MRI-Based Radiomics Model for Pretreatment Prediction of Pathological Lymph Node Metastasis in Neuroendocrine Carcinoma of the Cervix: Dual-Center Development and External Validation

Abstract

Objectives: To develop and externally validate MRI-based radiomics and clinicoradiomics models for predicting pathological lymph node metastasis (LNM) in neuroendocrine carcinoma of the cervix (NECC). Methods: This retrospective two-center study included 133 patients who underwent pretreatment pelvic MRI and pathological nodal assessment: 93 for development and 40 for external validation. Intratumoral and peritumoral features were extracted from T2-weighted, diffusion-weighted, and contrast-enhanced T1-weighted images. Stratified nested fivefold cross-validation evaluated the complete revised selection process. Locked models were assessed for discrimination, calibration, and exploratory net benefit, with conventional MRI nodal assessment as a comparator. Results: Pathological LNM was present in 54/133 patients (40.6%). The selected radiomics-only k-nearest neighbors (KNN) model had an external AUC of 0.783 (95% CI, 0.630 to 0.935), accuracy of 0.800, sensitivity of 0.813, and specificity of 0.792. The complete selection procedure yielded an out-of-fold AUC of 0.731. External AUCs were 0.727 for the clinical/MRI model and 0.767 for the clinicoradiomics model; neither fusion model differed significantly from the clinical/MRI comparator. Conventional MRI sensitivity was 0.563, with no significant paired difference from the selected model (p = 0.125). External calibration was imperfect (Brier score, 0.186; slope, 0.672). Conclusions: MRI radiomics showed moderate external discrimination of pathological LNM in NECC. Adding clinical variables did not demonstrate further benefit. Multicenter validation, recalibration, and clinical-impact assessment are needed before clinical application.