PaperlayerView exact revision on arXiv

Source-linked AI summary

An Evidence-Driven Multi-Agent Compound AI System for Head and Neck Proton Radiotherapy Planning

Qingqing Wang, Liqiang Xiao, Chang Chang

arXiv:2610.29010v1cs.AI

TL;DR

Treatment planning must balance competing clinical objectives, while existing approaches have limits in incorporating physician instructions and in domain-specific reasoning for plan refinement. This paper combines automated proton PBS planning with a multi-agent system whose post-trained planning agent recommends evidence-supported adjustments; compared with the base LLM, it improves instruction-following, retrieval, and clinical reasoning and supports more effective patient-specific trade-offs.

  • Problem

    Treatment planning must balance competing clinical objectives, while DRL-based methods lack flexibility for human instructions and LLM-based approaches have limited domain-specific reasoning.

    Source: §1 · pp. 3, 5

  • Method

    The system combines DRL-based automated proton PBS planning with a conversational multi-agent refinement platform and a post-trained planning agent that uses evidence from prior plans.

    Source: §1 · p. 6, §2.3 · p. 8

  • Results

    Post-training improved instruction-following, information retrieval, and clinical reasoning, with adjustments better improving user-prioritized objectives while limiting deterioration in competing objectives.

    Source: §5 · pp. 28–29

  • Takeaways & Limitations

    The framework offers a user-friendly workflow for physician-guided proton PBS plan refinement and a promising foundation for personalized planning.

    Source: §4 · p. 27, §5 · p. 29

    Automatic beam angle selection was not addressed, and planning-structure creation relied on empirical rules.

    Source: §4 · pp. 27–28

Abstract

from arXiv · show

Radiation treatment planning requires balancing competing objectives to satisfy clinical criteria while accommodating physicians' preferences. This study aims to develop an evidence-driven, multi-agent compound AI system that integrates DRL-based automated treatment planning with LLM-based conversational plan refinement for proton PBS therapy. A policy network trained using PPO first generates multiple plans without human intervention, from which one is selected as the target plan and the others serve as references for subsequent refinements. For challenging cases requiring further adjustments, an LLM-based multi-agent system enables plan refinement through natural language interactions. The system comprises a planning agent that recommends evidence-supported adjustments, a chatbot providing a unified conversational interface, and an evaluation agent that retrieves clinical metrics and identifies constraint violations. To enhance domain-specific reasoning, the planning agent based on Qwen3-14B is designed to justify its recommendations using evidence from previously generated plans and post-trained on task-specific data using SFT and GRPO. The framework was trained and evaluated using 21 patients with H&N cancers, with 5 PPO-generated plans per patient. Of these, 50 plans were selected for instruction-guided refinement. Compared with the base LLM, the post-trained Qwen3-14B demonstrated improved instruction-following, information retrieval and domain-specific reasoning capabilities. The resulting adjustments more effectively improved user-prioritized objectives while limiting deterioration in competing objectives, yielding more clinically appropriate trade-offs in patient-specific plan refinement. The proposed compound AI system provides a user-friendly planning platform, offering a promising approach to personalized proton PBS treatment planning.

1. Introduction

Treatment planning must balance target coverage and normal-tissue sparing, yet iterative refinement can take days and depend on planner experience. The paper proposes combining automated proton PBS planning with evidence-driven, physician-guided conversational refinement.

1. Introduction

  • Close or overlapping organs at risk and tumors make trade-offs between target coverage and normal-tissue sparing unavoidable.Clinical refinement can take several days, and plan quality can vary with planner experience, available time, and treatment-planning systems.Source: §1 · pp. 3–4
  • Knowledge-based planning can converge to inferior local minima, while multi-criteria optimization requires substantial manual effort to generate, explore, and select plans.These methods do not fully automate the time- and expertise-demanding planning process.Source: §1 · p. 4
  • DRL planning has challenges in generalizing across treatment sites and defining rewards, while base LLMs may recommend overly aggressive or ineffective adjustments that fail to balance competing objectives.The introduction also notes concerns about guideline-inconsistent clinical recommendations and limited adaptation to changing patient conditions.Source: §1 · pp. 4–5
  • Existing LLM-based planning frameworks emphasize initial plan generation over physician-guided iterative refinement, and proton therapy remains relatively underexplored.The refinement stage is difficult when objectives conflict and physicians prioritize them differently.Source: §1 · p. 6
  • The proposed framework combines DRL-generated initial proton PBS plans with interactive refinement by a domain-specific, post-trained LLM agent.The agent uses SFT and GRPO to recommend patient-specific adjustments supported by evidence from previously generated plans.Source: §1 · p. 6

2.1. Patient Cohorts

The retrospective cohort comprised 21 bilateral or ipsilateral head-and-neck cases treated at one proton center between 2017 and 2024.

2.1. Patient Cohorts

  • 21 bilateral or ipsilateral H&N cases treated at the California Protons Cancer Therapy Center from 2017 to 2024 were retrospectively selected.The study was conducted under institutional review board approval at the University of California San Diego.Source: §2.1 · p. 7
  • Treatment courses ranged from 5–60 fractions delivered twice daily, with no more than five beams per plan..Source: §2.1 · p. 7
  • Cases retained the delivered plans’ prescriptions, beam angles, target definitions, and field-specific targets.Each case had one to four prescription levels spanning 4.0–74.2 Gy.Source: §2.1 · p. 7

2.2. Automatic Treatment Planning Framework for Proton PBS

The framework first generates proton PBS plans automatically, then uses prior plans to support further refinement of challenging cases.

2.2. Automatic Treatment Planning Framework for Proton PBS

  • A PPO-trained adjustment policy automatically generates initial plans for uploaded patient cases without human intervention.The planning system uses an L2O inverse optimization engine as its backend.Source: §2.2 · p. 7
  • For challenging cases, an LLM-based multi-agent module adjusts a selected plan by reasoning over plans generated during initial planning.Easy cases may receive clinically accepted plans from the initial planning module.Source: §2.2 · p. 7
  • The framework uses the system depicted in Figure 2 to connect automatic initial planning with multi-agent refinement.The authors focus their evaluation on the multi-agent module and planning-agent post-training, noting prior demonstrations of the PPO policy and L2O optimizer.Source: §2.2 · p. 7, Figure 2 · p. 8

2.3. LLM-based Multi-agent Planning System

Three collaborative agents provide a conversational interface for plan refinement, clinical evaluation, and evidence-based adjustment recommendations.

2.3. LLM-based Multi-agent Planning System

  • The chatbot is the sole user interface, routing plan-refinement requests to the planning agent and clinical-metric or constraint queries to the evaluation agent.The system’s chatbot and evaluation agents use prompted GPT models, while the planning agent is post-trained Qwen3-14B.Source: §2.3 · p. 8
  • The evaluation agent retrieves clinical requirements, evaluates plan quality, checks constraint violations, and compares plans before and after adjustment.It uses a memory module or predefined APIs and provides feedback to the planning agent.Source: §2.3 · p. 8
  • For high-level instructions, the planning agent recommends adjustments and API calls through a predefined workflow, justifying recommendations with evidence from prior plan objective-metric mappings.For specific parameter commands, it invokes APIs directly and confirms with the user before new plans are generated.Source: §2.3 · p. 9
  • The memory module stores patient information, clinical criteria, and plan objective-metric mappings for retrieval-augmented generation, while APIs calculate metrics and modify or optimize plans.The paper argues that final simulated plan results should assess both targeted-objective improvement and deterioration in competing objectives, including trade-off reasonableness.Source: §2.3 · p. 9

2.4. Post-training of the Planning Agent

The planning agent uses current and prior-plan metrics to recommend adjustments aligned with user intent, then presents API calls for user confirmation. It is post-trained on manually checked refinement examples with SFT and GRPO.

2.4. Post-training of the Planning Agent

  • The planning agent retrieves current-plan metrics, selects a prior plan reflecting the user’s intended trade-off, compares both plans, and proposes adjustments for confirmation.It checks target volumes and targeted or critical OARs whose metrics fail or nearly fail clinical requirements.Source: §2.4 · p. 10

2.4.1. Data Collection

  • 50 refinement samples—42 for training and 8 for evaluation—were created from five PPO-generated plans per case across 21 cases.Queries and ground-truth responses were manually reviewed, with reference plans and adjustments checked for consistency with human planning practice.Source: §2.4.1 · p. 11

2.4.2. Supervised Fine-tuning

  • SFT trains the agent to retrieve plan evidence and recommend adjustments, while LoRA enables task-specific adaptation with the original model parameters frozen.The think block includes objective settings and clinical metrics from current and reference plans; the reported LoRA rank and scaling factor are 32 and 64.Source: §2.4.2 · pp. 11–12

2.4.3. Group Relative Policy Optimization (GRPO)

  • GRPO further trains the agent to distinguish case-specific adjustments by comparing group-normalized rewards and constraining policy updates with a clipped objective.A GPT-5.6-based judge scores generated responses against ground-truth responses by estimating proposed API calls’ clinical effects across four aspects.Source: §2.4.3 · pp. 12–13

2.5. API Tools and Memory

The API tools let the evaluation agent assess plan metrics and the planning agent modify planning objectives. Customized memory stores treatment information and prior-plan objective–metric mappings to support evidence-based recommendations.

2.5. API Tools and Memory

  • Evaluation APIs compute structure-level dose measures, including maximum, minimum, and mean dose, Dx/Dcc, and Vx metrics.These tools provide clinical plan metrics for evaluation.Source: §2.5 · p. 14
  • Planning APIs modify existing objective constraints or priorities and can adjust prescriptions or generate a new dose distribution through inverse optimization.Dose- and volume-based constraint tools target irradiated subregions, including hot or cold areas in target volumes or OARs.Source: §2.5 · pp. 14–15
  • Customized JSON memory stores patient and treatment details plus prior-plan objective–metric mappings that support evidence-backed adjustment recommendations.Stored information includes beam configurations, clinical criteria, structures, prescriptions, and fractionation schemes.Source: §2.5 · p. 15

3. Experiments

The experiments evaluate how domain-specific post-training affects the planning agent’s outputs and the clinical effects of recommended plan adjustments, including patient-specific trade-offs. They also describe the conversational framework’s support for plan selection, natural-language refinement, and questions about clinical criteria and plan metrics.

3.1. Comparison of LLM Outputs

  • Post-trained Qwen3-14B followed the required format, requested confirmation, retrieved current and reference-plan values accurately, and gave more clinically reasonable recommendations than the base model.The base model often omitted retrieved values, proposed generic adjustments, and could introduce conflicting objectives; post-training also improved reasoning through more complete retrieval and GRPO-optimized recommendations.Source: §3.1 · pp. 16, 18, Figure 3 · p. 17

3.2. Comparison of Recommended Adjustments

  • The evaluation applied LLM-recommended adjustments, then used inverse optimization and Monte Carlo dose calculation to assess gains toward physician intent alongside degradation in competing objectives.The study selected multiple plans with different trade-offs from challenging bilateral-cancer patients for adjustment.Source: §3.2 · pp. 18–19
  • For Patient B25 Plan 1, Parotid_R Dmean fell from 28.8 Gy to 26.01 Gy with base and 21.08 Gy post-trained, meeting the 26 Gy hard constraint only post-trained.Both adjusted plans slightly improved Parotid_L sparing and target coverage, at the cost of slightly increased but clinically acceptable CTV7000 hotspots.Source: §3.2 · p. 19
  • For B25 Plan 2, Glnd_Thyroid V30Gy was 55.63% at baseline, 60.53% after base adjustment, and 16.86% after post-trained adjustment.Post-trained Oropharynx Dmean was 45.02 Gy; the request also sought to preserve Parotid doses and CTV7000 coverage.Source: §3.2 · p. 19
  • For B26 Plan 1, post-trained adjustment reduced OpticNerve_L D0.03cc to 49.06 Gy and Parotid_L Dmean to 23.66 Gy, while .CTVhd V100% was 93.55%.The request prioritized sparing several OARs while maintaining target coverage; the reported post-trained values still include other constraint failures.Source: §3.2 · p. 20
  • For B25, the post-trained model improved Oropharynx and Glnd_Thyroid sparing while maintaining all three CTVs, unlike the base model’s CTV5400 coverage compromise.Post-trained mean doses to Oropharynx and Glnd_Thyroid were 42.81 Gy and 30 Gy; the base model reduced CTV5400 V99% and V100% by 2.33% and 2.54%, respectively.Source: §3.2 · p. 22
  • For B26’s first plan, post-training achieved greater OAR sparing with only a 0.6% reduction in .CTVhd V100%, whereas the base model’s coverage loss was clinically unacceptable.For the third plan, the base model’s maximum dose reached 121.09%, while the post-trained model reduced Len_L D0.03cc below the clinical constraint without adding hot spots.Source: §3.2 · pp. 22–23
  • Across I22 and I23, the post-trained model made conservative trade-offs when OAR constraints were already violated, but used available margins for more aggressive coverage gains.For I22, it reduced CTV 66 maximum dose to 113.50% of prescription while largely preserving OAR sparing; for I23, OAR doses remained within clinical constraints despite increases.Source: §3.2 · pp. 23–24
  • For I26, post-training reduced Parotid_L mean dose from 30.69 Gy to 21.82 Gy, with CTV_6600 V99% and V100% decreasing by 0.21% and 0.16%.The authors describe both models’ distinct trade-offs as potentially reflecting physician preferences and clinically reasonable.Source: §3.2 · pp. 24–25
  • Post-training improved planning-agent alignment with clinical practice and patient-specific adaptation, while base Qwen3-14B could make overly aggressive or ineffective adjustments.The authors note that suboptimal recommendations may require more iterations to reach a physician-satisfactory plan, prolonging planning time and increasing token costs.Source: §3.2 · p. 25

3.3. Conversational Planning Using Proposed Multi-agent framework

  • The conversational framework lets users select among DRL-generated plans, refine challenging cases with specific commands or high-level instructions, and ask questions in natural language.It also retrieves plan metrics and clinical constraint violations, with the potential to reduce treatment planners’ workload and overall planning time.Source: §3.3 · pp. 25–26

4. Discussion

The framework links automatic planning with evidence-supported, physician-guided refinement, but remains an initial effort that needs further development before clinical translation.

4. Discussion

  • The framework supports both automatic initial planning and interactive refinement, addressing two stages of the treatment-planning workflow.The planning agent can justify adjustments using objectives and clinical metrics from previously generated plans.Source: §4 · p. 27
  • This pilot framework still needs capability and user-experience improvements to facilitate clinical translation.The authors describe it as an initial effort to bridge automatic plan generation and instruction-guided refinement.Source: §4 · p. 27
  • Current structure APIs extract subregions by dose range or highly irradiated volume; auxiliary planning structures require more flexible creation methods.Proposed directions include geometric and Boolean-operation APIs or integration with prompt-guided medical image segmentation.Source: §4 · p. 27
  • Automatic beam angle selection is not addressed, and robust optimization is absent from the current framework.Repeated Monte Carlo dose calculations are identified as a computational challenge for robust optimization.Source: §4 · p. 28
  • The authors identify expanding the training dataset and supervising intermediate decisions, especially reference-plan selection, as possible ways to improve recommendation accuracy and consistency.These are proposed future directions rather than evaluated improvements.Source: §4 · p. 28

5. Conclusion

The study integrates DRL-based automated planning with multi-agent conversational refinement for proton PBS treatment planning. Post-training improved the planning agent’s capabilities and enabled more user-aligned adjustments while maintaining reasonable trade-offs among clinical objectives.

5. Conclusion

  • The compound AI system integrates DRL-based automated planning with multi-agent conversational refinement for proton PBS treatment planning.The proposed structured planning pipeline and SFT-GRPO strategy support evidence-driven, patient-specific plan adjustments.Source: §5 · p. 28
  • Post-training improved instruction-following, information retrieval, and clinical reasoning, enabling adjustments aligned with user intentions while maintaining reasonable trade-offs.The study presents this as an initial integration of automated initial planning and physician-guided interactive refinement.Source: §5 · pp. 28–29
  • The framework offers a foundation for more intelligent, interactive, and clinically adaptable agentic treatment-planning systems.The authors also emphasize a more flexible and user-friendly planning workflow.Source: §5 · p. 29

Related papers

Loading 2610.29010v1…