Peritumoral CT radiomics reveals CLDN18.2-related imaging phenotypes beyond the tumor boundary in gastric cancer.

BACKGROUND

CLDN18.2 is an actionable biomarker in gastric cancer and a tight-junction protein implicated in epithelial adhesion and barrier integrity. Alterations in CLDN18.2-related junctional biology may influence tumor invasion, epithelial remodeling, and tumor-host interaction at the invasive front. We hypothesized that CLDN18.2-associated biological alterations extend beyond the tumor boundary and may therefore be captured more effectively by peritumoral CT radiomics than by conventional intratumoral radiomics.

METHODS

This single-center retrospective study included 210 patients with pathologically confirmed gastric adenocarcinoma and available pre-treatment contrast-enhanced CT and CLDN18.2 immunohistochemistry results between January 2023 and December 2025. Patients were divided into a training cohort (n = 151) and a validation cohort (n = 59). CLDN18.2 expression was determined by immunohistochemistry using surgical resection or biopsy specimens as clinically available. For image analysis, 5-mm axial venous-phase reconstructed images were resampled to 1 × 1 × 1 mm3 isotropic voxels before slice-by-slice whole-volume 3D intratumoral and peritumoral ROI segmentation. Radiomics features were extracted using PyRadiomics, followed by variance filtering, Spearman redundancy filtering, and LASSO logistic regression. Intratumoral and peritumoral radiomics models were evaluated, and independent predictors were integrated into a clinicoradiomic nomogram as a clinical translation of the peritumoral imaging signature. Model assessment included ROC analysis, calibration analysis, decision curve analysis, and SHAP interpretation.

RESULTS

Among 210 patients, 106 were CLDN18.2-positive and 104 were CLDN18.2-negative. LASSO selection retained 4 intratumoral and 7 peritumoral radiomics features. The peritumoral model yielded AUCs of 0.761 in the training cohort and 0.784 in the validation cohort, compared with 0.756 and 0.761 for the intratumoral model, respectively. Peritumoral RadScore remained independently associated with CLDN18.2 positivity in multivariable analysis (OR, 3.23; 95% CI, 2.08-5.02; P < 0.001). A clinicoradiomic nomogram incorporating CA19-9, Borrmann type, and Peritumoral RadScore achieved AUCs of 0.818 in the training cohort and 0.820 in the validation cohort, with acceptable calibration and potential clinical utility. SHAP analysis indicated that peritumoral texture and intensity-distribution features contributed substantially to model prediction.

CONCLUSIONS

Peritumoral CT radiomics captured CLDN18.2-associated imaging phenotypes beyond the visible tumor boundary and showed numerically higher discrimination than conventional intratumoral radiomics. These findings suggest that the peritumoral region may provide an imaging window into features consistent with CLDN18.2-associated epithelial remodeling and tumor-host interaction. The clinicoradiomic nomogram provides a potential clinical translation of this peritumoral imaging phenotype for individualized prediction of CLDN18.2 positivity.