基于预测剂量约束引导的自动多目标调强放疗计划优化方法研究

Intensity modulated radiotherapy plays an important role in cancer treatment. Its planning efficiency and plan quality are the two main factors for evaluating a treatment planning system, as their significant impact on patient’s clinical efficacy. Optimizing both beam directions and intensity profiles under a condition of without knowing appropriate dosimetric constraint and multi-objectivity is a complex problem that cannot be solved directly. As a matter of fact, planners have to manually tweak dosimetric objectives repeatedly to achieve an optimal plan, thus limiting planning efficiency and besides, plan quality can vary significantly among institutions and even planners. To these questions, we first build a geometry-dosimetry correlation model using a machine learning method to predict appropriate dosimetric constraints; secondly propose a novel multi-criteria optimization model by adjusting constraints sequentially according to its clinical emphasis under aforementioned predicted initial constraints; afterwards, incorporate a voxel-independent based automatic FMO optimization algorithm and use GPU technique to accelerate the entire process. Thus to develop an automated multi-criteria optimization for intensity modulated radiotherapy, laying a foundation for developing fast and automatic treatment planning algorithm in the future.

调强放疗是恶性肿瘤放疗的重要手段，其计划设计的效率与生成计划的质量是决定计划系统性能的关键，同时也是影响患者临床疗效的重要因素。计划设计的核心是计划的优化，而当前优化存在着剂量约束在设计前未知、优化过程具有多目标性等特点，因此临床上只能采用人工试错的方式以期逼近最优计划，造成了计划质量具有显著差异以及计划设计过程效率低下等问题。对此，本项目拟针对调强放疗快速自动计划优化中的关键问题展开研究，首先，基于先验计划智能学习方法预测治疗计划的剂量学指征项，然后将该剂量学指征值作为初始约束条件，引导构建多目标计划的优化新模型，期间再耦合体素权重因子的自动调整机制，并采用GPU编程对整体进行加速，最终实现集自动约束设置与自动多目标优化为一体的快速自动多目标计划优化，为后续快速自动计划设计奠定基础，解决临床实际问题。

项目以建立并形成智能放疗计划设计与自动计划设计新解决方案为目标，建立了基于经验自动学习方法的三维剂量分布预测模型与配套的以剂量预测为引导的放疗计划逆向优化新方法。该方法首先应用人工神经元网络等对患者解剖结构与对应质优计划剂量特性进行关联，实现新患者计划设计前的剂量目标预测；然后根据预测目标形式设计特制的数值优化方法，有效利用所预测的剂量分布信息，同时保障输出计划的优质性。实验验证结果显示，所提出方法的预测精度小于5%。与此同时，提出的以该剂量学预测为引导的自动计划优化方法通过对传统优化目标函数进行改写或整合，可基本实现所生成计划的剂量满足甚至优于所预测的剂量分布，保证所生成计划优质性的同时大幅度降低人工试误的成本。此外，整个计划优化过程基本实现自动化，无需人工调整。该方法可有效改善现有临床存在的放疗计划设计效率不高、同质化程度较低的弊端。