3685 - LLM-Based Structured Clinical Information Integration, Multimodal Target Delineation, and Dose-Prescription Guidance in Nasopharyngeal Carcinoma
Presenter(s)
X. Zeng1, Y. Liu2, Z. Zhen3, Z. Liang3, L. Jia4, Y. Sun5, and L. Lin6; 1State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China, 2Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, 3State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China, Guangzhou, China, 4Shenzhen United Imaging Healthcare Co., Ltd., Shenzhen, Guangdong, China, 5Sun Yat-sen University Cancer Center; State Key Laboratory of Oncology in South China; Collaborative Innovation Center for Cancer Medicine; Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangzhou, China, 6State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, Guangdong, China
Purpose/Objective(s):
Accurate target delineation for nasopharyngeal carcinoma (NPC) relies on clinical information that is often unstructured and inconsistently documented. This study developed a three-stage pipeline consisting of LLM-based extraction of guideline-relevant clinical information, multimodal segmentation integrating CT imaging with structured profiles and clinical guidelines, and a guideline-informed decision tree for dose-prescription recommendations, and evaluated whether this framework improves accuracy and interpretability compared with imaging-only approaches.Materials/Methods:
The workflow included three components: extraction of structured, guideline-relevant information from raw clinical free text; incorporation of these elements into a multimodal segmentation model; and generation of interpretable, guideline-based dose-prescription recommendations through a rule-driven decision tree. A retrospective cohort of 132 NPC patients was analyzed. Simulation CT scans, expert-annotated target volumes, unprocessed clinical free text, and guideline-based delineation principles were collected. Clinical text was processed using a prompted LLM (Qwen3-256B) to extract key clinical elements into structured profiles. CT images underwent intensity normalization and resampling to 1×1×3 mm³. The multimodal framework employed a dual-branch architecture consisting of a 3D U-Net encoder–decoder for volumetric feature extraction and a text-embedding module for LLM-derived profiles. Target delineation followed the recent multinational consensus guidelines and contouring atlas for NPC published in The Lancet Oncology. Segmentation accuracy was assessed using the Dice similarity coefficient (DSC), and LLM extraction accuracy was validated against expert-defined structured labels.Results:
In the first stage, the LLM achieved high extraction accuracy (127/132, 96.2%), with errors mainly attributable to incomplete clinical information or heterogeneity in radiology report phrasing. In the second stage, incorporating LLM-derived structured clinical information alongside guideline-based principles improved clinical target volume delineation, with the multimodal model outperforming the imaging-only baseline (Dice 0.85 vs 0.82). In the final stage, a guideline-informed decision tree generated interpretable dose-prescription recommendations, and its reliability depended on the fidelity of the upstream clinical information extraction.Conclusion:
This study demonstrates that structured clinical information can be extracted from unprocessed medical text with high accuracy and effectively integrated into a multimodal DL framework for NPC target delineation. Incorporating LLM-derived structured profiles yielded measurable improvements in contouring performance. Future work will extend this framework to all target volumes and evaluate its generalizability in larger multicenter cohorts.