Main Session

Sep 29

PQA 07 - Head and Neck Cancer, Lung Cancer/Thoracic Malignancies, and Nursing and Supportive Care

3536 - NPC-SurvNet: A Dosimetry-Aware Longitudinal Multi-Modal Transformer for Precision Prognostication in Nasopharyngeal Carcinoma

03:45pm - 05:00pm ET

Poster Hall - Exhibit Hall A

Screen: 8

POSTER

Presenter(s)

Jiping Liu, DSc, RT - Zhejiang Cancer Hospital, Hangzhou, Zhejiang

J. Liu¹, Q. Zhou², M. Feng², D. Zheng³, Y. Dong¹, H. Chen¹, H. Zeng⁴, Z. Jia⁵, and X. Li²; ¹Zhejiang Cancer Hospital, Hangzhou, Zhejiang, China, ²University of Electronic Science and Technology of China, Chengdu, Sichuan, China, ³Southwest Petroleum University, Guangzhou, Guangdong, China, ⁴The First Affiliated Hospital of Yangtze University, Jingzhou, Hubei, China, ⁵The First People's Hospital of Kashi, Kashi, Xinjiang, China

Purpose/Objective(s):

Accurate prognostication for nasopharyngeal carcinoma (NPC) remains challenging due to the complex interplay between tumor burden and treatment intensity. Traditional TNM staging and static pre-treatment imaging often fail to capture the dynamic tumor response and the spatial heterogeneity of radiation dose deposition. This study aims to develop NPC-SurvNet, a novel longitudinal multi-modal Transformer that synergizes 3D dosimetric maps, anatomical CT, and longitudinal MRI data to achieve precision prediction of 5-year progression-free survival (PFS).

Materials/Methods:

A dual-center retrospective cohort of 515 patients with NPC (332 from Center A and 183 from Center B) was analyzed. We engineered NPC-SurvNet, a hierarchical framework designed to decode the spatiotemporal evolution of the tumor. The model uniquely integrates four distinct data streams: (1) Pre-treatment CT, (2) 3D Dose Distribution Maps (derived from TPS), (3) Longitudinal MRI (pre- and post-treatment), and (4) Clinical characteristics. A specialized Siamese Network was employed to extract longitudinal tumor response features, while a Feature-wise Linear Modulation (FiLM) mechanism dynamically fused anatomical and dosimetric features, conditioned on clinical context to enhance prognostic precision. The model was benchmarked against state-of-the-art deep learning architectures for predicting 5-year PFS, distant metastasis (DM), and locoregional recurrence (LRR).

Results:

NPC-SurvNet demonstrated superior prognostic performance, significantly outperforming all comparative baselines. In the external test cohort, the model achieved a robust AUROC of 0.819 for 5-year PFS and 0.856 for 5-year DM and 0.824 for 5-year LRR. Kaplan-Meier survival analysis confirmed the model’s clinical utility, effectively stratifying patients into distinct high- and low-risk groups with statistically significant separation (p < 0.05).

Conclusion:

NPC-SurvNet sets a new benchmark for NPC prognostication by effectively leveraging dosimetric data alongside longitudinal imaging dynamics. Our findings highlight the indispensable role of radiation dose distribution maps in deep learning-based survival analysis—a dimension often neglected in conventional radiomics. This high-precision model offers a powerful tool for early risk stratification, facilitating personalized surveillance strategies for high-risk patients. Future work will focus on validating these findings in large-scale prospective trials and exploring the integration of biological markers to further enhance predictive robustness.

Tab. 1 Benchmarking: Comparison of AUROC showing NPC-SurvNet outperforming conventional models.

Center A	Models	5-year PFS	5-year DM	5-year LRR
	CNN	0.836	0.830	0.763
	ResNet	0.805	0.836	0.765
	DenseNet	0.840	0.825	0.796
	ViT	0.828	0.836	0.750
	Ours	0.877	0.860	0.865
Center B	Ours	0.819	0.858	0.824