S2287
Physics - Machine learning and AI algorithms
ESTRO 2026
%, and difference of bladder V65Gy was 4.5 ± 1.5 %. The text-prompting mechanism successfully modulated the dose distribution. The prompt "aggressively spare the rectum" resulted in a statistically significant average reduction in rectum V50Gy of 4.2% (p < [0.01]) compared to “standard VMAT plan”. The average inference time to generate a full 3D dose map on a single NVIDIA RTX4090 GPU was 142.8 seconds. Conclusion: We present the first 3D diffusion model for high- fidelity, interactive radiotherapy dose prediction. It ensures full 3D dose continuity and enables text- guided plan refinement, seamlessly integrating AI automation with clinical expertise for rapid, clinician- driven adjustments to treatment plans.
Neutroneneinfangtherapie (DGBNCT), University Hospital Essen, Essen, Germany. 3 Computer Science, Julius Maximilian University of Würzburg, Würzburg, Germany Purpose/Objective: Large Language Models (LLMs) are increasingly being integrated into medical workflows. However, two critical challenges limit their safe deployment: data privacy risks and hallucinations (generating plausible but incorrect values). In radiotherapy, where precise radiobiological calculations directly impact treatment planning and patient outcomes, LLM errors can compromise patient safety. Our Objective is to develop a tool-augmented LLM (TALLM) that addresses both challenges by maintaining data privacy while incorporating radiobiological knowledge to ensure reliable and accurate clinical decision support. Material/Methods: We developed a TALLM (Figure 1) by selecting “Mistral”, an offline, lightweight, open-source LLM, as the core model. We then built radiobiological calculation tools implementing the Linear-Quadratic (LQ) model to compute Biologically Effective Dose (BED) and the Equivalent Dose in 2 Gy fractions (EQD2). For toxicity assessment, we developed tools incorporating the Normal Tissue Complication Probability (NTCP) model using the Lyman-Kutcher- Burman (LKB) equation with organ-specific parameters based on the QUANTEC database1. To optimize computational efficiency, we approximated Equivalent Uniform Dose (EUD) using dose metrics based on tissue architecture: maximum dose (Dmax) for serial tissues and mean dose (Dmean) for parallel tissues2. Finally, we integrated these tools with the core LLM using a Natural Language Processing layer with regular expressions to parse user queries and route requests to the appropriate calculators. We validated our TALLM against GPT-4-Plus and Gemini-Pro using 50 radiobiological test prompts. Performance was evaluated using the Mean Absolute Error (MAE). The system was implemented using Python 3.13.5.Figure 1
Figure 2: An example of dose prediction.
Keywords: Dose Prediction, Diffusion Model, Prostate Cancer
Digital Poster 1872 Tool-augmented large language model for accurate radiobiological calculations in radiotherapy Haitam Lamtai 1,2 , Maroua L' Banani 1 , Anup Marasu Renukaprasad 3 , Arnd Obert 1 , Karim Moutchou 1 , Kathrin Breuer 1 , Thomas Fischer 1 , Irina Filimonova 1 , Marcus Zimmermann 1 , Bülent Polat 1 , Andrea Wittig- Sauerwein 1 1 Department of Radiotherapy and Radiation Oncology, University Hospital Würzburg, Würzburg, Germany. 2 Deutsche Gesellschaft für Bor-
Made with FlippingBook - Share PDF online