To evaluate the efficacy of 2D and 3D radiomics models based on contrast-enhanced CT and a combined model incorporating clinical features for preoperative assessment of lymphovascular invasion(LVI) in colorectal cancer.

A retrospective study included 303 patients who underwent radical surgery for colorectal cancer at Shanghai General Hospital from January 2015 to December 2019. Patients were categorized into LVI-positive (n=122) and LVI-negative (n=181) groups. The cohort was randomly divided into a training set (n=212; 90 LVI-positive, 122 LVI-negative) and a test set (n=91; 32 LVI-positive, 59 LVI-negative) in a 7:3 ratio. Clinical predictors of LVI were identified using chi-square tests, Mann-Whitney U tests, and multivariate logistic regression. Two radiologists manually delineated 2D regions of interest (ROI) and 3D volumes of interest (VOI) on venous-phase contrast-enhanced CT images. Radiomic features were selected via Mann-Whitney U test, intraclass correlation coefficient(ICC) and least absolute shrinkage and selection operator(LASSO) regression. Based on selected features, support vector machine(SVM), logistic regression(LR), random forest(RF), and extreme gradient boosting(XGBoost) algorithms were employed to construct 2D, 3D radiomics models, and combined models. Model performance was assessed using receiver operating characteristic(ROC) curves, area under the curve(AUC), and the DeLong test. Calibration curves and decision curve analysis were used to evaluate diagnostic accuracy and clinical utility.

N stage, perineural invasion, and tumor deposits were identified as independent clinical risk factors for LVI. Models built with the LR classifier demonstrated superior performance. The 2D radiomics model achieved AUC of 0.734 (95%CI: 0.660~0.798) and 0.683 (95%CI: 0.562~0.805) in the training and test sets, respectively. The 3D radiomics model yielded AUC of 0.784 (95%CI: 0.720~0.843) and 0.726 (95%CI: 0.602~0.830). The 2D combined model (integrating N stage, perineural invasion, and tumor deposits) achieved AUC of 0.843 (95%CI: 0.786~0.899) and 0.819 (95%CI: 0.723~0.905), while the 3D combined model showed AUC of 0.907 (95%CI: 0.865~0.943) and 0.874 (95%CI: 0.790~0.941). Calibration curves indicated comparable predictive accuracy between 2D and 3D models. Decision curve analysis revealed greater clinical net benefit for the 3D combined model. DeLong tests showed no significant performance differences between classifiers for 2D/3D models, but the 3D combined model outperformed the 2D radiomics model (P<0.05).

While 2D and 3D segmentation methods exhibit distinct advantages and limitations, the 3D combined model constructed with LR demonstrats superior predictive performance. Both 2D and 3D radiomics models enable effective preoperative prediction of LVI in colorectal cancer, offering valuable guidance for clinical decision-making.