2873 - A Deep-Learning Model Integrating Multi-Modality Data for Predicting the Recurrence of Nasopharyngeal Carcinoma
Presenter(s)
Y. Wang1, Y. Zhao2, Y. N. Zhang2, K. Y. You3, Z. Q. Liu4, Y. He5, H. M. Wang6, S. S. Yang7, X. Lan8,9, Y. S. Wu10, J. Tang11, D. Li12, J. N. Liu13, S. Liang14, G. Zou11, W. J. Zhang6, J. G. Guo15, F. Y. Xie16, and P. Y. OuYang2; 1Department of Radiation Oncology, Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Collaborative Innovation Center of Cancer Medicine, Guangzhou, China, 2Department of Radiation Oncology, Sun Yat-sen University Cancer Center, State KeyLaboratory of Oncology in South China, Guangdong Key Laboratory of NasopharyngealCarcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center forCancer, Guangzhou, China, 3Department of Radiation Oncology, Sun Yat-sen Memorial Hospital, Sun Yat-sen University, Guangzhou, Guangdong, China, Guangzhou, China, 4Department of Nuclear Medicine,The Affiliated Hospital of Southwest Medical University, Luzhou, Sichuan, China, Luzhou, China, 5Department of Radiology,Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangzhou, Chin, Guangzhou, China, 6Department of Radiation Oncology, Affiliated Cancer Hospital & Institute of Guangzhou Medical University, Guangzhou, China., Guangzhou, China, 7Department of Radiation oncology, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, Shandong, China., Jinan, China, 8Guangdong Provincial Key Laboratory of Malignant Tumor Epigenetics and Gene Regulation, Medical Research Center, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China, 9Guangdong Provincial Key Laboratory of Malignant Tumor Epigenetics and Gene Regulation, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, China 510120., Guangzhou, China, 10Department of Nasopharyngeal Carcinoma, Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy,, Guangzhou, China, 11Department of Oncology, Affiliated Panyu Central Hospital of Guangzhou Medical University, Guangzhou, China., Guangzhou, China, 12State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, Guangdong, China, 13Department of Head and Neck Oncology, The Cancer Center of the Fifth Affiliated Hospital of Sun Yat-sen University, Zhuhai, China, Zhuhai, China, 14Department of Radiation Oncology, First People's Hospital of Foshan, Foshan, China, 15Department of Radiation Oncology, The First People's Hospital of Foshan, Foshan, China, Foshan, China, 16Department of Radiation Oncology, Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Guangzhou, China
Purpose/Objective(s): Predicting nasopharyngeal carcinoma (NPC) recurrence is critical for improving patient outcomes. Current post-radiotherapy follow-up strategies exhibit inefficiencies that may lead to both healthcare resource overutilization and delayed detection of recurrence. To address these challenges, we developed a Transformer-based deep learning model integrating multi-modality data (MRI, CT, dose maps, DVH, and clinical parameters) for NPC recurrence prediction and follow-up strategy optimization, followed by external validation.
Materials/Methods: The model architecture comprises three phases: feature extraction, fusion, and pattern learning. Multi-modality data were first processed through dedicated encoder blocks to extract modality-specific features. These features were then aligned and fused in a shared encoding space. A Transformer network with multi-head attention mechanisms subsequently captured complex inter-modality relationships, followed by an MLP layer generating recurrence risk scores. To rigorously evaluate generalizability, two independent datasets were utilized: an internal dataset with 1,806 samples from a single institution (training/validation/internal testing: 938/192/676) and an external dataset with 483 multicenter samples for external testing.
Results: The model demonstrated robust performance across both datasets. On internal validation, it achieved a C-index of 0.920 (95% CI: 0.817–0.968; mean: 0.919), an integrated Brier score (IBS) of 0.1037 indicating low probability estimation error, and an AUC of 0.9989 reflecting near-perfect classification accuracy. External validation showed a C-index of 0.824 (95% CI: 0.748–0.894; mean: 0.823), IBS of 0.1721, and AUC of 0.9884. Follow-up efficiency was substantially improved: for positive recurrence risk patients, follow-up frequency decreased from 14 to 5 sessions (internal dataset) and 14 to 6 sessions (external dataset), while negative-risk patients required no follow-up (0 sessions) in both cohorts.
Conclusion: We present a multimodal AI model integrating imaging, dosimetric, and clinical data for NPC recurrence prediction. The model achieved high prognostic accuracy (C-index >0.82) across internal and external validations while significantly optimizing follow-up strategies—reducing unnecessary examinations by 64–100%. This approach enhances early recurrence detection efficiency and alleviates healthcare resource burdens, demonstrating substantial clinical translation potential.