Author: Rex A. Cardan, Carlos E. Cardenas, Udbhav S. Ram 👨🔬
Affiliation: The University of Alabama at Birmingham, University of Alabama at Birmingham 🌍
Purpose: The AAPM TG-263 report provides nomenclature guidelines for target and normal tissue structures used in radiation oncology. Adherence to these guidelines is challenging for targets, as there is high variance for prescriptions/structures. Here, we investigate the utility of a locally hosted, large language models (LLM), combined with structure templates and automation scripts, can streamline and standardize target naming in radiotherapy planning.
Methods: We used locally-hosted versions of Llama3.2:3B, DeepSeek-r1:7B, and Mistral:7B, each guided by custom system prompts with few shot learning examples and structured JSON outputs. The model is prompted to choose from a rigid subset of options, given an input description. These subsets are determined by the software options available in a TG263 name generation tool. Descriptions are then converted to a set of JSON outputs, and the resulting options are inputted to the TG263 generator, generating the correct names based on the input description. To evaluate this approach, we generated 1,000 prompt-ID pairs (ex. “The vascular PTV prescribed to the highest dose level”, “PTVvas_High”). We then measured the Levenshtein distance between each model’s generated ID and the reference ID (distance=0 indicates an exact match).
Results: Across 1000 target structures, the mean Levenshtein distances were 2.212±3.53, 1.494±3.67, and 2.47±6.87 for Deepseek-r1, Llama3.2 and Mistral-7B, respectively, showing close agreement for most cases. Llama3.2 and Mistral-7B produced exact matches in 69% of cases, demonstrating strong agreement with the reference TG-263–compliant names.
Conclusion: Preliminary results suggest that locally hosted, open-source LLMs can promote consistency and efficiency in target naming. With further testing, these models may offer secure and practical automation solutions for TG-263 compliance in clinical workflows.