In this retrospective study, we differentiated between glioma and solitary brain metastasis from lung cancer and its subtypes by using a radiomics approach, and also adopted MLP, SVM, RF, and LR models. The results showed that the radiomics models had a high diagnostic efficiency in the differentiation of high grade glioma and solitary brain metastasis from lung cancer and its subtypes, in which, the MLP model had the best diagnostic efficiency in the validation groups.

One of the findings in this study was that the radiomics model was established for accurate preoperative differentiation of HGG and SBM. Since HGG had similar imaging findings to brain metastasis, especially solitary brain metastasis, the accuracy of preoperative diagnosis was low. Glioma and brain metastasis had different origins, as well as staging, screening, and treatment methods. Generally, gliomas only require focus on intracranial lesions, which rarely metastasize to other sites. As for brain metastasis with unknown primary lesions, key sites should be screened to identify the primary lesions for staging, to determine the subsequent treatment. Therefore, preoperative differentiation played an important role. Although the operation or aspiration biopsy can differentiate between the two tumors, when the tumor was close to or in a functional area of the brain or when the patient was too weak to undergo the operation, only the non-invasive mode could be selected for differentiation [15]. Conventional MRI is the most commonly used imaging differentiation method, however, it has low accuracy, for example, HGG was the most common type with peritumor enhancement of T1WI (about 40%), followed by tumor metastasis (about 30%) [16]. Some studies showed that multi-parameter MRI could better differentiate between the two tumors [5,6,7]. In the case of multiparametric joint application, Kerim et al. [6] established an optimal model for the differentiation of HGG and brain metastasis, with an AUC of 0.970. However, the results are affected by the scanning machine, scanning conditions, and the post-processing methods used in clinical practice [17].

The radiomics approach has become a promising method for tumor diagnosis, outcome prediction, and prognosis evaluation [18]. It was found to have outstanding advantages in the analysis of different types of tumors and their prognosis. Currently, studies have been conducted involving the lung, liver, cardiovascular, and neurological areas [19,20,21,22]. Moran et al. [23] analyzed 212 patients with glioblastoma and 227 patients with brain metastasis based on machine learning, and concluded that the SVM classifier could best differentiate between the glioblastoma and brain metastasis, with the average accuracy of 0.85, sensitivity of 0.86, specificity of 0.85, and AUC of 0.96. In this study, the accuracy and AUC of the four HGG/SBM classifiers were higher than those in the previous studies; MLP had the highest comprehensive diagnostic effectiveness, and the accuracy and AUC in the validation groups of the HGG/SBM were 0.992 (accuracy) and 1.000 (AUC), respectively. Karoline et al. [24] demonstrated that the heterogeneity of the solid part of the tumor had no significant difference; however, the heterogeneity of GBM peritumoral edema was significantly higher than that of solitary brain metastasis, with the sensitivity and specificity of 80% and 90%, respectively. In contrast to the study conducted by Karoline et al., Qin et al. [25] extracted the features of the solid part of the tumor, differentiated between glioblastoma and SBM based on machine learning, and obtained an AUC of 0.94 and accuracy of 95%. Our results are consistent with the results of their study. Machiko et al. [26] found that machine learning based on a combination of texture parameters from multiple sequences outperformed machine learning from individual sequences. The AUC of the model for differentiating HGG and brain metastasis in combination with T2WI, CE-T1WI, and ADC was 0.92, and the AUC of a simple sequence was inferior to the results of the combination. Studies have used radiomics methods based on CT, PET, and PET-CT fusion to predict the survival of patients with head and neck squamous cell carcinoma, and have achieved good results [27, 28]. Here, we studied the single sequence of CE-T1WI The reasons are as follows: (1) In the delineation and feature extraction of ROI, there are more single sequences with less workload and shorter time required. This study attempts to explore simpler, faster, and more efficient radiomics methods for distinguishing; (2) Some lung cancer patients only underwent a single sequence examination to determine whether there was intracranial metastasis. In order to obtain more data and make the model more accurate, this study first only used CE-T1WI images. We obtained better results. As for the reasons, firstly, the authoritative algorithm SVM was generally used in the past, however in this study, we used MLP, which was confirmed to be significantly superior to SVM. Secondly, previous studies focused on the differentiation of HGG and brain metastases of multiple primary foci (lung, breast, gastrointestinal tract, etc.); whereas in clinical practice, the larger SBM from lung cancer was harder to be differentiated from glioma, and there was a higher incidence of brain metastasis from lung cancer. Therefore, the results of this study are more accurate and have greater clinical significance.

As per the results of several studies, the most commonly used classifiers for differentiating glioma and tumor metastasis were support vector machine (SVM) and k-nearest neighbor (KNN); SVM performed well, and as a traditional machine learning method, it is a commonly used algorithm in neuroradiological predictive modeling by virtue of its simplicity and flexibility [29]. We used the MLP classifier in this study. With a wide application scope and strong scalability, MLP can fit complex functions with a generic function approximation method, and solve the non-linear classification problem. As no classifier is universally considered the best and accepted in clinical practice, we compared MLP and the three authoritative classifiers, and evaluated their diagnostic efficiency in this study. The results revealed that MLP had good diagnostic efficiency in the validation group, and the results of other classifiers were higher than the previous diagnostic efficiency. New directions and ideas have been provided for future radiomics research to distinguish between gliomas, metastatic tumors, and other brain tumors. Our research also utilized data augmentation methods to address the most common issue of small data volume in radiomics training models.

In this study, we explored the machine learning method for preoperative differentiation of low-grade glioma and brain metastasis. For most low-grade gliomas, especially those that are Grade I as classified by the WHO guidelines, the MRI showed a low signal without significant enhancement, and no edema, necrosis, or cystic areas. The imaging findings of these gliomas and brain metastases were significantly different, and they could be easily differentiated preoperatively. Some lesions tent to have similar imaging findings to brain metastases, such as ring enhancement and peritumoral edema. Peritumoral edema of low-grade glioma and brain metastasis may be caused by secondary ischemia due to the stress of the tumor on the peripheral cerebral parenchyma; therefore, they cannot be easily differentiated based on conventional MRI and even functional MRI (such as MRS and ASL). In this study, a total of 14 features extracted by machine learning were used for modeling and differentiating LGG/SBM, and a higher accuracy and AUC were obtained.

Another finding in this study is that the radiomics model based on machine learning can differentiate between HGG and brain metastasis from small cell lung cancer/non-small cell lung cancer. At present, there is no research on distinguishing subtypes of gliomas from solitary lung cancer brain metastases and lung cancer brain metastases. Small cell lung cancer and non-small cell lung cancer have different treatment modalities, and the treatment of brain metastasis may differ greatly. As for patients with asymptomatic brain metastasis from NSCLC, systemic therapy can be performed, and as for symptomatic patients with no more than three brain metastases, surgery, stereotactic radiotherapy (SRT), or SRT combined with whole brain radiotherapy (WBRT) can be performed. For patients with asymptomatic brain metastasis from SCLC, systemic chemotherapy can be performed, followed by WBRT, while for symptomatic patients, WBRT can be performed actively [30, 31]. Patients with brain metastasis from SCLC and NSCLC are treated with different chemotherapy drugs [32, 33]. In this study, we explored a method that can differentiate between brain metastasis from SCLC/NSCLC and glioma, and its diagnostic efficiency was high. Therefore, it can identify whether the primary focus in the lung is small cell lung cancer, thereby providing a reference for clinical treatment and medication; furthermore, some patients cannot undergo surgery or aspiration biopsy to determine the pathological types of lung cancer due to physical or other reasons.

This study has several limitations, firstly, as a retrospective study, the sample may be biased; secondly, the data volume of high/low grade glioma was small—there were only 34 patients with low grade glioma, and 172 patients with glioma, showing a big difference in data, however, the data enhancement method adopted in this study reduced the difference. ROI was outlined manually, showing strong subjectivity, great workload, and low efficiency. The data in the training and validation groups were collected from the same hospital, and were not validated by an external validation set; they had the same scanning parameters and a single source of patients. Another drawback of this study is that the clinical symptoms and laboratory indicators of gliomas and brain metastases were not included in the study. We only used machine learning methods and did not explore deep learning methods. In terms of experimental design, only a single sequence CE-TIWI was used to study the tumor enhancement area, without including edema around gliomas and metastatic tumors, adjacent cortical signal changes, other sequences other than CE-T1WI, or CT, PET, and other images. When delineating ROI, there is also uncertainty in determining the boundary of the necrotic zone. In the future, these shortcomings can be resolved by expanding the sample size, working with multiple centers, including the clinical symptoms and laboratory indicators of gliomas and brain metastases, and exploring automatic outlining of ROI [34,35,36].

To summarize, the MRI radiomics model has a certain application value for the differentiation of glioma and solitary brain metastasis from lung cancer and its subtypes, and can provide more information for clinical evaluation, staging, medication, and prognosis, thereby offering new ideas for precise differentiation of glioma and solitary brain metastasis from lung cancer and its subtypes. In the future, we can also conduct research on machine learning or deep learning based on tumor surrounding edema, adjacent cortical signal changes, sequences other than CE-T1WI, or images such as CT and PET.