Development and validation of a multi-task deep learning model integrating PET-CT radiomics, clinical variables, and EBV DNA for prognostic prediction in locally advanced nasopharyngeal carcinoma.

Background: Accurate prognostic prediction in locally advanced nasopharyngeal carcinoma (LA-NPC) remains challenging due to tumor heterogeneity and complex treatment responses. Existing prognostic models relying on single imaging modalities or traditional machine learning algorithms inadequately capture the complex non-linear interactions among metabolic, anatomical, and molecular biomarkers, and fail to leverage the synergistic information between functional metabolic data from positron emission tomography (PET) and anatomical structural information from computed tomography (CT). This study developed and validated a multi-task deep learning (MTDL) model integrating PET-CT imaging, clinical variables, and Epstein-Barr virus (EBV) DNA to predict overall survival (OS) and progression-free survival (PFS) in LA-NPC patients. Methods: A total of 200 LA-NPC patients [American Joint Committee on Cancer (AJCC) 8th edition stage III-IVA] aged 18-75 years with World Health Organization (WHO) type II/III histology, Eastern Cooperative Oncology Group (ECOG) performance status 0-1, and complete pre-treatment imaging data, who received concurrent chemoradiotherapy (CCRT) between January 2018 and December 2022 were retrospectively enrolled from a single institution. Patients were randomly divided into training (n=140, 70%) and internal validation (n=60, 30%) cohorts with stratification by tumor-node-metastasis (TNM) stage. The MTDL model incorporated dual-pathway three-dimensional (3D) convolutional neural networks for simultaneous extraction of metabolic features from PET and anatomical features from CT, integrated with clinical variables and EBV DNA levels. Pre-treatment 18F-fluorodeoxyglucose (18F-FDG) PET-CT images were preprocessed with standardized uptake value (SUV) normalization, and multi-modal feature fusion was performed through an attention mechanism. The model simultaneously predicted 3-year OS and PFS through shared feature representations with task-specific output heads. Results: The study included 200 patients with a median age of 47 years (range, 24-72 years) and a median follow-up of 36 months (range, 6-60 months). The MTDL model achieved superior predictive performance with concordance indices (C-indices) of 0.834 [95% confidence interval (CI): 0.805-0.863] for OS and 0.821 (95% CI: 0.790-0.852) for PFS in the internal validation cohort, significantly outperforming TNM staging (C-indices: 0.648 and 0.631, respectively); both P<0.001 and single-task models (C-index improvements: 0.063 and 0.071, respectively; both P<0.01). The model identified high-risk patients with significantly worse outcomes (3-year OS: 54.20% Conclusions: The MTDL model leveraging complementary metabolic and anatomical information from PET-CT demonstrated promising performance in predicting the prognosis of LA-NPC patients in an internal validation setting, outperforming conventional staging systems and single-modality approaches. However, prospective external validation in independent multi-center cohorts is warranted before clinical implementation. This model may facilitate personalized treatment strategies and risk-adapted therapy intensification.