In recent years, targeted therapy has been increasingly applied to cancer treatment. Many studies have shown that HER2 overexpression is associated with poor prognosis and low survival rate in patients with gastric cancer both before and after surgery, and such patients may benefit from HER2-targeted therapy[21, 22]. In addition, some scholars have confirmed that HER2 overexpression is not only related to the prognosis of patients with gastric cancer but is also closely related to the occurrence and development of gastric cancer [23]. Therefore, it can be seen that the expression level of HER2 is crucial in guiding the choice of individual treatment strategies in GC patients. However, not all patients are able to detect HER2 expression status due to the cost and equipment of IHC examination in daily clinical practice. Some lesions could not be sampled because of their location. In addition, the tumor heterogeneity of gastric cancer is relatively high, with inconsistencies in tissue, time, and space, which can be as high as 5–30% [24]. For example, different biopsy sites in the same lesion can lead to different HER2 expression states. The expression status of HER2 in primary and metastatic lesions is sometimes not identical. The HER2 status can be partially altered before and after treatment for gastric cancer. Moreover, the detection of HER2 by conventional IHC method is affected by different pathologists. Therefore, there is an urgent need to explore a new method that can accurately, standardizedly, noninvasively, and dynamically evaluate HER2 expression in patients with gastric cancer. 18F/68Ga-HER2 affibody PET/CT can detect HER2 expression in malignant tumors, however, it has not been widely applied in clinical practice to date.
In this study, we established and validated three predictive models of HER2 expression status in patients with GC based on 18F-FDG PET/CT radiomics features before treatment. It could be seen that among the three radiomics models, the LR and RF models had a better performance in predicting the expression status of HER2 in GC patients and also performed better in classification accuracy. After the data were balanced by SMOTE, the AUC of the three models increased in the training and test sets. All had good predictive performance, and the RF model showed the best performance. The results showed that the predictive models of HER2 expression status in GC patients based on 18F-FDG PET/CT radiomics features had relatively high sensitivity, specificity, and accuracy, and had potential value in clinical use.
Among many tumor markers, HER2 status is one of the most important factors in the clinical diagnosis and treatment of gastric cancer. At present, many radiomics studies on HER2 have made progress in breast cancer, esophageal cancer, and lung cancer [25]. Currently, the main applications of machine learning technology in tumor imaging include two categories: radiomics and convolutional neural networks [26]. convolutional neural network is prone to overfitting when making disease classification predictions when applied to a small sample size, resulting in poor generalization ability of the model [27]. However, in clinical practice, the sample size of complete medical image data is relatively small, and the use of convolutional neural network is limited to some extent [28]. Therefore, in this study, radiomics was used to construct the model. Reviewing relevant studies, we found that the sample size limit of radiomics is relatively low, and the performance of the model is better. Li et al. [29] combined a radiomics model based on CT image features with the level of Carcinoembryonic Antigen (CEA) and obtained the best model to predict HER2 expression in GC patients; the AUC values of the best model in the training and test sets were 0.799 and 0.771, respectively. Similarly, Wang et al. [30] conducted a retrospective analysis of 101 patients with adenocarcinoma of the esophagogastric junction and obtained 7 optimal features based on portal-phase CT images. The Nomogram, which consisted of the best features with T-staging from CT, also had very good predictive performance for HER2 status (AUC, 0.946 and 0.903 in the training and test sets, respectively). This is also similar to the predictive performance of the composite model constructed by Ma et al. [31] based on three stage CT images of 745 GC patients (AUC, 0.85 and 0.84 in the training and test sets, respectively). In addition, Wang et al. [32] retrospectively analyzed CT images of 132 GC patients and established a RF model to predict HER2 expression in GC patients. The model was built based on the radiomics features of arterial-phase CT, and the AUC values were 0.756 and 0.830 in the training and test sets, respectively. However, the above studies were mainly based on CT images, and PET images were not included in the study; therefore, the prediction model based on CT images has limited clinical application. In this study, three radiomics models based on 18F-FDG PET/CT imaging were constructed, and the results demonstrated that the three radiomics models had good predictive performance for HER2 status, especially the LR and RF models. The clinical variables are not included in the construction of the prediction model because the research results show that clinical variables such as gender and age do not have the ability to predict the expression of HER2.
Several studies have shown that 18F-FDG PET/CT plays an important role in predicting HER2 expression. Chen et al. [33] reviewed the 18F-FDG PET/CT images of 64 patients with gastric cancer before surgery. The results showed that the SUVmax of lesions in patients with HER2-positive gastric adenocarcinoma was significantly higher than that in HER2-negative lesions (SUVmax: 8.619 ± 5.878 vs 2.600 ± 2.036). This is similar to the results of Bai et al. [34], where the SUVmax was significantly different between HER2-positive and HER2-negative gastric cancer lesions. With the development of artificial intelligence, the applications of radiomics, machine learning, and deep learning in the medical field have become increasingly extensive and mature. At present, research on gastric cancer based on 18F-FDG PET/CT imaging has penetrated many aspects, such as the diagnosis and differential diagnosis of gastric cancer, staging, lymph node metastasis, and prognosis. However, no studies have reported on the prediction of HER2 expression in gastric cancer. In this study, PET imaging features were included to predict HER2 expression in the patients with gastric cancer. In addition, it can be seen from the study that for different sample proportions, the three machine learning models have their own advantages.
Radiomics features are usually divided into three categories: shape, first-order, and second-order features. Among the six best features selected in this study, SUVbwKurtosis and SUVbwExcessKurtosis are radiomics features based on PET images, reflecting cell metabolic activities. Kurtosis, also known as the peak coefficient, represents the peak of the probability density distribution curve at the mean value. Intuitively, the kurtosis reflects the cusp of the peak. Multiple studies have shown that kurtosis, as a texture feature, has high application value in cancer differentiation, tumor heterogeneity differentiation, and tumor staging [35, 36]. In this study, HER2-positive patients had higher kurtosis values than HER2-negative patients, which may be related to the heterogeneity of GC. The remaining four features were all based on the CT images. GLRLM is a basic feature of CT, which is obtained by recording the occurrence of multiple consecutive same pixel values in a one-dimensional direction in the image. A larger length distribution indicates a shorter run length and finer texture. GLRLM_LRE is a method for measuring the run-length distribution, with larger values indicating longer run lengths and coarser structural textures. GLRLM_GLN is always used to measure the intensity value of the image, and a lower GLN value is related to a higher intensity value of the image. GLRLM_RLN measures the similarity of the run lengths in the entire image, and a lower value indicates more uniform run lengths in the image. GLRLM_RLNU describes the length of the grayscale level inhomogeneity run, with lower values representing more uniform run lengths in the image. Lee et al. [37] divided thymus epithelial tumors (TETs) into three subgroups (low-risk thymoma, high-risk thymoma, and thymoma) based on pathological results, which showed differences in texture heterogeneity on PET/CT images. This study showed that most features derived from GLRLM showed good differentiation ability. Similarly, Kunimatsu et al. [38] compared glioblastoma and primary central nervous system lymphoma using texture analysis based on MRI images, and the results also proved that GLRLM features had good distinguishing performance.
In this study, PET/CT image data were obtained using two instruments with different parameters. A few studies have found that the radiomics features change with different parameters such as reconstruction layer thickness and algorithm, but a number of studies in recent years have shown that there is no significant statistical difference between the radiomics features extracted from two different devices [39, 40].
In conclusion, this study shows that 18F-FDG PET/CT radiomics features before treatment are of great value in predicting HER2 expression status and provide a noninvasive and dynamic detection method to determine HER2 expression levels. This has important significance in guiding the clinical formulation of individualized treatment schemes and in evaluating and predicting the efficacy of HER2-targeted therapy.