Examining Radiation Therapy Planning Knowledge in Large Language Models

Abstract

Personalized radiation therapy (RT) planning is crucial for minimizing toxicities. As large language models (LLMs) advance, it is vital to assess the RT planning knowledge across model types and sizes. We introduce a board-exam-style RT dataset and benchmark proprietary and open-weight LLMs in a zero-shot setting. Results show that state-of-the-art models exceed 80% accuracy, with larger models outperforming smaller ones, while fine-tuned compact models remain competitive, underscoring the potential of LLM-based decision support for clinical RT planning.

Ähnliche Arbeiten