Prediction of Lymph Node Metastasis Risk in Endometrial Cancer Using Multi-Sequence Magnetic Resonance Imaging Combined with Deep Transfer Learning

Purpose: To develop and externally validate a nomogram integrating clinical predictors, multisequence magnetic resonance imaging (MRI) radiomics, and deep learning (DL) features for preoperative lymph node metastasis (LNM) prediction in endometrial cancer (EC), and assess the incremental value of DL. Methods: This two-center retrospective study included 520 patients: 268 for training, 117 for held-out internal validation, and 135 for external validation. Radiomics features were extracted from four MRI sequences, and DL features from four ImageNet-pretrained 50-layer residual network (ResNet-50) branches fine-tuned in training. The Clinical+Radiomics model combined four clinical predictors with a radiomics signature; the nomogram additionally incorporated the DL signature. Final models were fitted using all training patients and evaluated unchanged in both validation cohorts. Assessments included DeLong tests, calibration, decision curve analysis, and prespecified sensitivity analysis in magnetic resonance (MR) report-based LNM-negative patients. Results: In internal and external validation, respectively, the nomogram achieved areas under the receiver operating characteristic curve (AUCs) of 0.866 and 0.886, with 95% confidence intervals (CIs) of 0.786–0.946 and 0.817–0.955. Corresponding Clinical+Radiomics AUCs were 0.733 (95% CI: 0.591–0.875) and 0.715 (95% CI: 0.594–0.836), favoring the nomogram ( P =0.024 and P <0.001, respectively). Brier scores were 0.110 and 0.083. Decision curves favored the nomogram over comparators at probability thresholds of 0.02–0.39 and 0.01–0.71, respectively. Nomogram AUCs in MR report-based LNM-negative subgroups were 0.843 and 0.858, respectively. Conclusion: Adding DL improved discrimination beyond Clinical+Radiomics. The nomogram may complement preoperative LNM risk stratification but should not replace pathological nodal staging. Prospective multicenter validation is required before clinical implementation.