Integrating Deep Learning Radiomics and Node-RADS for Interpretable Prediction of Colorectal Cancer Lymph Node Metastasis

Abstract Background and purpose: To develop and validate an interpretable multimodal fusion model integrating deep learning radiomics (DLR) and Node-RADS score for preoperative prediction of lymph node metastasis (LNM) in colorectal cancer (CRC). Methods: This retrospective multicenter study included 892 CRC patients (training: Center 1, n=684; external validation: Centers 2/3, n=208). Venous-phase CT tumor regions were segmented. 1394 radiomic and 320 ResNet-50 deep learning features were extracted; LASSO was used for feature selection. An XGBoost DLR model was integrated with Node-RADS score and clinical variables to construct a multimodal nomogram. Performance was assessed by AUC, calibration, decision curve analysis, and the DeLong test. SHAP analysis provided model interpretability. Results: LASSO regression identified 24 deep learning features and 9 radiomic features with non-zero coefficients, wherein deep learning features demonstrated higher predictive weights. The XGBoost-based DLR model achieved AUCs of 0.876 and 0.798 in training and external validation cohorts, respectively. The multimodal fusion model, incorporating Node-RADS score and selected imaging features, yielded significantly improved AUCs of 0.900 (95%CI:0.874–0.926) and 0.870 (95%CI:0.821–0.919), outperforming both the DLR model and the clinical model (AUC 0.751/0.641) (all DeLong P< 0.05). Calibration curves demonstrated excellent agreement (Hosmer–Lemeshow P > 0.05), and decision curve analysis confirmed superior net clinical benefit across wide threshold ranges. SHAP analysis identified DL_27 and lbp_3D_m2_glszm_ZoneEntropy as the dominant predictive features. Conclusion: This externally validated, interpretable CT-based multimodal model integrating DLR signatures and Node-RADS score enables accurate preoperative prediction of LNM in CRC, offering a noninvasive decision-support tool to guide personalized treatment.