Prediction of lymph node metastasis in colorectal cancer based on clinical data, body composition, and radiomics

Abstract

To develop and validate a positron emission tomography/computed tomography (PET/CT)-based model integrating clinical variables, body composition indices, and radiomics features for predicting lymph node metastasis (LNM) in colorectal cancer (CRC). This retrospective study included 233 CRC patients who underwent preoperative 18 F-FDG PET/CT. Body composition parameters were quantified from CT images, and radiomics features were extracted from primary tumors on PET and CT images. Only preoperatively available clinical variables were included. Extreme Gradient Boosting (XGBoost) models based on clinical, body composition, and radiomics features were constructed in the training set and evaluated in an independent internal test set. Model performance was assessed using the area under the curve (AUC), DeLong test, and decision curve analysis. LNM was present in 100/233 patients (42.9%). In the training set, multivariable logistic regression identified CEA, non-ulcerative macroscopic tumor type, PET/CT-reported lymph node status, and a lower SMA/SFA ratio as independent predictors of LNM. The integrated Body_Clinical_Radiomics model achieved an AUC of 0.794 (95% CI 0.724–0.860) in the training set and 0.780 (95% CI 0.659–0.887) in the internal test set. However, DeLong testing showed no statistically significant improvement over the Clinical_Radiomics model. The integrated PET/CT-based model showed promising performance for preoperative LNM prediction in CRC. SMA/SFA may provide complementary host-related information, although its incremental discriminatory value requires further validation.