Preoperative prediction of lymphovascular invasion and prognostic risk stratification in gastric cancer using multiparametric MRI radiomics and machine learning: a retrospective model development and internal validation study

Lymphovascular invasion (LVI) is associated with recurrence and adverse outcomes in gastric cancer, but it is generally established after surgery. This study developed and internally validated an interpretable model integrating clinical variables and multiparametric magnetic resonance imaging (MRI) radiomics for preoperative LVI prediction and explored its association with disease-free survival (DFS). In this retrospective single-center study, 458 patients were randomly assigned to a training cohort ( n = 320) and an internal test cohort ( n = 138). Radiomics features were extracted from fat-suppressed T2-weighted, apparent diffusion coefficient, and venous-phase contrast-enhanced MRI. Three logistic-regression models were constructed for predictor-set comparison, with the combined clinical-radiomics logistic-regression model designated as the primary model. Secondary algorithm comparisons used fold-restricted random oversampling. Performance was assessed by discrimination, calibration, and decision-curve analysis. Exploratory DFS groups used an X-tile cutoff selected in training and applied unchanged to testing; a training-median cutoff was examined as a sensitivity analysis. Four clinical predictors and six radiomics features were retained. The primary combined clinical-radiomics logistic-regression model achieved an AUC of 0.775 in training and 0.752 (bootstrap 95% CI, 0.663–0.840) in the internal test cohort. At the training-derived Youden threshold of 0.217, test sensitivity was 0.795, specificity 0.564, F1 score 0.583, balanced accuracy 0.680, and negative predictive value 0.855. The test calibration intercept, slope, and Brier score were 0.092, 0.965, and 0.180, respectively. Exploratory DFS groups were defined using the training-derived X-tile cutoff (0.59) and evaluated unchanged in testing; a training-median cutoff was added as a sensitivity analysis. Model probability remained associated with DFS in an adjusted complete-case Cox model (hazard ratio per SD, 1.57; 95% CI, 1.21–2.03). The clinical-radiomics model showed moderate discrimination and acceptable calibration in a same-center internal test cohort. Its high negative predictive value at a sensitivity-oriented threshold suggests a potential triage role, but the model is not ready for stand-alone clinical use and requires prospective external validation. The DFS findings are exploratory.