Preoperative prediction of HER2 expression in gastric cancer based on an enhanced CT radiomics model

Introduction

Gastric cancer remains one of the most common tumors worldwide, particularly in East Asia. The latest global cancer statistics report indicates that gastric cancer ranks fifth in incidence and fourth in mortality in China (1). The integrated treatment approach combining surgical resection with adjuvant chemotherapy and chemoradiotherapy remains the main treatment for advanced cancer.

However, the prognosis remains poor after treatment for patients with advanced cancer (2). Studies indicate that among all gastric cancer patients, 6% to 35% of gastric lesions overexpress human epidermal growth factor receptor 2 (HER2). This serves as both a key driver of tumorigenesis and a prognostic factor (3-5). Trastuzumab is a monoclonal antibody targeting HER2 that can inhibit tumor cell proliferation by blocking downstream signaling pathways. The median survival time for gastric cancer patients receiving only chemotherapy is less than one year (6). The Trastuzumab for Gastric Cancer (ToGA) trial showed that targeted therapy with trastuzumab combined with standard chemotherapy can extend the survival of HER2-positive patients (4). Currently, the assessment of HER2 status in gastric lesions via immunohistochemistry (IHC) or fluorescence in situ hybridization (FISH) is invasive. Gastroscopic biopsy struggles to reach the outer gastric wall, making it impractical for patients to undergo periodic follow-up evaluations of HER2 status during treatment (3,7). Therefore, accurate assessment of HER2 status is particularly significant for gastric cancer patients, and the adoption of new non-invasive methods to evaluate HER2 status is crucial.

This study aims to investigate the predictive value of radiomics derived from portal-phase-enhanced computed tomography (CT) for HER2 status in gastric cancer. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0034/rc).

Methods

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethical Committee of Cheng Fifth People’s Hospital (No. 20210306). Given the retrospective nature of the study, the requirement for written informed consent was waived by the ethical committee.

Patients

This study retrospectively enrolled patients diagnosed with gastric cancer who were consecutively admitted to Chengdu Fifth People’s Hospital from April 2022 to December 2023 and confirmed by postoperative pathological findings. After screening according to the inclusion and exclusion criteria, 100 gastric cancer patients were enrolled in this study. Inclusion criteria: (I) abdominal CT scans and contrast-enhanced imaging performed within 2 weeks before gastrectomy; (II) no preoperative radiotherapy or chemotherapy; (III) HER2 testing conducted via immunohistochemistry (IHC). Exclusion criteria: (I) gastric cavity insufficiently distended, inadequate contrast enhancement on CT images, respiratory or serve peristaltic artifacts; (II) lesions 5 mm that are unmeasurable or uncountable; (III) no serum tumor marker testing performed within one week before surgery; (IV) presence of other malignant tumors.

All patients were randomly split into a training cohort (n=70) and a validation cohort (n=30) at a ratio of 7:3. The training cohort was used to build the model, while the validation cohort was used to validate it. Collect patients’ clinicopathologic characteristics, including age, gender, preoperative carcinoembryonic antigen (CEA) level, tumor location, maximum tumor diameter, pathological information (differentiation, lymph node metastasis), CT-reported stage, and serosal invasion. The assessment of HER2 status follows the National Comprehensive Cancer Network (NCCN) guidelines (8), which deemed HER2 status as positive if the IHC score is 3+ or 2+, and negative if the score is 0 or 1+.

CT image acquisition protocol

The CT examination was performed using a 256-channel CT (Philips, Amsterdam, the Netherlands or GE Revolution Wisconsin, America) scanner. A lonic contrast material Ultravist (2.5 mL/s, 1.5 mL/kg, Bayer Pharmaceuticals) or iopamidol (3 mL/s, 1.5 mL/kg) was administered intravenously via the median cubital vein. Arterial phase scans were performed 30s post-injection, followed by portal venous phase scans at 60 s. The scan range extends from the diaphragm to the renal inferior pole, completed at the end of a quiet exhalation. Examination details and reconstruction parameters: tube current 280–600 mA or automatic mA adjustment technology, tube voltage 80–120 kV, matrix of 512×512, field of view 350 mm × 350 mm, pitch 0.9–1.0 mm, rotation time 0.8 s, window width 300–360 HU, window level 45–60 HU, reconstruction slice thickness 1.25–1.5 mm. All acquired image data were uploaded to the picture archiving and communication system (PACS) and the computer post-processing workstation.

Radiomics analytics

Since the peak time for tumors involving the muscularis layer (≥T2, advanced gastric cancer) on contrast-enhanced CT for gastric cancer generally occurs after the 60s to 70s, and often persists for a prolonged duration following mucosal enhancement, this study selected the portal venous phase as the target period for investigation (9). ITK-SNAP software (Version 3.6.0, http://www.itksnap.org) was used to construct the volume of interest (VOI) along the margin of tumor., avoiding surrounding blood vessels, adipose tissue, fluid, or gas. The segmentation process is illustrated in Figure 1. All VOIs were delineated by two radiologists: one with 6 years of experience in diagnosing gastrointestinal tumors and another with 13 years of abdominal imaging experience. To validate the stability and robustness of radiomic features, the intraclass correlation coefficient (ICC) was used to assess the inter-observer and intra-observer consistency of radiomic feature extraction (10). Using the Pyradiomics package on Python (version 3.9), Nonlinear intensity transformations were applied to image voxels (squared, square root, logarithmic, and exponential), Gaussian Laplace filters with sigma values of 1, 2, and 3, and eight wavelet transform algorithms were performed on FOF and texture features. To maximize the ability to retain descriptive features, remove the feature with the highest redundancy in the current set. Most features comply with the Image Biomarker Standardization Initiative (11). Perform z-score normalization on all features, conduct Mann-Whitney U tests, and feature selection on all radiomics features, retaining only radiomics features with P values less than 0.05. For highly repetitive features, calculate the Spearman correlation coefficient between each pair of features. Feature pairs with coefficients greater than 0.9 were deemed highly correlated. Using the least absolute shrinkage and selection operator (LASSO) with 10-fold cross-validation for feature selection.

Figure 1 An example of manual segmentation in gastric cancer.

Construction of radiomics nomogram

The filtered radiological features were processed through six machine learning algorithms, including logistic regression (LR), Naive Bayes (NB), support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), and extreme gradient boosting (XGBoost).

Predictive performance of radiomics nomogram

Based on the training cohort to construct a radiomics model and select the optimal radiomics model. Clinical independent predictors were identified through univariate logistic analysis and multivariate logistic analyses to establish a clinical prediction model. Clinical independent predictors were combined with radiomics features to establish a clinical-radiomics model, which was then tested on the validation cohort.

Statistical analysis

Statistical analysis was performed using IBM SPSS software (version 26.0; 0; IBM Corp, New York, NY, USA). Quantitative data, including age and maximum tumor diameter, were converted to categorical data and expressed as percentages. We used the chi-square test or Fisher’s exact test to analyze the count data. Statistical significance was set at P<0.05. The goodness-of-fit of the regression model was assessed using the Hosmer-Lemeshow test and calibration curves. The receiver operating characteristic curve (ROC) was plotted, and the model’s clinical utility was assessed through the area under the curve (AUC) and decision curve analysis (DCA). The DeLong test was used to compare AUC differences between models.

Results

Clinical features

Single-factor analysis results revealed statistically significant differences between HER2-positive and HER2-negative patients in the training and validation cohorts regarding serosal invasion (P<0.001) and tumor location (P=0.048). However, there were no significant differences in CEA levels (P=0.37), age (P=0.97), gender (P=0.91), tumor maximum diameter (P=0.59), differentiation (P=0.11), lymph node metastasis (P=0.37), and depth of invasion (P=0.59). The summaries of cohorts are shown in Table 1. The variable assignments are shown in Table 2, and the results of the multivariate analysis of factors affecting HER2 expression status in gastric cancer are shown in Table 3.

Feature extraction

Of the 1,834 radiomics features extracted from each VOI completed by segmentation, including 360 first-order features (FOF), 440 gray level co-occurrence matrices (GLCM), 280 gray level dependence matrices (GLDM), 320 gray level run length matrices (GLRLM), 320 gray level size zone matrices (GLSZM), 100 neighboring gray-tone difference matrix (NGTDM), and 14 shape features. 44 radiomics features with an ICC >0.75 were selected (Figure 2A). After excluding highly correlated features, a total of 14 radiomics features were retained (Figure 2B). By using the LASSO logistic regression method with ten-fold cross-validation, 8 radiomics features were retained (Figure 2C), the selected features were shown in Table 4.

Figure 2 Selection of radiomics characteristics from a portal venous phase CT image. (A) Tuning parameter (λ) in the LASSO using 10-fold cross-validation via minimum criteria. (B) Feature analysis profile of LASSO dimension reduction. (C) Histogram of feature weighting in radiomics. CT, computed tomography; LASSO, least absolute shrinkage and selection operator; MSE, mean squared error.

Model construction and predictive performance evaluation

We selected six machine learning algorithms, including LR, NB, SVM, KNN, RF, and XGBoost, to construct models with the final retained radiomics features (Table 5). The most effective machine learning algorithm was selected to build the radiomics model, with results indicating that the LR model demonstrated the highest predictive performance (Figure 3). By integrating clinical independent risk factors (serosal invasion), we constructed a clinical model and a clinical-radiomics model. The training and validation cohorts of the clinical model achieved AUC values of 0.932 and 0.944, respectively, with accuracies of 82.9% and 90.0%. The AUC values for the training and validation cohorts of the radiomics model were 0.823 and 0.790, respectively. Sensitivity and specificity were 66.7% vs. 66.7%, and 81.0%, and 81.0% respectively. The clinical-radiomics model achieved AUC values of 0.986 and 0.938 on the training and validation cohorts, with sensitivities of 91.7% and 66.7%, specificities of 93.1% and 92.6%, and accuracy of 92.9% and 90.0% (Figure 4).

Figure 3 ROC curve of logistic regression model. AUC, area under the curve; CI, confidence interval; LR, logistic regression; ROC, receiver operating characteristic.

Figure 4 Comparison of ROC curves between clinical models, radiomics models, and nomogram in training and validation cohorts. (A) The ROC curves in the training cohort. (B) The ROC curves in the validation cohort. AUC, area under the curve; CI, confidence interval; CT, computed tomography; ROC, receiver operating characteristic.

Using the DeLong test to compare the AUC values of the three models pairwise (Table 6), the combined model demonstrated superior predictive performance, with statistically significant differences in AUC compared to the clinical model and the radiomics model in the validation cohort (P=0.04, 0.03). The Hosmer-Lemeshow test was used to evaluate the diagnostic performance of the models. Results indicated that both the clinical model and the combined model demonstrated good fitness, with P values of 0.979 vs. 0.828 and 0.583 vs. 0.579 in the training and validation cohorts, respectively (Figure 5). The DCA curve indicates that the combined model demonstrates superior clinical utility (Figure 6). The combined model is visualized as a column chart (Figure 7).

Figure 5 Calibration curves of the training (A) and validation (B) cohorts generated by the three models. CT, computed tomography.

Figure 6 DCA curves of each prediction model in the training (A) and validation (B) cohorts. CT, computed tomography; DCA, decision curve analysis.

Figure 7 Clinical application of column chart in predicting the HER2 expression status. CT, computed tomography; HER2, human epidermal growth factor receptor 2.

Discussion

Research on HER2 expression had primarily focused on breast cancer; however, in recent years, it has been discovered that HER2 also plays a significant role in malignant tumors such as gastric cancer and colorectal cancer. HER2-positive expression correlates with lymph node metastasis, distant metastasis, and TNM staging in gastric cancer patients (12). Relevant literature reports that the overall survival rate of HER2-negative gastric cancer patients is significantly higher than that of HER2-positive gastric cancer patients (13).

The risk of postoperative recurrence and metastasis in HER2-positive gastric cancer patients is markedly higher than in HER2-negative gastric cancer patients. Chen et al. (14) found no association between serum CEA and HER2 overexpression, but Mavroeidis et al. (15) reported a correlation between serum CEA levels and HER2 status. Li et al. (16) retrospectively collected portal phase images from 134 gastric cancer patients. By integrating radiomics features with a prediction model constructed at the CEA levels. Results showed an AUC value of 0.799 [95% confidence interval (CI): 0.704–0.894] in the training cohort (95% CI: 0.704–0.894), with an AUC value of 0.771 (95% CI: 0.607–0.934). Given the inconsistent findings in the aforementioned studies, this research also examined the correlation between serum CEA levels and HER2 expression status. Results showed no statistically significant association between serum CEA levels and HER2 expression status (P=0.37). One contributing factor identified in the analysis was the limited number of HER2-positive samples included in this study, with an even smaller proportion of HER2 (3+) samples.

In this study, serosal invasion status in gastric cancer patients was assessed using the criteria of high-intensity serosal enhancement (9). Through univariate and multivariate analyses, the results demonstrated that serosal invasion (P<0.001) is an independent risk factor predicting gastric cancer prognosis, consistent with previous studies (17). The clinical model incorporating serosal invasion as an independent risk factor achieved AUC values of 0.932 and 0.944 in the training and validation cohorts, respectively, with accuracies of 82.9% and 90.0%. These results are consistent with those of previous studies (18,19).

Numerous studies have confirmed that the overall accuracy rate of CT scans in preoperative T staging for gastric cancer ranges from approximately 69% to 88% (20,21). However, CT stage in this study showed no correlation with gastric cancer prognosis. This may be because the T staging in this study was assessed by radiologists based on CT images, making it susceptible to observer subjectivity.

Currently, there are few reports on constructing line chart models for preoperative prediction of HER2 expression status in gastric cancer. Li et al. (16) established a scatter plot model demonstrating good discriminatory performance for HER2 expression status; however, this model was constructed solely using a logistic regression model. This study also incorporated images from gastric cancer patients in the portal vein phase, expanded the range of machine learning algorithms to identify the optimal radiomics model, and the constructed prediction model demonstrated superior performance.

There are certain limitations in this study. (I) This study investigates the predictive value of HER2 expression status, but serum tumor markers included only CEA, excluding other markers such as CA199, CA724, and CA125. (II) The retrospective, single-center nature of the samples included in this study may lead to model overfitting. Future prospective multicenter studies are needed to enhance the model’s accuracy. (III) The scanner models used in this study differed. Despite image preprocessing, some texture features may have been lost, and the preprocessing was insufficient to eliminate the effects of the center effect. (IV) In this study, due to the irregular morphology of the gastric wall, manual delineation of the VOI was employed. Manual delineation relies heavily on the radiologist’s experience, which can lead to errors. Future studies should strive to adopt semi-automated or fully automated delineation methods whenever possible. (V) This study included only portal phase images and did not incorporate arterial or venous phases into the analysis.

Conclusions

The radiomics model constructed based on radiomics features derived from portal phase-enhanced CT demonstrates favorable predictive performance for HER2 expression status in gastric cancer patients. The predictive model established by integrating clinically independent risk factors further enhances overall performance.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0034/rc

Data Sharing Statement: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0034/dss

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0034/prf

Funding: This study was supported by the Major Basic Research Project of Chengdu University of Traditional Chinese Medicine Science Foundation (No. YYZX2020016).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0034/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethical Committee of Cheng Fifth People’s Hospital (No. 20210306). Given the retrospective nature of the study, the requirement for written informed consent was waived by the ethical committee.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
2. Zhu ZG. Current status and prospect of perioperative therapy for locally advanced gastric cancer. Zhonghua Wei Chang Wai Ke Za Zhi 2021;24:101-6. [PubMed]
3. Shimozaki K, Fukuoka S, Ooki A, et al. HER2-low gastric cancer: is the subgroup targetable? ESMO Open 2024;9:103679. [Crossref] [PubMed]
4. Bang YJ, Van Cutsem E, Feyereislova A, et al. Trastuzumab in combination with chemotherapy versus chemotherapy alone for treatment of HER2-positive advanced gastric or gastro-oesophageal junction cancer (ToGA): a phase 3, open-label, randomised controlled trial. Lancet 2010;376:687-97. [Crossref] [PubMed]
5. Coit DG, Strong VE. Fifty years of progress in gastric cancer. J Surg Oncol 2022;126:865-71. [Crossref] [PubMed]
6. Janjigian YY, Kawazoe A, Bai Y, et al. Pembrolizumab plus trastuzumab and chemotherapy for HER2-positive gastric or gastro-oesophageal junction adenocarcinoma: interim analyses from the phase 3 KEYNOTE-811 randomised placebo-controlled trial. Lancet 2023;402:2197-208. [Crossref] [PubMed]
7. Hirai I, Tanese K, Nakamura Y, et al. Phase II clinical trial of docetaxel and trastuzumab for HER2-positive advanced extramammary Paget's disease (EMPD-HER2DOC). Oncologist 2024;29:e1201-8. [Crossref] [PubMed]
8. Ajani JA, D'Amico TA, Bentrem DJ, et al. Gastric Cancer, Version 2.2022, NCCN Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw 2022;20:167-92. [Crossref] [PubMed]
9. Kist JW, Prince JF, Lacle MM, et al. Letter to the Editor on "Radiologic serosal invasion sign as a new criterion of T4a gastric cancer on computed tomography: diagnostic performance and prognostic significance in patients with advanced gastric cancer". Abdom Radiol (NY) 2024;49:4185-6. [Crossref] [PubMed]
10. Hernández-Hermoso JA, Nescolarde L, Yañez-Siller F, et al. Combined femoral and tibial component total knee arthroplasty device rotation measurement is reliable and predicts clinical outcome. J Orthop Traumatol 2023;24:40. [Crossref] [PubMed]
11. Orlhac F, Eertink JJ, Cottereau AS, et al. A Guide to ComBat Harmonization of Imaging Biomarkers in Multicenter Studies. J Nucl Med 2022;63:172-9. [Crossref] [PubMed]
12. Cheng J, Cai M, Wang G, et al. HER2 becomes a novel survival biomarker for gastric cancer patients: a pooled analysis. Ther Adv Med Oncol 2024;16:17588359241271913. [Crossref] [PubMed]
13. Lv J, Yuan C, Zhao HY, et al. HER2 protein levels predict tumor-free survival and overall survival after radical gastrectomy for gastric cancer patients. Journal of Modern Oncology 2023;31:4567-71.
14. Chen XZ, Zhang WH, Chen HN, et al. Associations between serum CA724 and HER2 overexpression among stage II-III resectable gastric cancer patients: an observational study. Oncotarget 2016;7:23647-57. [Crossref] [PubMed]
15. Mavroeidis VK, Gkegkes ID, Saffioti F, et al. Parietal cell/oncocytic gastric carcinoma: systematic review and first-time assessment of HER2 status in two new cases. Ann R Coll Surg Engl 2020;102:300-7. [Crossref] [PubMed]
16. Li Y, Cheng Z, Gevaert O, et al. A CT-based radiomics nomogram for prediction of human epidermal growth factor receptor 2 status in patients with gastric cancer. Chin J Cancer Res 2020;32:62-71. [Crossref] [PubMed]
17. Nakamura K, Tajima K, Kanamori K, et al. Impact of Subclassification of Serosal Invasion on the Survival of Patients With T4a Gastric Cancer. In Vivo 2022;36:1923-9. [Crossref] [PubMed]
18. Wan C, Chen X, Yang Z, et al. Development and validation of models to predict serosa invasion in advanced gastric cancer using the enhanced CT imaging-based radiomics features and clinical feature. Chinese Journal of Medical Physics 2023;40:1518-22.
19. Tan HY, Xu J. The Predictive value of imageomics model based on enhanced CT portal venous phase images for serosal infiltration in advanced gastric adenocarcinoma. Medical Equipment 2023;36:17-20.
20. Wang K, Guo D, Yan H, et al. Enhancing Diagnostic Accuracy for Gastric Cancer: Integration of Multi-Slice Spiral CT and Gastrointestinal Angiography. Altern Ther Health Med 2024;30:118-23. [PubMed]
21. Bai L, Liu W, Di S, et al. Clinical study of CT enhanced scan in preoperative TNM staging of advanced gastric cancer and the effect of misdiagnosis rate. Panminerva Med 2023;65:259-60. [Crossref] [PubMed]