Comparison of CT-based radiomics models for the prediction of human epidermal growth factor receptor 2 status in gastric cancer.

Introduction
According to Global Cancer Statistics 2020, gastric cancer (GC) is one of the leading causes of cancer-related deaths, especially in eastern Asia [1]. In the past 10 years, breakthroughs in the basic biology of cancer have led to a new understanding of tumor targets, and researchers have developed many highly specific molecular targeted drugs, broadening the clinical diagnosis and treatment of cancer [2]. Clinically, human epidermal growth factor receptor 2 (HER2) has been established as a predictive biomarker for targeted therapy [3, 4]. In GC, it is a predictive biomarker for trastuzumab therapy. Because trastuzumab combined with chemotherapy significantly improves survival in patients with advanced HER2-positive gastric and gastroesophageal junction adenocarcinoma, combination therapy has become the standard treatment for HER2-positive advanced disease [5, 6]. Therefore, timely and accurate detection of HER2 status is essential for the treatment of GC.
In clinical work, HER2 status is mainly assessed by immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH), which is an invasive method involving tissue samples, and this technique is costly and time-consuming [7]. Consequently, HER2 status testing or follow-up evaluation is not routine during the treatment of patients with GC. Although certain studies have explored the potential relationship between HER2 status and noninvasive imaging tools, including positron emission tomography/computed tomography (PET/CT) and magnetic resonance imaging (MRI), no firm conclusions and reliable models have been established to date [8, 9]. In addition, PET/CT is not a routine test for all patients, and gastric gas affects MRI image quality. Therefore, new methods to assess HER2 status in GC are urgently needed.
Computed tomography (CT) is the preferred and conventional imaging method for the diagnosis, staging, curative effect evaluation and follow-up observation of GC recommended by the diagnosis and treatment guidelines or norms [10]. However, because the naked eye can recognize a limited number of features from the thousands of pixels in a CT image, the information that can be obtained is limited. In recent years, artificial intelligence (AI) technology has offered new ways to process images and turn them into quantitative data, identifying the microscopic features of tumors invisible to the naked eye [11]. Increasing evidence indicates that CT-based radiomics applied to various aspects of diagnosis, metastasis risk prediction, survival, and treatment response in patients with GC is promising [12–14]. Meanwhile, previous studies have also reported that HER2 status in GC can be predicted by contrast-enhanced CT-based radiomics features [15, 16].
While previous studies have demonstrated the feasibility of predicting HER2 status using radiomics, their reliance on an arbitrary or limited selection of classifiers poses a significant limitation. A critical gap remains: the lack of a systematic comparison across diverse machine learning families. This gap includes linear models such as logistic regression (LR), support vector machines (SVM), and powerful ensemble methods like gradient boosting decision tree (GBDT), random forest (RF), and extreme gradient boosting (XGBOOST). Conducting such a comparison is crucial to identify the most robust and generalizable model for this specific clinical task, as different classifiers have distinct inductive biases and sensitivities to dataset characteristics [17].
Therefore, this study aimed to use five machine learning classifiers to construct a contrast-enhanced CT-based radiomics model to predict HER2 status in GC patients. In addition, two other predictive models, including the clinical model and combined model were developed and validated, and the predictive performance of these models was systematically compared.

Materials and methods

Patient cohort
This retrospective study was approved by the local Institutional Review Board (IRB) (Jiangxi Provincial People’s Hospital), and informed written consent was waived by the IRB. All methods were performed in accordance with the relevant guidelines and regulations.
A total of 210 patients with GC diagnosed in our hospital between 2017 and 2023 were included in the study and enrolled into a training set (n = 147; HER2-negative: 111, HER2-positive: 36) and a testing set (n = 63; HER2-negative: 47, HER2-positive: 16) in a 7:3 ratio using a computer-generated random series. The inclusion criteria were as follows: (I) abdominal contrast-enhanced CT examination within one month prior to gastrectomy or endoscopic biopsy; (II) pathologically confirmed GC; (III) available HER2 status; and (IV) no radiotherapy or chemotherapy was performed prior to pathological examination. The exclusion criteria were as follows: (I) missing imaging data. (II) poor imaging quality: (i) respiratory artifacts or severe peristaltic in the image. (ii) the gastric cavity is not distended, making it difficult to delineate the ROI.

HER2 status assessment
HER2 status was assessed according to NCCN guidelines [18]. In brief, the HER2 status was detected by IHC and FISH. An IHC staining score of 3 + or 2 + with FISH + was considered positive. An IHC staining score of 0, 1+, or 2 + with FISH − was considered negative.

Clinical research materials
The following clinical features of each patient were collected by two radiologists (reader 1, P. H.; reader 2, W. D.) with 5 and 10 years of abdominal imaging experience, including age, gender, tumor location, tumor size, histological grade, pathological T stage, pathological N stage, AFP, CEA, CA199, and CA125. Blinded to the HER2 status data, the two radiologists interpreted the following features by consensus: tumor location (the main body of tumor is located gastric of fundus, body, antrum, respectively); tumor size (the maximum diameter of the tumor on the axial CT image); histological grade (low, middle, high; extracted from patients’ pathological data); and pathological T and N stage were determined in accordance with the eighth edition AJCC cancer staging manual [19]. Laboratory analysis of AFP, CEA, CA199, and CA125 was performed through routine blood tests within one week before gastrectomy or endoscopic biopsy. We used the median level of these laboratory analysis date as the cut-off value.

Clinical model construction
Univariate and multivariate logistic regression analyses of clinical features were performed to identify independent predictors of HER2 status in the training set. Then, these independent factors were used to establish the clinical model (clinical HER2 status prediction nomogram), and the clinical score was driven by Eq. (1).

CT Protocol
All patients were required to fast for at least 6 h and were instructed to drink 600–1000 mL of water before the contrast-enhanced CT examination. The CT scan covered the whole stomach region. Details of the CT imaging protocols are available in Table 1.

ROI segmentation
The DARWIN scientific research platform (Beijing Yizhun Intelligent Technology Co., LTD., https://arxiv.org/abs/2009.00908 ) was utilized to segment 3D-ROIs on the uploaded contrast‑enhanced CT of venous phase images. The ROIs were meticulously manually delineated on all slices by a radiologist who was blinded to the HER2 status (reader 1, P. H.). The volume of interest (VOI) of the lesions was automatically generated by the same platform. Meanwhile, another senior radiologist (reader 2, W. D.) reviewed all segmentation results. The intraclass correlation coefficient (ICC) was used to evaluate the intra- and interobserver consistency and reproducibility of radiomics feature extraction. We randomly selected 40 CT images (from the whole study cohort) for ROI re-segmentation by the senior radiologist (reader 2, W. D.). Features with ICC > 0.75 in both intra- and interobserver consistency analyses indicate good agreement of the feature extraction.

Radiomics feature extraction and selection
The Python PyRadiomics package (https://pyradiomics.readthedocs.io/en/latest/) was used to extract radiomics features. All feature extraction and processing steps were performed in accordance with the Imaging Biomarker Standardization Initiative (IBSI) guidelines. To minimize the feature variability introduced by the two different CT scanners, all images underwent a standardized preprocessing pipeline before feature extraction, which included image resampling to a uniform isotropic voxel size of 1.0 × 1.0 × 1.0 mm³.
Feature extraction was subsequently performed using the following key parameters: gray-level discretization with a fixed bin width of 25 Hounsfield Units to standardize the intensity scale, and a voxel array shift of -300 to ensure consistent feature calculation. A total of 1781 features were extracted per VOI, comprising 18 first-order statistics, 14 shape-based (3D) features, and 1749 texture features.
The computer-generated dataset was randomly assigned 70% of the data to the training set (HER2-negative: 111, HER2-positive: 36) and 30% to the testing set (HER2-negative: 47, HER2-positive: 16) in the same platform.
To make the algorithm converge faster and obtain more reasonable models, the maximum and minimum normalization was used to linearly stretch the features of each dimension to the interval [0,1]. Feature selection is crucial for training classifiers, reducing computational complexity and improving classification accuracy, so the optimal feature filter (sample variance F-value; https://scikit-learn.org/stable/modules/generated/sklearn.feature_selection.SelectKBest.html ) was used to evaluate the linear correlation between each feature and category label, and the 60 most relevant features were selected. The parameter k = 60 was empirically selected to retain a sufficiently large subset of potentially informative features for the subsequent least absolute shrinkage and selection operator (LASSO) regression, balancing computational efficiency with the need to avoid premature filtering of predictive features.
The LASSO regression algorithm was used to further select the optimal prediction features from the above features. The optimal penalty parameter (λ) for the LASSO regression was determined through 10-fold cross-validation on the training set, selecting the value of λ that yielded the minimum binomial deviance. The performance of the feature set was evaluated using nested cross-validation to prevent overfitting and data leakage during the feature selection process. Specifically, the LASSO regression with 10-fold cross-validation was applied to identify the most predictive features from the pre-selected pool. Finally, we extracted the 12 features that were most strongly associated with predicting HER2 status in GC, including 6 first-order statistical features and 6 texture features (Fig. 1).

Radiomics model construction
Based on the optimal radiomics features, five machine learning classifiers, including LR, SVM, RF, GBDT, and XGBOOST, were used to construct the radiomics model. The key hyperparameters for each classifier were set as follows to optimize performance and ensure reproducibility:
LR: L2 penalty was applied. The inverse regularization strength (C) was tuned, with the optimal value set to 1.0. The solver was ‘lbfgs’.
SVM: A radial basis function (RBF) kernel was used. The regularization parameter (C) and kernel coefficient (gamma) were tuned. The optimal values selected were C = 10 and gamma=’scale’.
RF: The number of trees (n_estimators) was set to 100. The maximum tree depth (max_depth) was tuned and set to ‘None’. The minimum samples to split a node (min_samples_split) was 2.
GBDT: The number of boosting stages (n_estimators) was 100. The learning rate and max_depth were tuned to 0.1 and 3, respectively.
XGBOOST: The number of trees (n_estimators) was 100. The learning rate (eta) and max_depth were tuned to 0.1 and 6, respectively. The objective was ‘binary: logistic’.
To validate the robustness and stability of each model, 10-fold cross-validation was utilized exclusively on the training set. The predictive performance of the five models was compared, the one with best performance was retained, and the radiomics score (Rad-score) was obtained according to Eq. (1).

Combined model construction
We developed a combined model (clinical-radiomics nomogram) that incorporated the independent HER2 status prediction factors and Rad-score via LR, which determined the prediction score with Eq. (1):

Note
represents the constant, stands for the LR coefficients, and stands for the value of chosen clinical features or Rad-score.

Model performance evaluation
To evaluate the predictive performance of the three models (clinical, radiomics, and combined models) in the training and testing sets, we quantified the area under the curve (AUC) of the receiver operating characteristic (ROC). DeLong ‘s test was used to compare the differences in AUC between the models. The calibration of the models was evaluated by calibration curves, and decision curve analysis (DCA) was used to measure the clinical net benefit of the models, reflecting the clinical practicability of the models. The workflow of this study is shown in Fig. 2.

Statistical analysis
Statistical analysis and graphical representations were conducted using SPSS 24.0 software and R software (Version 4.1.0). Pearson’s chi-square test was utilized in the comparison of categorical variables, reported as numbers and percentages. Student’s t-test was used for the analysis of normally distributed continuous variables, reported as the mean ± SD. LASSO regression was performed using the “glmnet (R)” package. The “partial Receiver Operating Characteristic (pROC)” package was used to plot ROC curves. The nomogram and calibration curves were constructed using the “Regression Modeling Strategy (RMS)” package. DCA was performed using the “risk model decision analysis (RMDA) " software package. The significance was set as two-sided P < 0.05.

Results

Clinical-pathological characteristics
In this study, no statistically significant difference was observed between the training and testing sets in terms of age, gender, tumor location, tumor size, histological grade, AFP, CA199, and CA125 (all P > 0.05). In addition, significant statistical differences were exhibited in pathological T stage (P = 0.044), pathological N stage (P = 0.005), and CEA (P = 0.009) in the training set but not in the testing set (all P > 0.05). More details are shown in Table 2.

Construction of the three HER2 status prediction models
After conducting univariable and multivariable LR analyses in the training set (Table 3), pathological T stage (OR, 3.86; 95% CI, 1.03–14.46; p = 0.045), pathological N stage (OR, 3.32; 95% CI, 1.05–10.49; p = 0.041), and CEA (OR, 2.86; 95% CI, 1.22–6.68; p = 0.015) were considered independent predictive factors of HER2 status in GC. Subsequently, the clinical model was constructed based on the predictive factors (Fig. 3a).
The predictive performance of the five machine learning models is presented in Fig. 4. Comparing the AUCs of the five models between the training and testing sets revealed that the predictive performance of the GBDT, RF, and XGBOOST classifiers in the training set was falsely high (all AUC = 1.000). However, the SVM classifier not only had good predictive performance in the training set (mean 10-fold cross-validation AUC = 0.869) but also had the best predictive performance (AUC = 0.823; 95% CI, 0.702–0.945) compared with the remaining classifiers in the testing set. Therefore, the SVM classifier was selected to construct the radiomics model. Figure 5 shows the distribution of the Rad-score in the training and testing sets, in which HER2-positive patients had a higher Rad-score than HER2-negative patients. The Rad-score was calculated using the following formula:
Similarly, the combined model, known as the clinical-radiomics HER2 status prediction nomogram, was constructed using LR analysis and by combining the Rad-score with the clinical score (Fig. 3b).

Validation and comparison of the three HER2 status prediction models
The combined model had the highest AUC value in the training set (AUC = 0.889; 95% CI, 0.826–0.951) and testing set (AUC = 0.826; 95% CI, 0.708–0.944), which clearly exceeded the clinical model and radiomics model (exhibited in Fig. 6). The results of DeLong’s test indicated statistically significant differences between the combined model and the clinical model in both the training set (p = 0.003) and testing set (p = 0.047), and the radiomics model and the clinical model also had significant differences in the training set (p = 0.015) (Table 4). The calibration curves showed good agreement among the predicted and actual values for all three HER2 status prediction models in both sets (Fig. 7a, b). The Hosmer-Lemeshow test yielded non-significant p-values (all p > 0.05), indicating no statistically significant deviation from perfect calibration and affirming the models’ reliability. The DCA revealed that the combined model had a higher overall net benefit in predicting HER2 status in GC than the clinical model and the radiomics model across most of the range of reasonable threshold probabilities (Fig. 7c, d). To provide a specific decision threshold for clinical use, the optimal cut-off probability for HER2 positivity was determined on the testing set using the Youden index. The optimal cut-off was 0.28, yielding a sensitivity of 81.3% and a specificity of 76.6%.

Discussion
The prediction of HER2 status in patients with GC is of great significance for clinical treatment and prognosis [5, 6]. In this study, we developed and validated three HER2 status prediction models based on optimal clinical and radiomics features, including a clinical model, a radiomics model, and a combined model. In general, the combined model showed the highest predicted efficacy and clinical net benefit of the three models, while the clinical model lagged.
Clinical features have been widely used to construct nomogram models to predict the gene expression of various tumors before surgery. Gastrointestinal tumors are no exception [20, 21]. In our study, a clinical model for predicting HER2 status in GC was constructed based on pathological T stage, pathological N stage, and CEA. Several past studies have noted that CEA is significantly correlated with HER2 expression in GC [15, 22, 23]. In addition, Park et al. [23] reported that the median level of serum CEA was correlated with HER2 status, which was consistent with our findings. Moreover, Zhao H et al. [15] found that clinical TNM staging (cTNM; I + II vs. III + IV) was a significant predictor of the HER2 status of GC. Similarly, the pathological T and N stages extracted from the pathological data of patients in our study were also demonstrated to be independent predictors of HER2 status. The above results indicated that the independent predictors used to construct the clinical model for predicting HER2 status in GC are effective and reliable. Although pathological T and N stages are typically obtained postoperatively and thus limit the clinical model’s utility for preoperative prediction, their inclusion served to establish a foundational understanding of the relationship between tumor pathology and HER2 status. This provides a benchmark for evaluating the superior, and truly non-invasive, predictive power offered by the CT radiomics signature. Furthermore, it highlights a clear pathway for clinical translation: future work should focus on developing a fully pre-operative model by integrating the radiomics signature with pre-therapeutic clinical staging (cTNM) and serum markers.
This study showed that the clinical model had AUCs of 0.730 and 0.638 in the training and testing sets, respectively, indicating its suitability for predicting HER2 status in patients with GC. In addition, the radiomics model presented better predictive efficacy in both the training set (AUC = 0.869 vs. 0.730) and the testing set (AUC = 0.823 vs. 0.638), and these differences reached statistical significance in the training set (DeLong’s test; P = 0.015). Similarly, compared with the combined model, AUC differences were statistically significant in both the training set (0.730 vs. 0.889, p = 0.003) and the testing set (0.638 vs. 0.826, p = 0.047). These findings proved that integrating radiomics into the clinical model led to improved predictive performance of the model. However, the difference in AUC values between the combined model (AUC = 0.826) and the radiomics model (AUC = 0.823) was extremely small, and DeLong’s test showed no statistical significance between the two (p = 0.890). Although the combined model did not show a statistically significant improvement over the radiomics model alone in the testing set, it achieved the highest numerical AUC. This suggests that the integration of clinical factors provides a stable, albeit minor, complementary effect to the radiomics signature. The model’s predictive accuracy and net benefit were also significantly improved with the inclusion of radiomics, further confirming that radiomics can increase accuracy and overall efficacy when combined with the clinical model.
This data-rich diagnostic investigation identified important associations between image-based features and HER2 status. Adding CT-based radiomics features to independent clinical predictors resulted in a combined model with better predictive power. The explanation proposed by Lambin P et al. [24] is that radiomics can capture and quantify intratumor heterogeneity in medical images, where information related to HER2 status is translated into quantitative features and further integrated into imaging phenotypes. Our findings align with the growing body of evidence linking radiomics phenotypes to underlying molecular characteristics in various cancers, including glioblastoma [25, 26], suggesting that the captured image heterogeneity may reflect the complex tumor biology associated with HER2 status.
Among the radiomics features included in our study, in addition to the first-order statistical features, we also used the platform to mine for richer high-order texture parameters inside the lesions, in which the neighboring gray tone difference matrix (NGTDM) accounted for the most significant proportion. NGTDM quantifies the sum of the differences between the gray level of a pixel or voxel and the average gray level of its adjacent pixels or voxels within a predefined distance, which describes the dynamic range of intensity at the local level, and better quantifies intratumor heterogeneity [27]. The prominence of NGTDM features in our model may offer a biological link to HER2 status. NGTDM quantifies the local intensity variation within the tumor, reflecting textural heterogeneity. In the context of HER2-positive GC, which is often associated with aggressive tumor behavior and specific histological patterns (e.g., intestinal-type, high-grade tumors) [7], this captured heterogeneity might correspond to underlying pathological phenomena. These could include variations in glandular structure, the coexistence of different cell populations, or regions of necrosis, all of which contribute to the macroscopic texture observed on CT imaging [28]. Thus, the NGTDM features may be serving as non-invasive surrogates for the complex intratumoral heterogeneity driven by HER2-driven molecular pathways [24, 28].
To ensure the robustness of our radiomics signature, we systematically evaluated five distinct machine learning classifiers. The five classifiers were selected to represent a spectrum of popular and high-performing algorithms in radiomics research: LR as a simple baseline [29], SVM for its effectiveness in high-dimensional spaces [30], and GBDT, RF, and XGBOOST as state-of-the-art ensemble methods known for capturing complex nonlinear relationships [31]. The performance comparison among classifiers revealed a critical finding: while complex ensemble tree-based methods (GBDT, RF, and XGBOOST) achieved perfect discrimination (AUC = 1.000) on the training set, their performance decreased notably in the testing set. This pattern is a classic indicator of overfitting, a known risk with such high-capacity algorithms, particularly with limited sample sizes [32]. In contrast, the SVM classifier demonstrated more balanced and robust generalization performance, yielding the best predictive efficacy (AUC = 0.823; 95% CI, 0.702–0.945) among the five classifiers in the testing set. This robustness of SVM can be attributed to its inherent maximum-margin principle, which helps prevent overfitting by seeking a decision boundary that maximizes the separation between classes, rather than perfectly fitting the training data [33, 34]. It works effectively even when the data dimension is larger than the number of samples, so it has a large advantage over many other classifiers in such scenarios [34, 35].
Looking forward, this study lays the groundwork for the future development of a fully pre-operative prediction tool. While our current model incorporates post-operative pathological staging to establish a robust benchmark for the radiomics signature, its primary value lies in validating the strong predictive power of CT-based radiomics for HER2 status. The clear next step is to translate this knowledge into a clinically applicable, pre-operative model. In a future clinical scenario, following a GC diagnosis and staging CT scan, the radiomics signature could be automatically extracted and integrated with genuinely pre-operative data (e.g., clinical TNM stage [cTNM] and serum CEA level) to generate a HER2 positivity probability. Such a tool would help clinicians identify high-priority candidates for confirmatory IHC/FISH testing, thereby optimizing resource allocation and potentially accelerating treatment decisions.
However, the translation of this vision into clinical practice must be tempered by an acknowledgment of the present study’s limitations. First, the model was developed and validated on a retrospective, single-center dataset with a relatively small sample size. This design, coupled with the lack of independent external validation, may introduce selection bias and limits the generalizability and stability of our findings, restricting their immediate clinical applicability. Therefore, external validation using multi-center datasets with larger sample sizes is imperative to verify the robustness and clinical utility of our models across diverse populations and healthcare settings. Second, the imbalance between HER2-positive and HER2-negative cases in our cohort, while reflective of clinical reality, is acknowledged. However, the use of AUC for evaluation and the robustness of the selected SVM classifier provide confidence in our reported performance. Third, manual delineation of the lesion may lead to errors and the loss of image information. Therefore, more accurate lesion contour delineation methods, such as semiautomatic segmentation, are needed in the future to extract more reliable radiomics features.

Conclusion
In this study, we constructed and validated three models for predicting HER2 status in GC. The combined model, which incorporated contrast-enhanced CT-based Rad-score with clinical factors, demonstrated the highest predictive efficacy and clinical net benefit among the models evaluated, thereby validating the significant value of radiomics and establishing a foundation for future development of a fully pre-operative prediction tool.