Our institutional review board approved this retrospective study, and the requirement for written informed consent was also waived. Initially, we performed a search of electronic medical records between January 2020 and July 2021 and identified a total of 1258 consecutive patients with surgically confirmed GC. Inclusion criteria were as follows: (a) underwent routine abdominal contrast-enhanced CT before surgical resection; (b) availability of IHC staining for analysis of MMR protein expression. Patients (a) who received neoadjuvant chemoradiotherapy prior to imaging; (b) with a history of other malignant tumors; (c) with distant metastasis during the operation; (d) with poor-quality images; (e) with invisible target lesions on CT images and (f) with diffuse or multiple GC lesions detected were excluded. The final study cohort comprised 53 dMMR patients and 106 MMR-proficient (pMMR) patients who were randomly selected ( Figure 1 ).

The postoperative pathological diagnosis and immunohistochemical staining were judged by two experienced pathologists. Tumor histological type, localization, ulceration, degree of differentiation, invasion depth and lymphovascular invasion were evaluated according to the 8th edition of AJCC cancer staging criteria of GC.

IHC for MMR protein (MLH1, MSH2, MSH6, and PMS2) expression was used to determine the MMR status and routinely performed by a standard streptavidin biotin-peroxidase procedure. Tumors displaying loss of at least one MMR protein were collectively referred to as dMMR, and those with intact expression were referred to as pMMR.

Abdominal contrast-enhanced CT examinations of patients were performed using a 64-channel CT (SOMATOM Definition Flash; Siemens Healthcare, Erlangen, Germany). Patients were instructed to fast for at least 6 hours prior to CT examination. Twenty minutes before scanning, patients were intravenously administered 10 mg of anisodamine to minimize gastrointestinal peristalsis and then drank 800-1000 ml water to expand the stomach. The CT scans were performed with the patient in the supine position and covered the entire abdomen from the diaphragmatic dome to the pubic symphysis. The images were acquired with the following parameters: a tube voltage 120 of kVp, a tube current of 200 mAs, detector collimation of 0.6 mm, an image matrix of 512 × 512, a slice thickness of 5 mm, a slice interval of 5 mm, and a pitch of 0.6 mm. After the unenhanced scan, 100 ml of iodinated contrast agent was administered at a rate of 3 ml/s into the antecubital vein with an automatic injector pump. The arterial and portal venous phases were obtained at 30-35 s and 65-70 s after injection of contrast material, respectively. The enhanced axial CT images were reconstructed with a section thickness of 1 mm for multiplanar reformation (MPR) reconstruction.

All the obtained images were evaluated retrospectively by two experienced abdominal radiologists (with 10 and 6 years of experience in abdominal radiology) who were blinded to the histopathological results. All variables included clinical data, morphological features, and CT quantitative parameters. Morphological feature analysis was independently performed by two radiologists, and the final results were determined by consensus, while the acquisition of quantitative parameters was jointly completed by both radiologists. The variables on CT were defined and measured as follows: tumor location (upper third of the stomach, middle third of the stomach and lower third of the stomach), tumor size and tumor thickness (the longest and thickest diameter of tumor on axial, sagittal, or coronal CT image), surface ulceration (absent or present), adjacent organ invasion (absent or present), growth pattern (localized or infiltrative type), and the largest lymph node size (short-axial diameters of the largest lymph node < 0.8 cm or ≥ 0.8 cm). A circular region of interest (ROI, a least 20 mm2 for large lesions and a circular with diameter slightly smaller than tumor thickness for small ones) was positioned to encompass as much of the most strongly enhanced portion of the tumor and the abdominal aorta as possible at the same slice of the portal venous phase by the two radiologists together. The average CT attenuation values of each tumor parenchyma (CTAVtumor) and abdominal aorta (CTAVaorto) were finally obtained to calculate the normalized tumor enhancement ratio (NTER) using the following formula: NTER (%) = CTAVtumor/CTAVaorto * 100. Additionally, other clinical data were recorded, including age, gender, hypertension and so on.

Continuous data distributions were verified using the Shapiro−Wilk test. Normally distributed data were presented as the mean ± standard deviation, and data with a nonnormal distribution were presented as medians and ranges (25th, 75th percentiles). Categorical variables were expressed as frequencies and percentages. In univariate analysis, differences between the two groups were compared using Student’s t test or the Mann−Whitney U test for continuous data and using the chi-square or Fisher’s exact test for categorical variables. Receiver operating characteristic (ROC) curve analysis was employed to evaluate the predictive value of each continuous parameter, which showed statistically significant differences and aided in determining their optimal cutoff value. Subsequently, the significant variables from the univariate analysis were entered into the multivariate binary logistic regression using the stepwise method. Finally, a nomogram was generated based on the identified independent predictors. The nomogram was internally validated by bootstrapping with 1000 resamples. To quantify the model performance in predicting MMR status, we assessed model discrimination using the AUC (area under the curve) and calibration using a calibration curve combined with the Hosmer−Lemeshow test. Additionally, the decision curve analysis (DCA) was utilized to determine the clinical practicability of the nomogram.

All data were processed in SPSS software (version 26.0.0; IBM Corp., Armonk, NY, USA) and R software (version 4.1.2; R Foundation for Statistical Computing, Vienna, Austria). A two-sided p value of < 0.05 indicated statistical significance.

Of the 1258 consecutive GCs, 97 (7.7%) patients had dMMR status. A total of 159 patients, comprising 53 with dMMR and 106 with pMMR, were enrolled in this study. The demographic and clinical characteristics of the patients are summarized in Table 1 . GCs with dMMR were more likely to be older, female and have a N0 stage. There were no significant differences with regard to hypertension, histological differentiation degree or T stage between groups.

Differences in CT features between dMMR and pMMR GC patients are presented in Table 2 . Significant differences were observed in location, surface ulceration, adjacent organ invasion, tumor size and NTER. The tumor size of the dMMR tumors was substantially larger, and the NTER was lower than that of pMMR tumors. Well-defined margins, thicker lesions and short-axial diameters of the largest lymph node ≥ 0.8 cm were found more often in dMMR patients, but the difference was not statistically significant (P = 0.094, P= 0.084, and P = 0.050, respectively).

Post-hoc comparison revealed that dMMR tumors were significantly more frequently found in the upper stomach than in the middle and lower regions, while there was no significant differences between tumors in the middle and lower stomach. Thus, the latter two subgroups were pooled into a single group for subsequent comparisons, hereafter named the “middle-lower” group. Then, we organized the continuous variables into meaningful and dichotomous categories according to the optimal cutoff value by ROC curve ( Table 3 ).

To select final predictors, eight candidate predictors with a P ≤ 0.05 in the univariate logistic analysis were included when performing the stepwise logistic regression analysis. Gender, age, tumor size and NTER emerged as significant independent predictors of the dMMR phenotype in GC patients ( Table 4 ). Based on the multivariate analysis results, a predictive nomogram incorporating the above four factors was then constructed ( Figure 2 ).

Using an optimal cutoff value of 6.6 points, the nomogram provided better AUC values than each risk factor for dMMR (AUC = 0.895) ( Figure 3 ), with a sensitivity, specificity, NPV, PPV and accuracy of 84.91%, 81.13%, 91.49%, 69.23% and 82.39%, respectively. To further evaluate the predictive ability of this model without overfitting, the corrected AUC obtained by internal validation using the bootstrap method was 0.896 (95% CI 0.837-0.950). The calibration curve demonstrated good consistency between the predicted risk and observed dMMR probability ( Figure 4 ), and the Hosmer−Lemeshow test also indicated that there was no departure from a perfect fit (P = 0.945). Furthermore, the DCA justified relatively good performance for the nomogram model according to clinical application ( Figure 5 ).