In this study, we constructed three kinds of radiomics models: (1) intratumoral model (VOI_I model); (2) peritumoral model (VOI_P model); (3) intratumoral and peritumoral model (combined model). We found that combined models showed great promise in predicting the EGFR mutation status of lung adenocarcinoma patients. The best prediction performance was obtained by VOI4 model, with the highest AUCs of 0.877, 0.727, and 0.701 in the training and validation sets, the internal testing set, and the external testing set, respectively.

To our knowledge, few studies have revealed the added value of peritumoral radiomics in predicting EGFR mutation status in lung cancer. Choe et al. demonstrated that the predictive model combining intratumoral and peritumoral radiomic features performed slightly better in the training set than the intratumoral model, but the difference was not statistically significant (AUC = 0.66 vs. 0.64, p = 0.504), whereas, in the validation set, the AUC was lower than that of the intratumoral model (AUC = 0.56 vs. 0.62) [21]. Another study showed that compared to intratumoral radiomics alone, combining intratumoral and peritumoral 3 mm radiomic features significantly improved the predictive performance of EGFR mutation status in primary lung cancer (AUC = 0.730 vs. 0.774, p < 0.001), and in lung adenocarcinoma only (AUC = 0.687 vs. 0.630, p < 0.001) [22]. However, this study did not determine whether the 3 mm peritumoral region was optimal for evaluating peritumoral features. Ideally, to determine the best peritumoral range, we should extract features from different peritumoral ranges to construct models separately and compare their predictive performance. A recent study compared radiomic features of multiple peritumoral regions (3 mm, 5 mm, 7 mm) and constructed three machine learning models to predict EGFR mutation status in NSCLC. The results showed that combining intratumoral and peritumoral 3 mm radiomic features could better distinguish EGFR+ from EGFR− groups than 5 mm and 7 mm (training, p = 0.0000, test, p = 0.0025), but this study included only 164 patients and did not validate models with an external dataset [23]. Based on this, we expanded VOI_I outwards by 1 mm, 2 mm, 3 mm, 4 mm, 5 mm, 10 mm, and 15 mm to identify seven peritumoral regions and combined them with intratumoral regions to generate seven intratumoral and peritumoral regions, respectively, to compare the complementary value of different peritumoral regions to the predictive performance of radiomic models. In addition, compared to the previous studies, our study used a larger training cohort and was tested in an independent internal testing set and an external testing set. As a result, our model may be more effective in illustrating the differences in radiomic features between EGFR+ and EGFR− groups.

According to the results, the peritumoral region of lung adenocarcinoma may also provide important predictive information about EGFR mutations, with the best predictive performance achieved by combining intratumoral and peritumoral 4 mm radiomic features. Tumor cells are usually highly invasive and tend to migrate from the primary tumor to the surrounding parenchyma, disrupting the normal structure and causing morphological and textural changes in the peritumoral region. These changes are difficult to detect on medical images, whereas radiomic features extracted from CT images can quantitatively reflect subtle changes in the microenvironment surrounding the tumor that cannot be recognized by the naked eye, this may be the pathophysiological basis for the improved predictive performance of the combined models over the VOI_I model. Lung adenocarcinomas have obvious cellular and mutational heterogeneity. The concept of tumor heterogeneity applies not only to tumor epithelial cells but also to the various microenvironments with which the tumor cells interact, such as vasculature, cancer-associated fibroblasts, extracellular matrix, and infiltrating immune cells. Tumor cells can influence their microenvironment by releasing cell signaling molecules that promote tumor angiogenesis and induce immunological tolerance. Meanwhile, immunocytes infiltrated in the tumor microenvironment can secrete a large number of cytokines and chemokines to promote the epithelial-mesenchymal transition of tumor cells, which allows tumor cells to invade and metastasis [24].

The tumor margin is an important meeting place in the tumor microenvironment where immune and stromal cells are highly active and interact with the tumor. The microenvironment at tumor invasion edges differs from that of the tumor core. Hypoxia tends to be associated with the center of the tumor, whereas oxygen is primarily present at the tumor periphery. Monocytes in the blood are recruited around tumor cells by various chemokines and cytokines, thus becoming tumor-associated macrophages, which can promote the invasion of tumor cells by supplying pro-migratory factors such as epidermal growth factor, or by promoting extracellular matrix proteolytic remodeling, and play an important role in the invasion process of the tumor margin. Furthermore, under hypoxic conditions, tumor-associated macrophages promote tumor cell release of vascular endothelial growth factor and platelet-derived growth factor via the activation of the hypoxia-inducible factor-1 pathway, thus promoting tumor angiogenesis, providing oxygen and nutrients for tumor growth, and contributing to tumor cell invasion and metastasis. In addition, tumor-associated fibroblasts are also abundant at the tumor margin, promoting tumor proliferation, angiogenesis, invasion, and metastasis by secreting various growth factors, cytokines, and inflammatory chemokines [25, 26].

As in several previous studies, the most predictive radiomic features finally selected in our study included a significant number of texture features (235 in total), which reflect the pattern and spatial distribution of voxel intensities within the VOI, indicating its biological heterogeneity [15]. Therefore, our results may suggest that tumor heterogeneity is associated with EGFR mutation status in lung adenocarcinoma. Regarding the shape features, the shape_Flatness feature was found in all of the final selected features of 15 VOIs, which shows the relationship between the largest and smallest principal components in the VOI shape, suggesting that this feature plays an important role in predicting EGFR mutation status. However, unlike most other studies [16, 22, 27, 28] there were no first-order features in our best predictive model (VOI4). The first-order features describe the distribution of voxel intensities within the target region through commonly used and basic metrics, but it is difficult to measure the spatial distribution characteristics of voxels without considering the neighborhood relationship between voxels [29]. In our best predictive model, they are not critical predictive features.

In addition, we found that features from independent peritumoral regions also had predictive value for the prediction of EGFR mutations. Compared to other peritumoral radiomics models, the model based on the peritumoral 15 mm (VOI_P15) features achieved the best performance in the training and validation sets, the internal testing set, and the external testing set, with AUCs of 0.861, 0.716, and 0.704, respectively. However, this was inconsistent with findings that as peritumoral distance increased, the VOI comprised more normal lung tissue and relatively less tumor tissue, making the predictive performance of the model decreased [30]. The probable explanation was that radiomic features were more stable as peritumoral distance increased [31]. Tunali et al. also demonstrated that some radiomic features, including statistical features, histograms, and some texture features (GLCM, GLRLM, GLSZM, and NGTDM), had good stability and reproducibility regardless of peritumoral distance, indicating that these features were less influenced by changes in the size or shape of peritumoral regions caused by different segmentation and image acquisition [31]. It was generally consistent with the features eventually selected in our study, and these stable and reproducible features were more likely to construct robust radiomics models, allowing multicenter studies to maximize the clinical utility of radiomics models [32].

To achieve more generalizable and impactful results in radiomics, researchers need to obtain large patient cohorts by combining images from multiple institutions. However, most current radiomics studies collect imaging data retrospectively, and image acquisition protocols, processing or reconstruction settings, and imaging scanners may be different from different institutions, resulting in poor reproducibility and repeatability of radiomic features [33,34,35]. Therefore, in order to discover more reliable and stable radiomic features and apply them in multicenter clinical practice, image consistency must be improved by controlling imaging protocols in order to build a public database with a large amount of high-quality data [36]. In addition, several studies have demonstrated that the use of harmonization methods in the image domain (prior to feature extraction) or spatial domain (within or after feature extraction) would be beneficial in the design of multicenter studies. According to recent studies, ComBat harmonization is a fast and easy-to-use feature harmonization method in the feature domain that allows the correction of radiomic features to reduce the variation caused by different imaging protocols [37,38,39]. It was first proposed by Johnsond et al. [40] for genetic studies and was later used by Fortin et al. for medical imaging applications [41], and by Orlhac et al. [42] for PET radiomics studies, and had produced great results in several subsequent studies [39, 43, 44]. Among them, Shiri et al. demonstrated that ComBat harmonization could significantly improve the prediction performance when radiomics to predict EGFR mutation status in NSCLC, and the range of mean AUC increased from 0.87–0.90 to 0.92–0.94, which proved the effectiveness of ComBat harmonization [43]. Therefore, we can try to apply ComBat harmonization to further improve the prediction performance of the model in future.

Despite the encouraging results, there are still some limitations. First, we included some lung adenocarcinoma patients as an external testing set to validate the reliability and stability of the model, however, due to the small sample size, its predictive efficiency may be limited, and multi-institutional image data are needed to assess the generalizability of our findings in future; second, the incidence of EGFR mutation varies greatly across different races, with a significantly higher incidence in Asian populations [45]. The patients used for model training in our study were all Asians, making the results lacking in generalizability and requiring further validation in patients of other races; furthermore, some other potentially valuable factors such as smoking status and gender were not included in this study, and we will combine radiomic features with these clinical features for further research to improve the predictive performance of the model in future.

In conclusion, radiomic features extracted from the peritumoral region can add extra value in predicting the EGFR mutation status of lung adenocarcinoma patients, with the optimal peritumoral range of 4 mm. This may partially prove the clinical value of peritumoral microenvironment in cancer diagnosis.