ESTRO 2025 - Abstract Book

S3764

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Purpose/Objective: Radiation therapy is one of the main treatment modalities for nasopharyngeal cancer. However, the recurrence rate of nasopharyngeal cancer after radiotherapy is around 10% and is increasing year by year. Based on the continuous development of deep learning in the field of computer vision, this study aims to develop a novel prediction model by combining the patient's imaging information and clinical information to predict the early recurrence after radiation therapy, so as to provide timely and relevant interventions and a second course of radiotherapy.

Material/Methods:

A total of 468 nasopharyngeal cancer patients from Fudan University Shanghai Cancer Center who had been treated with IMRT between 2009 and 2020 were included in this study. The Transfomer-based 3D predictive model,TransPred, was developed in the training and validation cohorts after a five-fold cross-validation and independently tested on an internal test cohert. The model uses CT images, Radiotherapy dose distribution images, PTV (Planning Target Volume) outlining, and relevant clinical variables as the main inputs to predict the probability of recurrence in patients after primary treatment with radiotherapy. In this study, comparative experiments were conducted on the TransPred model with inconsistent inputs as well as the traditional deep learning model 3D Resnet and machine learning methods. In addition, based on ensuring the accuracy of the prediction, 3D visualization and analysis of the model was performed by the Grad-CAM method to further localize the high-risk areas of possible recurrence in patients at high risk of recurrence of nasopharyngeal cancer. Results: After independent testing, the TransPred model containing CT images, dose distribution images, tumor target area outlining, and relevant clinical variables achieved the best results on the validation set and the internal test set, with the average AUC values of 0.92 and 0.88, and average kappa coefficients of 0.81 and 0.80, respectively.The traditional machine learning approach is based on univariate logistic regression analysis of dose D90, D95 variable , with average AUC values of 0.53 and 0.80 and kappa coefficients of 0.02 and 0.01 on the validation set for univariate analyses of D90 and D95, respectively. Conclusion: The baseline model in this study initially demonstrated an exemplary predictive capability, proving that the model is a promising recurrence prediction model, making the prediction of subsequent recurrence in the pre-radiotherapy period using the available data a promising reality, and thus providing a certain degree of assistance in the subsequent treatment of patients.