An Iterative LLM Framework for SIBT Utilizing RAG-Based Adaptive Weight Optimization
Date:
Talk at 2025 Annual Meeting of Chinese College of Interventionalists (CCI 2025), International Youth Conference Hotel, Nanjing, China
This work was presented at CCI 2025 at the International Youth Conference Hotel in Nanjing. The study introduced an iterative framework that combines a locally deployed large language model, retrieval-augmented generation, and automatic plan optimization to adapt objective-function weights for seed implant brachytherapy.
Evaluated on 23 head-and-neck cases, the framework generated plans in 3.7 ± 0.7 minutes. Compared with clinical plans, it achieved comparable DVH metrics with fewer needles and lower variability; compared with fixed-weight plans, it reduced high-dose regions and improved organ-at-risk sparing, demonstrating the potential to reduce manual intervention in clinical planning.
