Improving portability of knowledge-based planning using an LLM-driven plan refinement framework in lung radiotherapy.

backgroundKnowledge-based planning (KBP) has improved the quality and efficiency of radiotherapy treatment planning. However, its broader clinical adoption remains limited because effective deployment often requires institution-specific model training and tuning. Publicly available KBP models provide a convenient starting point but may not consistently meet local clinical objectives across institutions. PURPOSE: We developed and evaluated the Planning Copilot, a large language model (LLM)-guided plan refinement framework designed to operate as a model-agnostic post-processing layer for KBP.

methodsThe Planning Copilot is a closed-loop, multi-agent system that iteratively refines KBP-generated plans through structured dosimetric feedback and the selection of clinically validated optimization actions within a treatment planning system. For each case, an initial step-and-shoot IMRT plan was generated with each of three RapidPlan models, including a publicly available model and two institutional models with different optimization constraints. To assess whether the refinement depends on KBP, we additionally evaluated PlanningCopilot starting from a non-KBP fixed objective template applied identically to all cases. The PlanningCopilot was applied without model-specific tuning to 62 retrospective locally advanced NSCLC cases. Clinical goal achievement rates and clinically relevant dose-volume metrics were compared between the initial KBP plans and the refined plans.

resultsAcross all three KBP models, the PlanningCopilot substantially improved plan quality. Clinical goal achievement increased from 79% to 98% for the UCSD model, from 73% to 97% for Institutional T1, and from 69% to 98% for Institutional T2. Starting from the non-KBP fixed template, the achievement rate increased from 68% to 97%, comparable to the KBP initializations. Significant reductions were observed in key lung dose metrics, including lung Dmean across all models and lung V20 in the Institutional T1 model and template initializations, while target coverage and doses to critical structures were maintained. Notably, the KBP model that prioritized OAR sparing, which exhibited the lowest initial pass rate, showed the highest rescue rate after refinement.

conclusionsAn LLM-guided refinement layer can improve the success rate and portability of KBP across heterogeneous models without retraining the underlying KBP system. This approach provides a practical strategy to enhance the reliability of KBP and supports the use of off-the-shelf models through automated, model-agnostic post-processing.