To evaluate the predictive value of CT radiomics features derived from the primary tumor in discriminating occult peritoneal metastasis (PM) in advanced gastric cancer (AGC).
Preoperative CT images of 233 patients with AGC were retrospectively analyzed. The region of interest (ROI) was manually drawn along the margin of the lesion on the largest slice of venous CT images, and a total of 539 quantified features were extracted automatically. The intra-class correlation coefficient (ICC) and the absolute correlation coefficient (ACC) were calculated for selecting influential features. A multivariate logistic regression model was constructed based on the training cohort, and the testing cohort validated the reliability of the model. Additionally, another model based on the preoperative clinic-pathological features was also developed. The comparison of the diagnostic performance between the two models was performed using ROC analysis and the Akaike information criterion (AIC) value.
Six radiomics features (ID_Energy, LoG(0.5)_Energy, Compactness2, Max Diameter, Orientation, and Surface Area Density) differed significantly between AGCs with and without PM and performed well in distinguishing AGCs with PM from those without PM in the primary cohort (AUC = 0.618-0.658). The radiomics model showed a higher AUC value than each single radiomics feature in the primary cohort (0.741 vs. 0.618-0.658) and similar diagnosis performance in the validation cohort. The radiomics model showed slightly worse diagnostic efficacy than the clinic-pathological model (AUC, 0.724 vs. 0.762).
Venous CT radiomics analysis based on the primary tumor provided valuable information for predicting occult PM in AGCs.
- Venous CT radiomics analysis provided valuable information for predicting occult peritoneal metastases in advanced gastric cancer.
- CT-based T stage was an independent predictive factor of occult peritoneal metastases in advanced gastric cancer.
- A radiomics model showed slightly worse diagnostic efficacy than a clinic-pathological model.