Automated Stereotactic Radiosurgery Planning Using a Reasoning-Based Large Language Model Agent
Abstract
Purpose
Automated stereotactic radiosurgery (SRS) planning often fails to match the high conformity and complex trade-off logic of expert human dosimetrists. We hypothesized that a reasoning-based AI agent, SAGE (Secure Agent for Generative dose Expertise), could generate single-fraction SRS plans dosimetrically equivalent or superior to physician-approved clinical plans.
Methods
SAGE was retrospectively evaluated using 41 patients with brain metastases treated with 18 Gy single-fraction SRS. The agent’s core consists of non-reasoning (large language model, LLM) or reasoning (large reasoning model, LRM) models which can be invoked by the human-in-the-loop. The reasoning architecture was based on the Qwen QwQ-32B LRM to explicitly verify constraints and deliberate trade-offs within the Varian Eclipse Treatment Planning System (TPS). AI-generated plans were compared to treated clinical plans using paired non-parametric Wilcoxon signed-rank tests (p0.05). The reasoning engine drove this performance by verifying an average of 11.1 constraints and 14.9 trade-offs per plan. Furthermore, incorporating human-in-the-loop natural language prompts enabled iterative refinement, significantly improving CI (p<0.05) relative to the initial AI-generated plans.
Conclusion
We demonstrate that a reasoning-based AI agent can independently generate SRS plans that are clinically indistinguishable from human expert plans. By mimicking the explicit verification logic of physicists, this framework presents a transparent, high-quality solution for automated radiosurgery.