| 刘嘉雯,李永宝,李华莉,等.基于梯度增强群体智能算法的预后引导肺癌调强放疗计划自动优化方法[J].中华放射医学与防护杂志,2025,45(4):302-308.Liu Jiawen,Li Yongbao,Li Huali,et al.Automatic optimization of prognosis-guided intensity-modulated radiation therapy plans for lung cancer based on a gradient-enhanced swarm intelligence algorithm[J].Chin J Radiol Med Prot,2025,45(4):302-308 |
| 基于梯度增强群体智能算法的预后引导肺癌调强放疗计划自动优化方法 |
| Automatic optimization of prognosis-guided intensity-modulated radiation therapy plans for lung cancer based on a gradient-enhanced swarm intelligence algorithm |
| 投稿时间：2024-09-10 |
| DOI：10.3760/cma.j.cn112271-20240910-00343 |
| 中文关键词: 调强放疗 预后引导放疗 大规模非线性规划 群体智能算法 |
| 英文关键词:Intensity-modulated radiation therapy Prognosis-guided radiotherapy Large-scale nonlinear programming Swarm intelligence algorithm |
| 基金项目:国家自然科学基金(82472117);广东省基础与应用基础研究基金(2024A1515011831,2024A1515010820) |
|
| 摘要点击次数: 262 |
| 全文下载次数: 76 |
| 中文摘要: |
| 目的 针对预后引导调强放疗(IMRT)计划优化中的大规模非线性规划问题,提出了一种基于梯度增强的群体智能算法GradRCIPSO,以实现临床高效场景下的预后计划全局优化寻解。方法 GradRCIPSO优化算法的核心思想是通过让粒子同时学习群体交互信息和梯度信息实现在全局范围的快速收敛。其中,交互信息从群体中的优秀个体处获得,用于让粒子高效搜索整个解空间;而梯度信息表征最速下降方向,用以帮助粒子快速探索当前所在邻域。为评估方法的有效性,本研究选取10例非小细胞肺癌(NSCLC)IMRT计划,对其用GradRCIPSO生成预后引导的放疗计划后,将该预后计划和原临床计划的质量进行对比。此外,还分别用内点法、序列二次规划、积极集、梯度下降法和随机对比交互粒子群算法(RCIPSO)作为预后计划的优化引擎,并比较其和GradRCIPSO在优化效率、精度方面的优劣。结果 GradRCIPSO可以成功地为患者生成预后引导的放疗计划,所得计划相较原临床计划剂量统计水平相当,且预期放疗总风险从1.22(0.84, 1.51)降至0.93(0.80, 1.29)(z=2.81,P<0.01)。GradRCIPSO的求解质量稳定,求解精度优于4种梯度算法(z=2.80~2.81,P<0.01)。和未使用梯度增强的群体智能算法RCIPSO相比,GradRCIPSO可在求解质量相似(P>0.05)的同时,提升约3倍的优化速度。结论 建立了一种新的优化算法GradRCIPSO,并验证了该方法在预后引导IMRT计划优化中的可行性及性能,为预后引导肺癌调强计划设计的广泛临床应用奠定技术基础。 |
| 英文摘要: |
| Objective To address large-scale nonlinear programming challenges in optimizing prognosis-guided intensity-modulated radiation therapy (IMRT) plans, to propose gradient-enhanced random contrastive interaction particle swarm optimization (GradRCIPSO). This gradient-enhanced swarm intelligence algorithm aims to enable global optimization of prognostic treatment plans in clinically efficient scenarios. Methods The core concept of GradRCIPSO lied in achieving rapid global convergence by allowing particles to learn both swarm interaction and gradient information. Specifically, the interaction information was obtained from elite individuals in the swarm, enabling the particles to efficiently search the entire solution space, whereas the gradient information represents the direction of the steepest descent, enabling the particles to quickly explore the current neighborhood. To assess the effectiveness of the methodology, the IMRT plans for 10 cases of non-small cell lung cancer (NSCLC) were selected in this study. They were compared with the GradRCIPSO-generated prognosis-guided IMRT plans. Moreover, the interior-point method, sequential quadratic programming, active set, gradient descent method, and random contrastive interaction particle swarm optimization (RCIPSO) were employed as optimization engines and compared with GradRCIPSO in terms of optimization efficiency and accuracy. Results GradRCIPSO successfully generated clinically viable prognosis-guided IMRT plans with comparable dosimetric statistics to original plans, while significantly reducing predicted total radiotherapy risk from 1.22(0.84,1.51) to 0.93(0.80,1.29) (z=2.81,P<0.01). It demonstrated superior accuracy over the above four gradient-based method (z=2.80-2.81,P<0.01) and achieved threefold acceleration versus RCIPSO while maintaining equivalent solution quality(P>0.05). Conclusions The proposed GradRCIPSO demonstrates high feasibility and performance in optimizing prognosis-guided IMRT plans, laying the technical foundation for the broad clinical application of prognosis-guided IMRT plans for lung cancer. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|