Related Experiment Video
Updated: Oct 29, 2026
Positron Emission Tomography-based Dose Painting Radiation Therapy in a Glioblastoma Rat Model using the Small Animal Radiation Research Platform
Published on: March 24, 2022
AI-assisted Treatment Planning - Only a Vision? A Laboratory Proof-of-Concept in Complex Radiotherapy Treatment
Claudia Sperling1, Maik Stille2, Annalena Kohnert2
1Department of Radiation Oncology, University of Lübeck, Lübeck, Germany.
Background/Aim:
The present study investigates whether volumes of interest (VOIs) derived from local dose deficits in VMAT-based craniospinal irradiation (CSI) can be used for dose optimization within an LLM-guided treatment planning process. Artificial intelligence (AI) may transform radiotherapy treatment planning, but current commercial treatment planning systems (TPS) remain primarily rule- and goal-based and lack dialogue-based decision making during ongoing optimization.
Materials And Methods:
This laboratory proof-of-concept investigated an AI-assisted planning workflow using a large language model (LLM) as a decision-making instance (DMI). Volumetric modulated arc therapy (VMAT) CSI was selected as a challenging scenario. Three strategies were evaluated: a baseline without volumes of interest (VOIs), Experiment A with two predefined VOIs, and Experiment B with VOIs introduced dynamically in response to the evolving dose distribution.
Results:
The DMI iteratively evaluated dose distributions and planning parameters and recommended objective adjustments to the planning target volume (PTV) and auxiliary structures. The estimated Van't Riet conformity index increased from 0.84 in the baseline experiment to 0.98 in Experiment A and 0.99 in Experiment B. In the baseline experiment, the final PTV D98% was 94.8%, while local dose deficits of approximately 80% remained within the junction regions. In Experiment A, the final D98% values were 98.6% for the PTV, 95.2% for VOI1, and 94.4% for VOI2. In Experiment B, three VOIs were sequentially generated during optimization, and final D98% values reached 96.4% for the PTV, 95.8% for VOI1, 95.6% for VOI2, and 95.0% for VOI3.
Conclusion:
Under laboratory conditions, the LLM-based DMI structured and guided an iterative dose-optimization process, supporting the feasibility of LLM-based planning assistance in radiotherapy.
