optimized radiation oncology, and (2) to provide decision support tools to guide patient-specific tradeoffs between local control and toxicity in lung cancer radiotherapy.