3759 - TransPred: A Novel 3D Deep Learning Model for Predicting Recurrence Risk of Nasopharyngeal Cancer after Radiotherapy

Purpose/Objective(s): Radiotherapy is one of the main treatment methods for nasopharyngeal carcinoma (NPC). However, approximately 10% of patients experience recurrence at the primary site or regional lymph nodes after radiotherapy. With the continuous development of deep learning in the field of computer vision, this study aims to develop a novel prediction model for NPC recurrence by integrating patients' imaging and clinical information, in order to predict early recurrence after radiotherapy, identify high-risk recurrence patients, and facilitate timely intervention and re-irradiation.

Materials/Methods: This study is a retrospective multi-center experiment. A total of 586 nasopharyngeal carcinoma (NPC) patients who underwent IMRT treatment were included. The training and internal test data were from patients at Fudan University Shanghai Cancer Center from 2009 to 2019, while the external test data were from NPC patients at Quzhou People's Hospital (2023-2024). The 3D prediction model, TransPred, was developed using five-fold cross-validation on the training and validation cohorts and was independently tested on the internal and external test sets. The model uses patients' initial CT images, dose distribution images, PTV (planning target volume) delineation data, and relevant clinical features as primary inputs to predict the risk of NPC recurrence after radiotherapy. In addition to ensuring the model's accuracy, SHAP and Grad-CAM methods were used to perform visual analyses of the imaging and clinical features, respectively.

Results: After independent testing, the model developed in this study, TransPred, achieved an average AUC of 0.89 and 0.79 on the internal and external test sets, respectively, with average Kappa coefficients of 0.54 and 0.36. In the comparison experiments of input variables, the model without clinical variables performed the worst (AUC=0.58), while the model using only dose distribution images (AUC=0.80) outperformed the model using CT images alone (AUC=0.78). In the model comparison experiments, TransPred showed a significant improvement over the traditional radiomics model (AUC=0.799) and the ResNet model (AUC=0.77), with AUC values increased by approximately 0.12 and 0.09, respectively.

Conclusion: The 3D prediction model proposed in this study achieved good predictive performance, demonstrating that radiotherapy-related imaging data and relevant clinical features of patients can provide sufficient and effective information for predicting nasopharyngeal carcinoma recurrence based on this model. It holds promise for supporting clinical decision-making.