Deep learning model using contrast-enhanced CT for predicting overall survival in oropharyngeal squamous cell carcinoma: a prospective multicenter study.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, The Affiliated Hospital of Qingdao University, Qingdao, China.
- Department of Pathology, The Affiliated Hospital of Qingdao University, Qingdao, China.
- Department of Radiology, The First Affiliated Hospital of Xiamen University, Xiamen, China.
- Department of Radiology, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, China.
- College of Computer Science and Technology, Qingdao University, Qingdao, China. [email protected].
- Department of Radiology, The Affiliated Hospital of Qingdao University, Qingdao, China. [email protected].
Abstract
Accurate risk stratification for overall survival (OS) in patients with oropharyngeal squamous cell carcinoma (OPSCC) is critical for guiding personalized treatment and surveillance. A deep learning (DL) model was developed and validated to estimate OS in OPSCC patients based on contrast-enhanced computed tomography (CECT). A total of 269 patients from three centers were retrospectively enrolled and divided into training (n = 144), internal validation (n = 56), and external validation (n = 69) cohorts. An additional prospective cohort (n = 50) was used for HPV-based subgroup analysis. Clinical and semantic CT features were selected via multivariate Cox regression. Radiomic features were extracted using PyRadiomics. A Swin Transformer V2-based DL model was developed to estimate OS risk. SHAP values and Grad-CAM were used to assess model interpretability. Combined models were constructed by integrating clinical factors, handcrafted radiomics features, and DL score values. Predictive accuracy and clinical utility of the model were evaluated through C-index, time-dependent ROC analysis, calibration assessment, and decision curve analysis. Kaplan-Meier analysis was performed based on the optimal cutoff identified by the X-tile. Deep learning-clinical signature (DLCS) achieved the best performance, with C-indices of 0.856 (internal) and 0.783 (external). It provided accurate 1-, 3-, and 5-year OS predictions, robust calibration, and clinical benefit. DLCS effectively stratified patients into high- and low-risk groups and maintained predictive consistency across HPV subgroups. The CECT-based DLCS offers reliable OS prediction and risk stratification in OPSCC, supporting individualized clinical decision-making and precision oncology. Question Prognostic prediction in OPSCC is crucial for personalized treatment, yet existing clinical tools lack accuracy and non-invasiveness. Findings The DLCS integrates imaging and clinical data to achieve precise OS prediction and robust risk stratification in OPSCC. Clinical relevance The DLCS offers a novel, non-invasive, and accurate prognostic tool for OPSCC, enabling differentiation between high- and low-risk patients. This supports personalized treatment strategies and optimizes clinical decision-making.