Interpretable intratumoral and peritumoral radiomics using SHAP for predicting microsatellite status in gastric adenocarcinoma: a dual-center study.

OBJECTIVE

The purpose of this study was to develop and validate a non-invasive radiomics-based model utilizing contrast-enhanced CT imaging of both intratumoral and peritumoral regions to predict preoperative microsatellite instability (MSI) status in gastric adenocarcinoma (GAC).

METHODS

A retrospective cohort comprising 193 patients with histologically confirmed GAC from two separate institutions (Centre 1: n = 115; Centre 2: n = 78) was enrolled. All patients underwent preoperative enhanced CT scans and immunohistochemical assays to determine MSI status. Tumor regions of interest (ROIs), specifically the intratumoral region (IR) and an extended area including the intratumoral and surrounding 3-mm peritumoral regions (IPR), were manually segmented on portal-phase CT images. Radiomics features were extracted from these defined ROIs. Following feature standardization, selection was conducted through inter-observer consistency (ICC), pairwise correlation analyses, and L1-regularized logistic regression. A radiomics signature was subsequently constructed via a support-vector machine (SVM) classifier. Feature contributions were quantified using SHapley Additive exPlanations (SHAP). Independent clinical variables and semantic features derived from CT were employed to develop a separate clinical model. A combined model was then formulated by integrating both radiomics and clinical variables. Receiver operating characteristic (ROC) curves, calibration plots, and decision curve analysis (DCA) were used to evaluate the predictive performance, calibration, and clinical utility of each model.

RESULTS

The combined radiomics-clinical model integrating intratumoral and peritumoral features exhibited superior predictive capability (AUC = 0.891) compared to the standalone clinical model (AUC = 0.771), peritumoral radiomics model (AUC = 0.780), and tumor-clinical integrated model (AUC = 0.784). Calibration curves and decision-curve analysis suggested favorable calibration and potential clinical net benefit of the integrated model.

CONCLUSION

The integrated model combining clinical variables with radiomic characteristics extracted from the intratumoral plus 3-mm peritumoral region showed promising ability for preoperative MSI prediction in GAC. Because this retrospective study included few MSI-H cases and used IHC as the reference standard, the model should be regarded as a preliminary imaging biomarker that requires prospective multicenter validation before clinical use.