Automated stereotactic radiosurgery planning using a human-in-the-loop reasoning large language model agent

TL;DR

This study investigates whether a System-2 reasoning-enabled LLM planner (SAGE) can autonomously generate stereotactic radiosurgery plans for brain metastases that are non-inferior to those created by human dosimetrists. By directly comparing a reasoning-enabled model to a non-reasoning counterpart in 41 retrospective cases, the authors show comparable target coverage and dosimetry while achieving significantly better right cochlear sparing. The reasoning agent also produces explicit planning traces—constraint verification, causal explanations, and forward simulations—facilitating auditable decision logs. These findings suggest that deliberative AI architectures can enhance plan quality and transparency, potentially enabling safer, more scalable automated SRS planning within clinical workflows.

Abstract

Stereotactic radiosurgery (SRS) demands precise dose shaping around critical structures, yet black-box AI systems have limited clinical adoption due to opacity concerns. We tested whether chain-of-thought reasoning improves agentic planning in a retrospective cohort of 41 patients with brain metastases treated with 18 Gy single-fraction SRS. We developed SAGE (Secure Agent for Generative Dose Expertise), an LLM-based planning agent for automated SRS treatment planning. Two variants generated plans for each case: one using a non-reasoning model, one using a reasoning model. The reasoning variant showed comparable plan dosimetry relative to human planners on primary endpoints (PTV coverage, maximum dose, conformity index, gradient index; all p > 0.21) while reducing cochlear dose below human baselines (p = 0.022). When prompted to improve conformity, the reasoning model demonstrated systematic planning behaviors including prospective constraint verification (457 instances) and trade-off deliberation (609 instances), while the standard model exhibited none of these deliberative processes (0 and 7 instances, respectively). Content analysis revealed that constraint verification and causal explanation concentrated in the reasoning agent. The optimization traces serve as auditable logs, offering a path toward transparent automated planning.

Figures (7)

• Figure 1: The agent receives clinical inputs (patient anatomy, prescription, physician constraints) and current optimizer state. Two model variants are shown: non-reasoning (top) and reasoning (bottom). Each executes iterative optimization cycles comprising LLM-based parameter adjustment, dose calculation, plan evaluation, and objective updates. The optimal plan proceeds to human review, where it is either accepted or returned with refinement feedback.
• Figure 2: PTV coverage (left) and maximum dose (right) for clinical plans (grey), non-reasoning model (red), and reasoning model (blue). Violin contours represent kernel density estimates. White boxes indicate IQR; red points indicate median values; black points represent individual patients (n = 41). Dashed lines denote clinical acceptance thresholds (coverage > 95%; maximum dose < 21.6 Gy). Brackets indicate comparisons between reasoning and clinical groups that remained significant after BH correction (q < 0.05).
• Figure 3: Maximum doses to brainstem and optic chiasm, and V12Gy for normal brain (defined as brain minus gross tumor volume (GTV)) across clinical plans (grey), non-reasoning model (red), and reasoning model (blue). Violin contours represent kernel density estimates. White boxes indicate IQR; red points indicate median values; black points represent individual patients (n = 41). Brackets indicate comparisons between reasoning and clinical groups that remained significant after BH correction (q < 0.05).
• Figure 4: Maximum doses to bilateral optic nerves and cochleae across clinical plans (grey), non-reasoning model (red), and reasoning model (blue). Violin contours represent kernel density estimates. White boxes indicate IQR; red points indicate median values; black points represent individual patients (n = 41). The reasoning model achieved significantly lower right cochlear doses compared to clinical plans. All other comparisons were not significant.
• Figure 5: Conformity index values before (Round 1) and after (Round 2) the refinement prompt for reasoning (blue) and non-reasoning (yellow) model variants. Both models showed significant improvement following refinement (non-reasoning: p = 0.007; reasoning: p < 0.001, paired Wilcoxon signed-rank test). The reasoning model achieved median conformity approaching clinical benchmark values. Boxes indicate IQR; horizontal lines indicate median; whiskers extend to a 1.5 multiple of IQR.