Development and validation of the risk stratification based on deep learning and radiomics to predict survival of advanced cervical cancer.
Authors
Affiliations (6)
Affiliations (6)
- Department of Obstetrics and Gynecology, Qilu Hospital of Shandong University, No.107 Wenhua West Road, Jinan City, 250012, Shandong Province, China.
- Department of Obstetrics and Gynecology, School of medicine, University of Lubumbashi, Lubumbashi, Democratic Republic of Congo.
- Medical Management Service Center of Health Commission of Shandong Province, Jinan, China.
- Department of Medical Imaging, Affiliated Hospital of Jining Medical University, Jining, China.
- Department of Gynecology, Weifang People's Hospital, Weifang, China.
- Department of Obstetrics and Gynecology, Qilu Hospital of Shandong University, No.107 Wenhua West Road, Jinan City, 250012, Shandong Province, China. [email protected].
Abstract
Advanced cervical cancer (aCC) is associated with a poor prognosis. This study aimed to develop and validate a deep learning-based risk stratification model to predict overall survival in aCC patients using pre-treatment CT images. A total of 396 patients with aCC were retrospectively enrolled and randomly allocated into training (n = 198) and validation (n = 198) cohorts. A deep learning model integrating a Vision Transformer (ViT) for feature extraction and a Recurrent Neural Network (RNN) for sequence modeling was developed to generate a prognostic radiomic signature (Rad-score) from baseline CT scans. The Rad-score was incorporated into a Cox proportional hazards model alongside clinical variables to construct an integrative nomogram. The model's performance was evaluated using the concordance index (C-index), time-dependent receiver operating characteristic (ROC) analysis, calibration curves, and decision curve analysis (DCA). Multivariate Cox regression identified the Rad-score as a strong independent prognostic factor (Hazard Ratio [HR] = 4.06, 95% confidence interval [CI]: 2.46-6.70, p < 0.001). The integrative nomogram achieved C-indexes of 0.784 (95% CI: 0.733-0.835) and 0.726 (95% CI: 0.677-0.785) in the training and validation cohorts, respectively. Calibration and DCA curves indicated good clinical utility. Kaplan-Meier analysis confirmed that the model-based risk stratification significantly discriminated between high- and low-risk patients (p < 0.001). The proposed deep learning-based nomogram offers a non-invasive and reproducible tool for predicting survival in aCC patients. It shows potential for assisting clinicians in making personalized treatment decisions and warrants further validation in prospective, multi-center studies before widespread clinical application.