Autonomous Stereotactic Radiosurgery Treatment Planning using a Large Language Model Agent: Proof of Concept
Recommended Citation
Nusrat H, Luo B, Hall R, Siddiqui M, Kim JP, Bagher-Ebadian H, Doemer A, Movsas B, Thind K. Autonomous Stereotactic Radiosurgery Treatment Planning using a Large Language Model Agent: Proof of Concept. Med Phys 2025; 52(8):16.
Document Type
Conference Proceeding
Publication Date
8-14-2025
Publication Title
Med Phys
Abstract
Purpose: Brain stereotactic radiosurgery requires precise planning to optimize tumor eradication while sparing healthy tissue, yet current manual methods are labor-intensive and time-consuming. This study evaluates an innovative AI-driven planning agent integrating large language models (LLMs) to autonomously optimize brain SRS plans, enhancing efficiency, consistency, and dosimetric quality while preserving patient data privacy. Methods: An autonomous agent integrating a state-of-the-art large language model (LLaMa3.1-8B) was developed, utilizing chain-of-thought prompting, retrieval-augmented generation (RAG), and reinforcement learning (RL) techniques to iteratively optimize radiotherapy treatment objectives. The agent was given 10 iterations to reach clinical goals per TG101. A scoring system quantified plan quality, awarding points for achieving the target coverage goal and deducting points for exceeding OAR constraints or target hotspots. The agent was retrospectively validated on 12 brain SRS cases. The autonomous agent also provided self-assessed confidence ratings at each iteration, scored on a scale from 0 (lowest) to 10 (highest), reflecting internal evaluation of its approach. Results: Agent optimized plans significantly improved over 10 iterations, with mean plan scores rising from an initial average of 10.5 to 72.5 at the final iteration (SEM ±25.0). The mean reported confidence across all optimization iterations was 7.2 ± 1.5, indicating a generally high level of internal certainty in decision-making. Conclusions: The AI-driven planning agent has the potential to enhance efficiency, consistency, and dosimetric quality in brain SRS planning. The agent's self-reported confidence ratings provide transparency into the LLM's decision-making processes, potentially enhancing clinical trust and adoption.
Volume
52
Issue
8
First Page
16
