A Stacking Fusion Model Integrating MRI Radiomics and Deep Learning for Prediction of Occult Lymph Node Metastasis in Cervical Cancer

Abstract Background To develop and validate a stacking model integrating MRI-based radiomics and deep learning model outputs for the preoperative prediction of occult lymph node metastasis (OLNM) in early-stage cervical cancer (CC), and to explore its potential complementary value in preoperative risk stratification. Methods This dual-center retrospective study included 392 patients with early-stage CC. Radiomics features were extracted from preoperative T2WI-FS and DWI sequences, and feature selection was conducted using intraclass correlation coefficient (ICC), Spearman correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression to construct the radiomics model. Deep learning models, including 2D-DL and multichannel 2.5D-DL, were developed separately for T2WI-FS and DWI sequences based on the ResNet50 architecture. A stacking model was subsequently established by integrating the output probabilities of the radiomics, 2D-DL, and multichannel 2.5D-DL models using an ensemble learning strategy. Model performance was evaluated using the AUC, and differences in AUCs between models were compared using the DeLong test. Calibration was assessed using calibration curves. Results The incidence of OLNM was 14% in the training cohort and 16% in the external validation cohort. The stacking model achieved the highest AUCs in the training, internal validation, and external validation cohorts (0.983, 0.932, and 0.868, respectively), and was numerically superior to multiple single-sequence models. The calibration curve demonstrated good agreement between the predicted probabilities and actual outcomes, and the decision curve analysis indicated a favorable net clinical benefit. Conclusions The stacking model proposed in this study demonstrated good discriminative ability and stability for the preoperative assessment of OLNM in patients with early-stage CC. By integrating outputs from radiomics and deep learning models, this approach may, to some extent, complement conventional imaging assessment and provide additional information for preoperative OLNM risk stratification.