S2289
Physics - Machine learning and AI algorithms
ESTRO 2026
Conclusion: VQ-DoseNet introduces a clinically meaningful stochastic dose prediction approach that models real planning variability while preserving anatomical fidelity. By generating diverse yet realistic dose distributions, the framework enhances the reliability and interpretability of AI-based dose prediction, supporting downstream applications such as automated plan generation, multi-plan evaluation, and clinical decision support across disease sites. Keywords: Stochastic dose prediction, Vector- quantization Foundation Model–Driven Regions of Interests Classification and Renaming in Radiotherapy: A Customizable, Retraining-Free Workflow Across Institutions Dong Yang 1 , Mingjun Lei 2 , Qiangxing Yang 3 , Zihan Sun 3 , Xuewen Hou 4 , Weigang Hu 1 , Jiazhou Wang 1 1 Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai, China. 2 Radiation Oncology, Xiangya Hospital, Changsha, China. 3 Radiation Oncology, The first affiliated hospital of Zhejiang University school of medicine, Hangzhou, China. 4 Radio Therapy Business Unit, Shanghai United Imaging Healthcare Co., Shanghai, China Proffered Paper 2069
Results: On the in-house nasopharynx dataset, VQ-DoseNet achieved a mean absolute error (MAE) of 0.106 Gy. On the OpenKBP dataset, the average dose and DVH scores were 3.608 ± 1.267 Gy and 1.329 ± 1.934 Gy, respectively. Across test cases with 50 stochastic predictions, all predicted dose distributions maintained clinically acceptable PTV coverage and OAR constraints. The range of stochastic predictions closely matched the variability observed among clinical re-plans.In the cross-site validation, results remained consistent across rectal and breast datasets, with no significant violation of clinical constraints. These findings indicate robust model generalization across different anatomical sites and planning geometries.
Made with FlippingBook - Share PDF online