文献速递胸部肿瘤放疗放疗物理与技术

大语言模型引导的模块化多智能体强化学习框架:提升肺癌放疗计划质量与效率

A Modular Multi-Agent Reinforcement Learning Framework Guided by Large Language Models: Improving Quality and Efficiency of Treatment Planning in Lung Radiation Therapy.
PubMed 文献 · Int J Radiat Oncol Biol Phys · 2026-09-17

该研究提出一种由大语言模型(LLM)监督的模块化强化学习(RL)框架,用于自动化肺IMRT计划设计。在62例局部晚期NSCLC回顾性病例中,该框架与KBP及纯LLM系统对比,所有LLM引导系统临床目标达成率98%(KBP为74%),混合框架在需KBP后优化的病例中以更少LLM调用和更短时间达标,且肺保留模式可进一步降低肺V20和平均肺剂量。

要点

  • 混合LLM-RL框架在62例NSCLC中临床目标达成率98%,KBP为74%。
  • 需KBP后优化的16例中,混合框架LLM调用1.7±0.7次、耗时20.6±15.7分钟。
  • 肺保留模式较常规模式降低肺V20(21.3±8.4% vs 25.2±10.1%,P<.001)。

核心结论

  • 模块化RL优化与LLM监督结合,较纯LLM自动计划改善肺保护并提升计划效率。

研究设计与样本

  • 采用两级分层架构:任务特异性RL智能体执行剂量修改,LLM监督层(Planner、Evaluator、Supervisor)协调。
  • 模块化并行OAR、串行OAR和靶区覆盖智能体使用SARSA与线性函数逼近训练,类别级权重共享实现跨OAR泛化。
  • 在62例回顾性局部晚期NSCLC病例上评估两种混合配置,对比机构KBP模型和纯LLM系统。

主要发现

  • 所有3种LLM引导系统临床目标达成率均为98%,KBP单独为74%。
  • 16例需KBP后优化的病例中,混合框架LLM监督调用次数较纯LLM系统减少75%(1.7±0.7 vs 6.7±5.7),时间更短(20.6±15.7 vs 33.4±32.1分钟),最差病例从109.8分钟降至58.5分钟。
  • 肺保留模式较常规模式降低肺V20(21.3±8.4% vs 25.2±10.1%;P<.001)和平均肺剂量(14.2±4.4 Gy vs 15.0±4.8 Gy;P<.001),脊髓和食管最大剂量小幅增加但仍在约束内。
  • 最终计划肺剂量低于同患者临床实际计划(V20 21.3% vs 24.6%;平均剂量14.2 Gy vs 14.9 Gy)。

关键数值与推荐

  • 使用机构剂量体积约束评估临床目标达成率。
  • 肺保留模式在完全满足约束后可选进入,通过单步LLM监督迭代降低肺剂量。

临床怎么用

  • 该框架可集成到自动计划流程中,减少人工优化和LLM监督调用,提升肺IMRT计划效率与肺保护。
  • 类别级权重共享降低训练集需求,便于跨OAR规模化部署。

局限性

  • 研究为回顾性单机构评估,样本量62例,需外部验证。
  • 肺保留模式可能小幅增加脊髓和食管最大剂量,需临床权衡。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

PURPOSE: Knowledge-based planning (KBP) frequently requires manual refinement to satisfy institution-specific clinical constraints, while existing automated approaches based on large language models (LLMs) or reinforcement learning (RL) each face distinct limitations in long-term optimization awareness and scalability. We propose a hybrid LLM-guided modular RL framework that integrates the strengths of both approaches for automated lung intensity modulated radiation therapy treatment planning. METHODS AND MATERIALS: The framework employs a 2-level hierarchical architecture in which task-specific RL agents execute dose modifications under the coordination of an LLM supervisory layer comprising Planner, Evaluator, and Supervisor agents. Modular parallel organ-at-risk (OAR), serial OAR, and target coverage agents are trained using state-action-reward-state-action with linear function approximation, with category-level weight sharing enabling cross-OAR generalization from a minimal training set. On achieving full constraint compliance, the system optionally enters Lung Sparing Mode, in which the lung-RL agent continues to reduce lung dose through single-step LLM-supervised iterations. The 2 hybrid configurations using the proposed framework were evaluated on 62 retrospective locally advanced non-small cell lung cancer cases against an institutional KBP model and an LLM-only system using institutional dose-volume constraints. RESULTS: All 3 LLM-guided systems achieved a 98% clinical goal achievement rate, compared with 74% for KBP alone. Among the 16 cases requiring post-KBP refinement, the hybrid framework reached the clinical goal with 75% fewer LLM supervisory calls than the LLM-only system (1.7 ± 0.7 vs 6.7 ± 5.7) and in less time (20.6 ± 15.7 vs 33.4 ± 32.1 minutes), with the worst case reduced from 109.8 to 58.5 minutes. Enabling Lung Sparing Mode further reduced lung V20 (21.3 ± 8.4% vs 25.2 ± 10.1%; P < .001) and mean lung dose (14.2 ± 4.4 vs 15.0 ± 4.8 Gy; P < .001) relative to Normal mode, at the cost of small increases in spinal cord and esophagus maximum dose that remained well within constraints. The resulting plans achieved lower lung dose than the clinically delivered plans for the same patients (V20 21.3% vs 24.6%; mean 14.2 vs 14.9 Gy). CONCLUSIONS: Integrating modular RL optimization with LLM-based supervision improves both lung sparing and planning efficiency relative to LLM-only automated planning, while enabling scalable deployment through cross-OAR weight sharing with reduced training requirements.

来源:Int J Radiat Oncol Biol Phys,查看原文。版权归原作者所有。

引用

  1. Wang Z, Guo H, Lei Y, et al. A Modular Multi-Agent Reinforcement Learning Framework Guided by Large Language Models: Improving Quality and Efficiency of Treatment Planning in Lung Radiation Therapy. Int J Radiat Oncol Biol Phys. 2026. PMID: 42754176. doi:10.1016/j.ijrobp.2026.09.008

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。