S1857
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
Figure 1: The framework of the universal text-driven automated treatment planning Results: Using LLM-extracted planning expertise, UText- Planner successfully generated automated plans for all 75 patients, averaging six rounds of optimization per plan. Figure 2 illustrates an instance of adjusting objective parameters and the resulting dose volume histogram (DVH) metric over iterations. In the internal validation set, automated plans significantly improved the target homogeneity index and reduced the mean dose to the femoral heads and red marrow by 0.60 Gy and 2.10 Gy, respectively. In the blinded evaluation on the external validation set, automated plans were rated as superior or equivalent to manual plans in 86.7% to 88.9% of cases across three independent experts.
Figure 2: The trajectory of objectives adjustment and resulting DVH metrics throughout the optimization iterations Conclusion: This study demonstrated the potential of an LLM- based, text-driven ATP framework to generate high- quality plans with limited data and enhance generalizability. References: [1] Petit S, Franco P, Heukelom J, Callens D. Increasing cancer incidence and workforce shortages - It is time to act now. Radiother Oncol. 2025;211:111057.[2] van Genderingen J, Nguyen D, Knuth F, et al. Deep learning
dose prediction to approach Erasmus-iCycle dosimetric plan quality within seconds for
instantaneous treatment planning. Radiother Oncol. 2025;203:110662.[3] Hirotaki K, Tomizawa K, Moriya S, et al. Fully automated volumetric modulated arc therapy planning for locally advanced rectal cancer: feasibility and efficiency. Radiat Oncol. 2023;18(1):147. Keywords: Large Language Models, Generalizability
Made with FlippingBook - Share PDF online