Development and validation of a model for predicting central lymph node metastasis in colorectal cancer based on CT textural features and clinical indicators.

Background: Central lymph node metastasis (CLNM) in colorectal cancer (CRC) is an important factor affecting patient prognosis and the assessment of surgical extent. However, conventional imaging remains limited in the preoperative assessment of LNM. This study aimed to investigate the value of a combined model based on computed tomography (CT) texture features and clinical indicators for the preoperative identification of CLNM in patients with CRC. Methods: In this retrospective study, 1,187 patients who underwent laparoscopic radical resection for CRC at two hospitals were screened, and 228 were ultimately included. Based on postoperative pathology, patients were divided into the CLNM-positive group (n=76) and the CLNM-negative group (n=152). Patients from one hospital were randomly assigned to a training set (n=126) and a test set (n=54) in a 7:3 ratio, while patients from the other hospital were used for the external validation set (n=48). Radiomic features extracted using MaZda software were used to construct the radiomics model, while preoperative clinical indicators were used to construct the clinical model. Subsequently, radiomic features and clinical indicators were incorporated to construct five combined predictive models, including logistic regression (LR), support vector machine (SVM), neural network (NN), random forest (RF), and gradient boosting machine (GBM). Their performance was compared using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). Results: Significant differences in T stage, N stage, preoperative serum carcinoembryonic antigen (CEA), and neutrophil-to-lymphocyte ratio (NLR) were observed between the CLNM-positive and CLNM-negative groups (all P<0.05). S(1,0)DifEntrp and S(0,1)SumEntrp were selected for the radiomics model, while CEA and NLR were selected for the clinical model. The combined model incorporated all four predictors and performed better than the clinical and radiomics models in both the test and external validation sets. It achieved areas under the curve (AUCs) of 0.816 and 0.775, respectively. Among the five models, GBM, RF, and NN showed larger AUC differences between the test set and the external validation set, with differences of 0.349, 0.208, and 0.096. In contrast, the differences were smaller for LR and SVM, at 0.041 and 0.003, respectively. Based on the ROC curves, calibration curves, and DCA, the LR model showed better overall predictive performance and stability in this study. Conclusions: The combined predictive model based on CT texture features and clinical indicators showed potential value for the preoperative identification of CLNM in CRC. It may provide an additional reference for the preoperative assessment of CLNM in patients with CRC.