A real-time framework for image-guided adaptive radiotherapy

电离辐射是癌症治疗的一种重要方式，其目的是提供高度局部的体积杀瘤剂量，同时限制对周围正常组织的毒性。通常，调强放疗（IMRT）计划过程中的治疗参数经过优化，并在预先计划的过程中提供，可以从几天延长到数周。然而，由于软组织变形、肿瘤消退和器官运动导致的日常解剖变化，需要频繁监测患者并进行适当调整，以确保有效性并降低暴露风险。尽管机载成像技术取得了技术进步，但鲁棒性和计算效率高的算法仍然严重滞后，无法实现这些技术的最佳利用，例如能够实时地将预处理计划图像配准到日常扫描中，使剂量变形，并且仍然可以重新优化计划以满足原始治疗多目标。简单的权衡方法已经提出，可能导致有害的陷阱。为了克服这些挑战，我们提出了一种基于统计学习技术的自适应反馈框架。这里的关键思想是保留先前IMRT计划解决方案的最优性条件，同时“绝热”（以最小的损失或增益扰动）修改当前解决方案。具体来说，我们计划在流处理器上设计和实现在线可变形图像配准和IMRT优化，使用快速统计增量学习方法，旨在使IMRT计划实时适应“日常解剖”，并将这种方法扩展到运动补偿交付。我们将根据放射生物学和剂量学原理推导和评估适应决策的新标准。我们将使用来自不同癌症部位的幻影和研究案例严格评估所提出的治疗计划适应反馈系统的性能。从技术上讲，这一提议将带来动态在线优化、可变形图像跟踪、运动预测和决策制定的新方法，同时在临床上允许最佳的个性化治疗计划适应和改善结果。

Ionizing radiation is an important modality for cancer treatment, which aims to deliver a highly localized volumetric tumoricidal doses while limiting toxicity to surrounding normal tissues. Typically, treatment parameters in an Intensity Modulated Radiotherapy (IMRT) planning process are optimized and delivered over a pre-planned course that can extend from few days to multiple weeks. However, daily changes in anatomy due to soft-tissue deformation, tumor regression, and organ motion require frequent monitoring of patients and proper adjustment to ensure effectiveness and reduce exposure risks. Despite technological advances in on-board imaging technologies, robust and computationally efficient algorithms are still severely lagging to achieve the optimal utilization of these technologies such as the ability in real-time to register the pre-treatment planning images to the daily scans, deform the dose, and still re-optimize the plan to satisfy the original treatment multi-objective goals. Simple trade-off methods have been proposed that can lead to detrimental pitfalls. To overcome these challenges, we propose a new adaptive feedback framework based on statistical learning techniques. The key idea here is to retain the optimality conditions of the previous IMRT plan solution while "adiabatically" (perturbing with minimal loss or gain) modifying the current solution. Specifically, we plan to design and implement on stream processors online deformable image registration and IMRT optimization using fast statistical incremental learning approaches that aims to adapt IMRT plans to 'anatomy of the day' in real-time and extend this approach to motion-compensated delivery. We will derive and evaluate new criteria for adaptation decision-making based on radiobiological and dosimetric principles. We will rigorously evaluate the performance of the proposed feedback system for treatment planning adaptation using phantoms and study cases from different cancer sites. Technically, this proposal will result in new methods for dynamical online optimization, deformable image tracking, motion prediction, and decision making while clinically allowing optimal personalized treatment planning adaptation and improved outcomes.