ESTRO 2025 - Abstract Book

S3121

Physics - Inter-fraction motion management and offline adaptive radiotherapy

ESTRO 2025

1941

Digital Poster Predicting anatomical changes between radiotherapy fractions in nasopharyngeal carcinoma patients using generative modeling

zou yue, jiazhou wang, weigang hu, zhenhao li, dong yang, xiaojie yin, ziwei li Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai, China

Purpose/Objective: Inter-fractional organ motion during radiation therapy can introduce errors in the delivered radiation dose, potentially impacting treatment efficacy. To address this issue, this study proposes a population-based predictive model to estimate organ motion between treatment fractions for individual patients. Generative modeling techniques were employed to construct a predictive model of anatomical changes in patients, specifically designed to forecast potential anatomical variations in nasopharyngeal carcinoma patients during radiotherapy. Material/Methods: We incorporated the Adaptive Instance Normalization (AdaIN) algorithm, to propose a generative model for predicting anatomical changes in patients, based on the VQ-VAE (Vector Quantized Variational Autoencoder) framework. By leveraging the style features extracted from the planning CT scans using AdaIN, these features are applied within the VQ-VAE framework to generate images, thus learning the potential anatomical style variations that may occur between treatment fractions for individual patients.We trained the model using data from 90 patients, consisting of 496 planning CT scans and 2,518 paired FBCT images after registration. For the validation set, which includes 78 CT scans, we conducted a comparative analysis of the distribution characteristics of the parotid glands and body regions between the ground-truth and the generated images.

Results: For individual patients, this generative model effectively predicts potential anatomical variations (as shown in the figure). For the validation cohort, the results from the error bar plots indicate that the differences in the mean and standard deviation between the real and generated images across different ROIs are minimal. The discrepancies between the mean, standard deviation, and true values for each ROI in the validation set are smaller compared to