A machine-learning study published in Cancer Imaging reports that radiomics features extracted from CT venous-phase images can distinguish early-stage gastric cancer (T1–T2) from advanced disease (T3–T4) before surgery, with a logistic regression model achieving the strongest discriminative performance among four tested classifiers.
The retrospective study enrolled 223 patients with pathologically confirmed gastric cancer, 66 with early-stage and 157 with advanced-stage disease, treated between January 2022 and May 2025. All patients underwent enhanced CT within one week before surgery. The cohort was split by stratified random sampling into a training set of 156 patients and an internal hold-out test set of 67 patients at a 7:3 ratio.
Venous-phase images were selected for analysis, and three-dimensional regions of interest covering each tumor were manually delineated using 3D Slicer. From these regions, 107 radiomics features were extracted, and LASSO regression reduced them to 23 key features for model building.
To guard against data leakage, the researchers performed all preprocessing steps—feature standardization, LASSO feature selection, hyperparameter tuning, and threshold determination—exclusively within the training set, using a pipeline-based approach with 5-fold cross-validation. Feature selection stability was additionally assessed through 50 iterations of repeated stratified cross-validation. Two features were selected in all 50 iterations, one was retained in at least 98 percent, six reached a selection frequency of at least 80 percent, and eight reached at least 50 percent, while the remaining features showed considerable variability.
Four machine-learning models were constructed: logistic regression (LR), random forest (RF), XGBoost, and support vector machine (SVM). Performance was evaluated with ROC curves, precision-recall curves, confusion matrices, and decision curve analysis, and SHAP analysis was applied to the LR model for interpretability.
In the internal hold-out test set, the LR model achieved the highest AUC of 0.926 (95% CI: 0.860–0.978), followed by SVM at 0.915 (95% CI: 0.848–0.973), RF at 0.893 (95% CI: 0.814–0.960), and XGBoost at 0.889 (95% CI: 0.806–0.956). DeLong’s test found no statistically significant differences in AUC among the four models (all P > 0.05); the marginal difference between LR and RF (AUC difference = 0.033, P = 0.051) represented only a weak numerical gap without statistical superiority of LR.
The LR model showed balanced classification metrics, with accuracy 0.836, sensitivity 0.830, specificity 0.850, positive predictive value 0.929, negative predictive value 0.680, F1-score 0.876, and average precision 0.972. Notably, it yielded the lowest false-negative rate at 17.0 percent, which the authors note minimizes the clinical risk of undertreating advanced gastric cancer. Calibration testing revealed good fit for the random forest (P = 0.586), XGBoost (P = 0.081), and SVM (P = 0.213) models, whereas the logistic regression model showed evidence of miscalibration (P < 0.001). Despite this limitation, the authors argue that its strong discriminative performance supports its utility as a binary classifier. Decision curve analysis indicated that all four models provided clinical utility, with LR offering favorable net benefit across most threshold probabilities.
SHAP analysis confirmed that the contribution directions of selected features such as LeastAxisLength and SurfaceVolumeRatio were consistent with pathological interpretations, supporting the model’s interpretability.
The authors conclude that all four CT venous-phase radiomics models showed comparable discriminative power, and that the LR model’s higher clinical translation potential stems from its transparent coefficient structure, balanced classification metrics, and minimal false-negative risk rather than from statistically superior predictive efficacy. They position the model as a quantitative auxiliary tool for preoperative staging and treatment decision-making, while stressing that further multi-center external prospective validation is mandatory before formal clinical deployment.
The findings suggest that routine venous-phase CT images, already acquired during standard enhanced CT examinations, could be leveraged to provide an objective, quantitative adjunct to preoperative assessment of gastric cancer invasion depth, but clinical use will depend on the outcome of the planned external validation.
Subject of Research: Machine-learning radiomics models based on CT venous-phase images for preoperative differentiation of early and advanced gastric cancer
Article Title: CT venous-phase radiomics for preoperative differentiation of early and advanced gastric cancer: construction, comparison, and validation of machine-learning models
Article References: Zhang, H., Liu, Q., Lian, W., Huang, R., Zhao, X., Li, L., Zhang, L., & Zhang, G. (2026). CT venous-phase radiomics for preoperative differentiation of early and advanced gastric cancer: construction, comparison, and validation of machine-learning models. Cancer Imaging. https://doi.org/10.1186/s40644-026-01132-7
Image Credits: AI Generated
DOI: 10.1186/s40644-026-01132-7
Keywords: gastric cancer, radiomics, CT venous phase, machine learning, logistic regression, T staging, preoperative staging, LASSO feature selection, SHAP interpretability, retrospective study, invasion depth, diagnostic model
News Source: Ophelia Keating. (October 29, 2026). CT venous-phase radiomics models distinguish early from advanced gastric cancer before surgery. Scienmag.
