Preoperative prediction of lymphatic metastasis in rectal cancer using a fusion model based on multiparameter magnetic resonance imaging: a retrospective validation study.

Background: To validate an MRI-based deep learning algorithm for predicting lymphatic metastasis in rectal cancer (RC) and to construct an integrated fusion model combining imaging and clinicopathological factors to improve preoperative diagnostic performance. Methods: This study retrospectively included 127 patients with RC as a primary cohort and 33 patients from two other centers as an external validation cohort. All patients underwent radical resection for RC without preoperative radiotherapy or chemotherapy. Based on the MR images, the lymph nodes were interpreted by a previously constructed prediction algorithm and two radiologists independently. The clinical factors associated with lymph node metastasis (LNM) were screened by a logistic regression model and then combined with the prediction algorithm to construct a fusion model. The area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA) was used to evaluate the predictive power and clinical utility. Results: In the primary cohort, the prediction algorithm achieved an AUC of 0.760 (95% CI: 0.674-0.846), significantly outperforming the two radiologists [AUC: 0.665 and 0.676; interobserver kappa = 0.241]. Multivariate analysis identified a carcinoembryonic antigen (CEA) level >5 µg/L and poor differentiation as independent risk factors for LNM. The integrated fusion model (algorithm + CEA + differentiation) demonstrated superior performance with an AUC of 0.873 (95% CI: 0.804-0.972). In the external validation cohort, the fusion model showed promising diagnostic efficacy with an AUC of 0.838 (95% CI: 0.697-0.979). Decision curve analysis further confirmed that the fusion model provided higher clinical net benefit than default treatment strategies across a wide range of threshold probabilities. Conclusion: The preoperative fusion model significantly improves the accuracy of N-staging in rectal cancer. By outperforming manual interpretation and demonstrating encouraging preliminary generalizability in an external cohort, this model shows potential as a clinical aid for identifying patients who will benefit from neoadjuvant therapy, thereby facilitating personalized clinical decision-making and reducing interobserver variability.