Primary tumor-derived, multiparametric MRI-based deep learning-radiomics-clinical model for predicting lymph node metastasis in early-stage cervical cancer

Abstract Objectives To develop and validate a primary tumor-derived, multiparametric MRI-based deep learning-radiomics-clinical (DLRC) model for predicting pelvic lymph node metastasis (LNM) in early-stage cervical cancer. Materials and methods This retrospective five-center study selected 1095 patients (Jan 2020-Dec 2022), divided into training (n = 481), internal validation (n = 204), and external validation (n = 410) cohorts. Radiomics and deep learning (DL) features were extracted from the volumetric segmentations of the primary cervical tumors on three MRI sequences (CE-T1WI, DWI, FS-T2WI). After constructing individual radiomics and DL models, the DLRC model was developed by integrating the radiomicsₛcore, optimal DL model, and significant clinical features. Model performance was evaluated using ROC analysis, calibration curves, and decision curve analysis. Results The DLRC model demonstrated superior predictive performance, achieving AUCs of 0. 807 (95% CI: 0. 766–0. 849) in the training cohort, 0. 789 (95% CI: 0. 721–0. 857) in the internal validation cohort, and 0. 807 (95% CI: 0. 761–0. 853) in the external validation cohort. It significantly outperformed both the radiomics model and the optimal DL model (all p < 0. 001) in the external validation cohort. The calibration curves indicated good agreement between predictions and observations. The decision curve analysis showed that the DLRC model provided the highest net clinical benefit across most decision thresholds. Conclusions The DLRC model, which integrates tumor-derived multiparametric MRI features with clinical features, represents a robust and generalizable tool for the preoperative prediction of LNM. Its comparable accuracy to standardized radiological assessment and clinical utility shows potential to aid in personalized treatment planning for patients with early-stage cervical cancer. Critical relevance statement The combined model (DLRC) integrating deep learning and radiomics features from the primary tumor with clinical characteristics enables preoperative LNM risk stratification, supporting personalized surgical planning and reducing unnecessary lymphadenectomy. Key Points Accurate preoperative prediction of lymph node metastasis in early-stage cervical cancer remains a significant clinical challenge. The model integrating deep learning and radiomics features derived from the primary tumor with clinical features achieved robust and generalizable predictive performance. The accuracy of a deep learning-radiomics-clinical nomogram for lymph node metastasis risk stratification in early-stage cervical cancer is comparable to standardized radiological assessment. Graphical Abstract