Lung cancer patients frequently develop brain metastases (BMs), and these patients account for 51% of all BM patients (1). Epidermal growth factor receptor (EGFR) mutations are detected in 10%–60% of all non-small cell lung cancer (NSCLC) patients (2), and are associated with poor survival (3). Ligand binding to EGFR leads to receptor tyrosine kinase activation and mediates cell proliferation and invasion (4). Previous studies have shown that EGFR tyrosine kinase inhibitor treatment improves survival in patients with advanced NSCLC and sensitive EGFR mutations (5, 6). Thus, the determination of EGFR mutation status is critical for prognosis and treatment.

Discordance in EGFR status between primary lung tumors and BMs has been increasingly reported (7–9), indicating that it is not completely accurate to determine the EGFR status of BMs based on the status of the primary tumor. Therefore, molecular diagnostic tests are now recommended by clinical guidelines, to determine the eligibility of patients with advanced NSCLC for targeted therapies (10, 11). However, barriers remain to defining the EGFR status of BMs. First, magnetic resonance imaging (MRI) is the preferred method for BM screening, diagnosis, response evaluation, and follow-up, as radiologists can use it to depict the distribution and morphological characteristics of the BMs. However, MRI cannot be used to determine the molecular status of the BM. Second, obtaining BM materials by biopsy or resection may not be feasible depending on the patient’s status. Additionally, the risks of neurosurgery, sampling bias, and the fact that the procedure does not always provide an accurate account of the intrinsic intertumor and intratumor heterogeneity must be considered. These issues emphasize the need to develop an innovative approach for deriving biomarkers of metastasis. Radiomics is an emerging technology that extracts high-dimensional features from images to mine the potential biological characteristics of tumors. Studies have evaluated the relationship of radiomics features with the isocitrate dehydrogenase gene status of gliomas (12) or the BRAF gene status of melanoma BMs (13). Although several studies have applied radiomics to identify EGFR mutations in either BMs or primary lung cancers using brain MRI, the study populations were relatively small, especially for patients with EGFR mutations, or the EGFR mutation status of the BMs was determined based on the primary tumor status, rather than samples obtained from the BMs (14–23).

Therefore, the aim of this study was to establish a radiomic model via machine learning to predict the EGFR status of BMs confirmed by postoperative histopathology, using preoperative brain MRI sequences. We hypothesized that differential EGFR expression levels in BMs could be captured by radiomic signatures.

In this proof-of-concept study, we extracted radiomic features from multiple MRI sequence images (T1CE, T2WI, and T2 FLAIR) of excised BMs originating from lung cancer and used these features to build machine-learning models for the classification of EGFR mutation status in BMs. Compared with a single sequence, the combination model, which extracted 10 key features from three sequences, achieved higher overall identification performance, yielding an AUC value of 0.81 in the independent test set. Additionally, the rate of discordance of EGFR mutation status between primary lung tumors and paired BMs was 15.4% in the 52 patients who underwent EGFR gene testing in both the primary tumor and the BM. Our findings indicate that the proposed radiomics signatures based on brain MRI can distinguish between mutant and wild-type EGFR in BMs, and the switch in EGFR status observed between the primary tumor and the BMs also indicates the importance of considering that the EGFR gene mutation status may differ between the metastases and the primary tumor.

New molecular agents targeting specific pathways have been developed and key molecules in tumor growth and progression have been identified. A typical example of such a target is the EGFR gene, which is an indicator of targeted treatment, an independent predictor of the treatment response, and a predictor of outcomes (26–28). Given the inconsistencies in target gene expression between primary tumors and their distant metastases, molecular diagnostic testing is now recommended for metastases in patients with advanced NSCLC whenever possible, to determine their eligibility for targeted therapies. Such assessments are recommended by the American Society of Clinical Oncology (29) and the European Association of Neuro-Oncology-European Society for Medical Oncology (10). Currently, however, it is not always practical to obtain a specimen of the BM by biopsy or surgery.

Therefore, several studies have used radiomics models to noninvasively predict the EGFR mutation status of lung cancer or BMs using brain MRI (15, 21). Ahn et al. extracted features from T1CE of 61 patients comprising 210 BMs with a size > 5 mm, and used several machine-learning algorithms to predict the EGFR gene mutation status of primary lung cancer, reaching an accuracy of 86.7% (AUC, 0.868) (15). In a similar study, Chen et al. built a model based on radiomic features generated by T1CE and T2 FLAIR (110 patients with 452 lesions, of whom 75 were EGFR positive) and clinical data using random forest classifiers, to classify EGFR, anaplastic lymphoma kinase, and Kirsten rat sarcoma virus gene mutation status in primary lung tumors and generated AUC values of 0.91, 0.92, and 0.99, respectively (21). However, both of these previous studies assumed an identical molecular profile in the BMs, thus overlooking possible discordances in EGFR mutation status between the lung cancer and the BMs. Additionally, there was no separate test set to validate the model performance, which may have led to overfitting.

Limited efforts have been focused on radiomics signatures to detect EGFR mutation status in BMs. Wang et al. analyzed four sequences (T1CE, T2WI, T2 FLAIR, and diffusion tensor images [DWI]) collected from 52 lung adenocarcinoma patients (28 with mutant EGFR, 24 with wild-type EGFR) (23). Although they concluded that the radiomics signature of T2 FLAIR achieved an AUC of 0.871, an accuracy of 0.845, a sensitivity of 0.901, and a specificity of 0.891 for discriminating EGFR mutation status using an independent testing data set, they also assumed that EGFR expression was consistent between the metastatic tumor and the primary tumor, which may not be accurate as discussed above. Haim et al. applied a deep-learning approach, using a ResNet-50 convolutional neural network, to predict EGFR mutation status in NSCLC BMs based on the EGFR testing results from resected BMs (20). However, they used data from a small cohort of 59 patients, of which only 16 patients were EGFR-positive. Moreover, they cropped regions of interest of the mid-tumor region and ± two slices for each patient. Such areas may be not sufficient to represent the entire tumor and may miss the three-dimensional features of the tumor. In contrast to previous studies, we enrolled, to the best of our knowledge, the largest reported study population of patients who underwent resection of their lung cancer BMs, to propose a radiomics signature based on multiple sequences of brain MRI. Moreover, despite adenocarcinomas showing the highest EGFR mutation rate among all histological cancer types, we included all patients with lung cancer, unlike other studies that exclusively selected patients with NSCLC or adenocarcinoma. Furthermore, we evaluated the EGFR mutation status in resected brain samples, which may better reflect the real mutation status. In addition, we used an open-source tool, Pyradiomics, for radiomics feature extraction, which may have improved the reproducibility of the feature extraction process.

We also found that the combination of features from multiple sequences had better classification performance than a single sequence, which was consistent with the study of Park et al. (18). Compared to single sequence, they reported that features extracted from the integration of T1CE and diffusion tensor images improved the capacity to determine the EGFR mutation status of BMs derived from lung cancer. Of the 10 features analyzed in our study, the biggest contribution came from T2WI. Furthermore, more second-order features than first-order features were selected, implying that multiparametric high-throughput characteristics enable a more accurate assessment than single parameters. Of the single sequences used to predict EGFR status, we found that the radiomic signatures of T2WI had the best performance. This differs from the result reported by Wang et al. (23), who found that T2-FLAIR yielded better EGFR mutation discrimination than TICE, T2WI, and DWI. Our results indicate that multiple sequences have higher predictive value than single sequences for the determination of EGFR mutation status.

Another finding in our study was that the discordance rate between the primary tumors and the corresponding BMs reached 15.4%. These results were comparable to those of previous studies that have reported heterogeneity in EGFR mutations between primary tumors and BMs, with variability rates ranging from 12% to 33% (8, 9, 30). Discordance between primary and metastatic tumors may be explained by clonal selection and intratumor heterogeneity (31). Clonal selection during the multistep metastatic process, combined with the potential effects of the tumor microenvironment and/or the treatment, may explain the discordance observed in metachronous metastases. Moreover, cancer is a highly heterogeneous disease, and polyclonal cell lines may exist with various EGFR statuses. Finally, the effect of different techniques on discordance cannot be excluded (32). Notably, two rare mutations were found in our study. A male patient with adenocarcinoma had both a deletion in exon 19 and an L858R missense mutation in exon 21 in the primary tumor, but the mutation in exon 21 was lost in the BM. Another adenocarcinoma in a female patient was found to have an S768I insertion in exon 20 and a G719X missense mutation in exon 18, but, similarly, the insertion was lost in the BM. The mechanism responsible for these changes will be investigated in future studies. We did not observe any EGFR-positive primary tumors that switched to an EGFR-negative form in BMs. Our data suggest that gaining EGFR mutations or switching EGFR subtypes may be more frequent than the complete loss of EFGR mutations when the primary tumors metastasize to the brain (negative to positive vs. positive to negative, 17.6% vs. 0%, Yates’ continuity correction, P = 0.567; change mutation type vs. positive to negative, 11.1% vs. 0%, Fisher’s exact test, P = 0.486), but these differences did not reach statistical significance, possibly due to the small number of samples.

This study has several limitations. First, this was a retrospective single-center design, which may have created selection bias. The performance of the model should be validated using a larger prospective multi-center dataset. Nonetheless, this is the largest reported cohort exploring the feasibility of classifying EGFR expression in BMs based on radiomics. Second, as in most previous studies, a region of interest was delineated for the entire metastasis. We did not analyze the subregional features of the tumor, e.g., the areas with enhancement, necrosis, hemorrhage, or edema. Third, more novel techniques such as deep learning or functional MRI were not applied to extract features. However, using an open-source Python package to extract features may have improved the reproducibility. In addition, conventional MRI sequences have wider adaptability in clinical practice. Finally, we did not distinguish between mutation subtypes, e.g., common vs. rare or sensitive vs. resistant mutations, given the limited number of samples with rare and resistant mutations.

We demonstrated that it is feasible to apply a multi-sequence radiomic model to noninvasively predict the EGFR mutation status of lung cancer BMs. Moreover, the discordance observed between the primary tumors and the BMs indicates that EGFR alterations in metastases should be considered when a molecular targeted treatment is to be implemented.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Institutional Review Board of Sun Yat-Sen University Cancer Center, approval No. B2021-198-01. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

Conceptualization: LZ, HX, XL, HL, CX. Literature search: XL. Study design: LZ, HX, XL, HL. Data curation: HX, XL, YZ, YL. Software: HX, XL. Supervision: HL, CX. Visualization: LZ, HX, XL. Writing – original draft: LZ. Editing: HX, XL. Review & approval: all authors

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.