2569 - Agentic LLM-Based Auto-contouring With Clinical Context for Lung SBRT Targets

Purpose/Objective(s): Auto-segmentation of gross tumor volumes (GTV) for lung stereotactic body radiation therapy (SBRT) using traditional image-only deep learning alone lacks integration with clinical context, potentially limiting performance in challenging cases.

Materials/Methods: We evaluated 27 lung SBRT cases with physician-delineated GTV contours across diverse histologies (adenocarcinoma, squamous cell, small cell, NSCLC NOS, metastatic) at one institution. Agent-ROCCK used a code-agent LLM that ingested clinical data (consult notes, pathology and oncology reports) and followed a domain-specific knowledge base encoding the sequence of actions for radiation oncology contouring in thoracic oncology to identify tumor locations, orchestrate imaging tools including nnU-Net, and iteratively refine contours. This was compared to a standalone nnU-Net trained on an institutional lung tumor dataset. Dice similarity coefficient (DSC) was compared using paired superiority testing (one-sided paired t-test, Wilcoxon signed-rank). Subgroup analyses examined performance by tumor diameter (median split: 2.0cm) and histology.

Results: Agent-ROCCK achieved significantly higher mean DSC than standalone nnU-Net overall (0.79 vs 0.73; one-sided p=0.029) with lower variance (SD 0.15 vs 0.24). A significant size-method interaction was observed (p=0.013): for small tumors (=2cm), Agent-ROCCK was superior (delta +0.17, p=0.019), while standalone nnU-Net trended better for large tumors. DSC did not vary by histology for either method (p>0.2).

Conclusion: An agentic LLM framework that orchestrates deep learning segmentation with clinical context achieved superior lung SBRT GTV contouring compared to standalone nnU-Net, driven by markedly better performance on small tumors. Integrating clinical reasoning into segmentation pipelines may improve auto-contouring robustness for challenging lesions where imaging features alone are insufficient.