Models Developed Based on Baseline Gastric Cancer and Metastatic Lymph Node CT Radiomics and Clinical Features for Predicting Early Postoperative Lymph Node Recurrence

Objective: To develop models based on baseline clinical and computed tomography (CT) radiomic features of primary tumors and metastatic lymph nodes to predict early lymph node recurrence after radical gastrectomy in patients with gastric cancer. Methods: The preoperative computed tomography (CT) and clinical data of 200 consecutive patients diagnosed with gastric cancer and lymph node metastasis who underwent radical surgery at Medical Centers 1 and 2 were collected retrospectively. Cases from Medical Center 1 were randomly assigned to a training group (n = 110) and an internal validation group (n=48) in a 7:3 ratio. Cases from Medical Center 2 were assigned to an external validation group (n = 42). The regions of interest of the primary tumors and metastatic lymph nodes were marked on the CT images, and their corresponding features were extracted. Using an independent sample t-test or U-test, statistically significant radiomics and clinical features were selected. LASSO regression analysis was used to obtain the core features of the primary tumors and metastatic lymph nodes. Subsequently, a clinical model, radiomic models of the primary tumors and metastatic lymph nodes, and radiomic models for the primary tumors and metastatic lymph nodes individually combined with clinical features were constructed. The predictive performance of the models was evaluated and compared using the area under the receiver operating characteristic (ROC) curve, sensitivity, specificity, DeLong’s test, and calibration curve. Results: Fourteen radiomic features of the primary tumor, 12 radiomic features of metastatic lymph nodes, and three clinical features were selected to construct individual prediction models. The clinical features included the number of metastatic lymph nodes, lymph node morphology, and tumor markers. The AUC, sensitivity, and specificity of the radiomics model established based on the primary tumor were 0.844, 0.868, and 0.706 in the training group; 0.802, 0.879, and 0.600 in the internal validation group; and 0.791, 0.714, and 0.786 in the external validation group, respectively. The AUC, sensitivity, and specificity of the radiomics model based on metastatic lymph nodes were 0.898, 0.753, and 0.941 in the training group, 0.842, 0.879, and 0.667 in the internal validation group, and 0.825, 0.828, and 0.769 in the external validation group, respectively. In the training, internal validation, and external validation groups, the DeLong test showed that the AUC values of the combined model integrating primary tumor radiomic features and clinical features were 0.970, 0.961, and 0.976, respectively. The AUC values of the combined model integrating metastatic lymph node radiomic features and clinical features were 0.943, 0.957, and 0.977, respectively. In the training, internal validation, and external validation groups, there were significant differences in the AUC between the primary tumor radiomics model, metastatic lymph node radiomics model, clinical model, and the combined models by integrating the primary tumor radiomics features or the metastatic lymph node radiomics features with clinical features (P < 0.05). Conclusion: The preoperative metastatic lymph node CT radiomics model was more effective than the primary tumor radiomics model in predicting early lymph node recurrence after radical gastrectomy for gastric cancer.