Development and validation of an interpretable CT-based scoring model for gastric cancer aggressiveness.

objectivesAccurate prediction of adverse histopathological status (AHS) in gastric cancer (GC) is clinically crucial due to its strong association with poor prognosis. This study aims to develop and validate a CT-based machine learning model for AHS prediction, and establish an interpretable imaging score (I-score) for prognostic stratification.

methodsIn this dual-center retrospective study, 1164 GC patients undergoing radical gastrectomy between 2014 and 2023 were included. Four semantic CT features, including clinical lymph node status (cN), longest diameter (LD), tumor thickness (TT), and serosal status, were independently evaluated by radiologists. Radiomic features were extracted from arterial and portal venous phase volumes of interest. Logistic regression, support vector machine, random forest, and extreme gradient boosting (XGBoost) models were trained to predict AHS. The best-performing model was used to construct the I-score, which was validated in an independent cohort for prognostic assessment.

resultsAmong the 1164 patients (median age: 62 years old; 847 men, 317 women), data were divided into training (n = 396), test (n = 618), and validation (n = 150) cohorts. The XGBoost model using semantic features achieved the highest predictive performance (AUC: 0.803, 0.848 and 0.764, respectively). The derived I-score stratified patients into low-risk (43%) and high-risk (57%) groups, with significantly poorer 1000-day overall survival in the high-risk group (55.7% vs. 73.8%, p < 0.001). Multivariate Cox analysis confirmed the I-score as an independent prognostic factor (HR = 1.02, p = 0.004).

conclusionThe simplified CT-based machine learning model using semantic imaging features achieved high predictive performance for AHS in GC. The interpretable I-score enables effective preoperative risk stratification and individualized treatment planning.