In this present study, we analyzed brain MR images from four sequences in 122 patients with BMs of lung adenocarcinoma. 960 radiomics features were extracted from each normalized CE-T1WI, FLAIR, DWI and CE-SWI sequence, and then were reduced to 27 key features (9 from CE-T1WI, 10 from FLAIR, 6 from DWI, and 2 from CE-SWI) by the LASSO selection method. We built machine-learning models to predict EGFR mutation status from selected features on the training set and evaluated the predictive ability of models on the testing set. We found that radiomics features from CE-T1WI, FLAIR, DWI, and CE-SWI sequences of brain MR could differentiate wild-type and mutant EGFR in BMs of lung adenocarcinoma.
Relevant research suggested that the age distribution of EGFR mutations status may count on different types [32, 33]. In this study, no statistical differences in EGFR mutation status were found by statistical analysis for gender and age, which may be due to the limitation of sample size in our study. The number of BM lesions characterized by statistical analysis was statistically different in terms of EGFR mutation status.
We extracted 960 imaging radiomics features from each MRI sequence, and selected 27 statistically significant radiomics features after Mann-Whitney U-test and LASSO feature selection. We visualized the feature importance of the extracted features on each sequence in the form of a bar chart and performed feature correlation analysis, showing the correlation between the features in the form of a heat map. From Fig. 2, it could be seen that most of the color blocks have low color saturation, the correlation between features was low and most of the redundant features have been removed.
The CE-T1WI + FLAIR + DWI + CE-SWI model showed excellent performance which yielded an ACC of 0.9167, AUC of 0.9720, Sensitivity of 0.9167, and Specificity of 0.9015 on our testing set, which was much higher than that of the CE-T1WI model (ACC: 0.7917, AUC:0.8531), CE-T1WI + FLAIR model (ACC:0.9167, AUC:0.9231), and CE-T1WI + FLAIR + DWI model (ACC: 0.8333, AUC:0.9371).
According to previous studies, radiomics has been used to assess the mutation status of distant metastases in primary cancers. Related studies have assessed the ability to identify the mutation status based on the radiomics features from primary lung cancer CT images [34–36]. However, with the advancement of imaging technology, a growing number of small brain metastases can be detected in some asymptomatic patients before the primary lung cancer is diagnosed. It is essential to build new radiomics models for rapid identification of mutation status rather than models based on primary tumor images. Furthermore, it is impractical to build radiomics models by extracting features from primary tumor lesions in patients who have received appropriate treatment (e.g., radiotherapy, chemotherapy, or surgical resection). Therefore, there is a need to develop an effective method based on conventional brain MRI from patients with BMs of lung adenocarcinoma, which will facilitate the development of personal treatment strategies. Recently, Wang et al. [12] built a radiomics signature with the logistic regression based on T2-FLAIR sequence to predict EGFR mutation status, yielding an AUC of 0.987 and an accuracy of 0.991 in their validation set. Ahn et al. [37] explored radiomics features from brain CE-T1WI images to predict EGFR mutation status in patients of BMs with primary lung cancer and the highest diagnostic performance was reaching an AUC of 0.8681 in the test set. Li et al. [38] used radiomics analysis to differentiate the EGFR and ALK gene mutations and their T2-FLAIR model achieved an AUC of 0.950. However, few studies have used all CE-T1WI, FLAIR, DWI, and CE-SWI sequences from BMs of lung adenocarcinoma to identify EGFR mutation status. As far as we know, our work is the first study involving the CE-SWI sequence and directly uses four brain MRI sequences (CE-T1WI + FLAIR + DWI + CE-SWI) to distinguish EGFR mutation status in BMs of lung adenocarcinoma. Our model can integrate multimodal complementary information to achieve more accurate prediction accuracy.
Our study remained some limitations. First of all, it was a retrospective study based on a single center and more prospective multi-center validation will have to be done in the future. Furthermore, due to the insufficient sample size, only wild-type and mutant EGFR mutation status were distinguished in this study, and more data will have to be collected to subdivide the mutation types and develop a radiomics signature to distinguish common EGFR mutations from rare mutations.
In conclusion, MRI sequence-based radiomics features are valuable in distinguishing the mutant EGFR and wild-type EGFR in patients with BMs of lung adenocarcinoma, especially features from the CE-SWI sequence, and our CE-T1WI + FLAIR + DWI + CE-SWI model achieved the best prediction result, which can serve as a non-invasive method to assist radiologists in judging EGFR mutation status.