An Mri-Based Machine-Learning Radiomics Approach for Preoperative Evaluation of Lymph Node Metastasis in Hilar Cholangiocarcinoma

Background: Synchronous lymph node metastasis (LNM) affects therapeutic regimes and prognosis for hilar cholangiocarcinoma (hCCA), which is evaluated by conventional imaging examinations with a high false-negative rate. Thus, a preoperative tool capable of predicting LNM status is urgently needed. The aim of this study was to develop an MRI-based radiomics-clinicopathologic approach for preoperative evaluation of LNM status in hCCA. Methods: 325 hCCA patients were enrolled to validate the prognostic effect of LNM on hCCA patients, and 134 cases who underwent preoperative MRI scans, primary surgery and lymph nodes dissection with pathological examination were enrolled to develop the radiomics-clinicopathologic models. The radiomics features were extracted using 3D Slicer software and reproducible features were used to construct radiomics signatures with several typical machine-leraning approaches. Its clinical applicability was compared with already established models for CCA, and its role in prognosis risk stratification was further investigated. Findings: The prognostic impact of LNM was validated in the Sun Yat-sen Memorial Hospital (SUSYM) cohort (hazard ratio [HR], 1.79 (1.36-2.34)). The optimal radiomics signature was based on an elastic network approach, distinguishing LNM status with parallel AUCs of 0.888 and 0.885 in the training and validation cohort, respectively. The radiomics-clinicopathologic model integrated the radiomics signature, carbohydrate antigen 199 level and MRI-reported LNM status, demonstrating better discriminating ability with AUC of 0.905, accuracy of 85.1%, sensitivity of 83.3%, and specificity of 88.0% in the entire cohort. Furthermore, the radiomics-clinicopathologic model could stratify patients into subgroups with different risks for overall survival (hazard ratio [HR], 2.29 (1.41-3.74)). The decision curve analysis showed that our radiomics-clinicopathologic nomogram held better clinical usefulness than clinicopathologic or radiomics signature alone, and excelled those established models for other cholangiocarcinoma.  Interpretation: The generated model was promising in non-invasive LNM evaluation and personalized therapeutic strategy decision-making. Funding: This study was supported by grants from China Postdoctoral Science Foundation (2021M693629), the National Natural Science Foundation of China (81972255, 81672412 and 81772597), the Guangdong Basic and Applied Basic Research Foundation (2021A1515010095). Declaration of Interest: We declare that we have no conflicts of interest. Ethical Approval: Ethical approval was obtained from the institutional review board of Sun Yat-sen Memorial Hospital (SYSMH) of Sun Yat-sen University.