Local large language model-guided optimization for prostate radiotherapy planning

Abstract

Background

Radiotherapy planning is time-intensive, iterative, and operator-dependent. Locally deployed large language-model agents may support planning automation while avoiding transmission of patient data to external model providers.

Purpose

The purpose of this work was to develop and evaluate SAGE, a locally deployed LLM agent for autonomous VMAT optimization-priority tuning, and to investigate the impact of model size (8B vs. 70B) and intra-case memory implemented through retrieval augmentation.

Methods

The SAGE agent (Llama 3.1 8B/70B) was interfaced with a clinical TPS to tune optimization priority values for 17 retrospective prostate cancer patients (60 Gy/20fx). All SAGE priority values were initialized to 1, while the study-defined optimization objective set and beam geometry were held fixed. We compared four agent variants (8B, 70B, 8B-RAG, 70B-RAG) to the clinical plan using paired Wilcoxon tests with FDR control.

Results

No statistically significant difference in target (PTV V58.5 Gy) coverage was observed between SAGE variants and clinical plans. The 70B-RAG agent significantly improved penile bulb sparing (V22Gy) over clinical plans (p = 0.014). Enabling intra-case memory provided a statistically significant improvement in penile bulb sparing, while no significant effect was observed from model size alone. While clinical plans were statistically better for some OARs, all SAGE 70B-RAG plans remained clinically acceptable. The most complex version of SAGE (70B-RAG) completed ten autonomous iterations in a median compute time of 45 min (IQR: 34–51). For reference, the clinical planning workflow turnaround (ARIA-recorded task duration) median time was 2700 min (IQR: 2040–3060).

Conclusions

This proof-of-concept study demonstrates that a local LLM agent can autonomously tune optimization priority values to generate clinically acceptable prostate VMAT plans within a fixed, study-defined objective space. Intra-case memory produced a statistically significant, endpoint-specific improvement (penile bulb V22Gy). These findings suggest that smaller, efficient models may be viable for constrained treatment-planning optimization tasks.