Differentiating Glioblastoma Multiforme from Brain Metastases Using Multidimensional Radiomics Features Derived from MRI and Multiple Machine Learning Models

Due to different treatment strategies, it is extremely important to differentiate between glioblastoma multiforme (GBM) and brain metastases (MET). It often proves difficult to distinguish between GBM and MET using MRI due to their similar appearance on the imaging modalities. Surgical methods are still necessary for definitive diagnosis, despite the importance of magnetic resonance imaging in detecting, characterizing, and monitoring brain tumors. We introduced an accurate, convenient, and user-friendly method to differentiate between GBM and MET through routine MRI sequence and radiomics analyses. We collected 91 patients from one institution, including 50 with GBM and 41 with MET, which were proven pathologically. The tumors separately were segmented on all MRI images (T1-weighted imaging (T1WI), contrast-enhanced T1-weighted imaging (T1C), T2-weighted imaging (T2WI), and fluid-attenuated inversion recovery (FLAIR)) to form the volume of interest (VOI). Eight ML models and feature reduction strategies were evaluated using routine MRI sequences (T1W, T2W, T1-CE, and FLAIR) in two methods with (second model) and without wavelet transform (first model) radiomics. The optimal model was selected based on each model's accuracy, AUC-roc, and F1-score values. In this study, we have achieved the result of 0.98, 0.99, and 0.98 percent for accuracy, AUC-roc, and F1-score, respectively, which have yielded a better result than the first model. In most investigated models, there were significant improvements in the multidimensional wavelets model compared to the non-multidimensional wavelets model. Multidimensional discrete wavelet transform can analyze hidden features of the MRI from a different perspective and generate accurate features which are highly correlated with the model accuracy.


Introduction
Glioblastoma multiforme (GBM) and brain metastases (MET) are the most common malignant brain tumors in adults [1,2]. The distinction between these two types of tumors is crucial to subsequent diagnostic and therapeutic planning [3,4]. An accurate diagnosis of the tumor's source and extent is crucial [5,6]. In addition, treatment strategies for these tumors differ; total en bloc resection is preferred for MET, while stereotactic radiosurgery can be used for MET less than 3 to 4 cm, whereas GBM should be treated with maximal resection followed by molecular classification and simultaneous chemoradiotherapy [7][8][9]. Diagnosis of GBM and MET is based on histopathology biopsy [9,10]. It is particularly dangerous for older adults and tumors near eloquent areas [9,10]. In routine magnetic resonance imaging (MRI), GBM and MET show peripheral edema and ring enhancement [11,12]. Even though the two lesions had different treatment strategies, similar radiological appearances made it difficult to differentiate them. Many years of radiological research have focused on accurately distinguishing these two lesions [11,13,14].
A radiomics study may provide pathophysiological insights otherwise hidden in quantitative imaging data [12]. In radiomics research, some features are used, such as size, shape, image intensity, and voxel relationships [4][5][6][7][8]. A wide range of problems can be solved using multidimensional wavelets [15,16]. An effective way to reveal hidden characteristics of signals is to use discrete wavelets (DWT) [15,16].The wavelet transform is a robust tool in signal processing, and it could provide us with deep and precise insight into the structure of the signals [15,16].
It has been reported that advanced techniques of MRI, such as perfusion-weighted imaging (PWI), diffusion tensor imaging (DTI), MR spectroscopy (MRS), and amide proton transfer-weighted imaging, play vital roles in the diagnosis of GBMs compared with MET; however, these advanced techniques may not be included in all standard MRI protocols [3,12,20]. Any single finding cannot guide clinical practice in some cases due to diagnostic uncertainty. Many previous studies showed that combined conventional MRI (cMRI), diffusion-weighted imaging (DWI), and 18F-FDG-PET images to establish different radiomics models to differentiate MET from GBM and found that the integrated model based on cMRI, DWI, and 18F-FDG-PET had the best discriminatory power. In contrast, advanced sequences like DWI are not widely available in the clinic as cMRI [3,12,20]. Consequently, the radiomics literature shows that different classifiers have different outputs [1][2][3][4][5][6]. Choosing the best model is complex [3][4][5][6]. It is necessary to build various models to achieve a more excellent result [3,12,20].
In this study, we have used standard MRI sequences (T1weighted imaging (T1WI), contrast-enhanced T1-weighted imaging (T1C), T2-weighted imaging (T2WI), and fluidattenuated inversion recovery (FLAIR)) to extract the features from the dataset. By combining these simple features with DWT, we developed suitable feature vectors that can be used in machine learning algorithms to differentiate between GBM and MET patients. The objective of our study was to develop a convenient, accurate, and stable predictive model to aid clinical investigations in differentiating MET from GBM without surgery.

Dataset and Patient
Population. This retrospective study was approved by the Tarbiat Modares University Institutional Review Board, and informed consent from the patients was waived (IR.MODARES.REC.1400.076).
In all cases, the pathology diagnosis was based on WHO standards and was obtained from Hazrate Rasool Akram Hospital, affiliated with Tarbiat Modares University [21]. We have collected MRI data from 91 patients (GBM: 51, MET: 40) based on their pathological confirmation.
Selected features from LASSO were imported to this ML, and the predictive ability of each algorithm was primarily assessed using the AUC of receiver operating characteristic (ROC) curve analysis [15,16,25,26,38].

Second Radiomics Model Establishment (Wavelet-Based
Features). Multidimensional wavelet transforms were created by importing selected features from LASSO. This means that the low pass filter generates the approximate coefficient, and the high pass filter would result in the detail coefficients. The approximate coefficient is the most similar signal to the original signal. The detail coefficient consisted of three matrices: vertical, horizontal, and diagonal [15,16]. We have considered 31 different wavelet filter banks from four distinct families to provide a wide range of feature vectors from different wavelet filter banks. The 31 different wavelet filter banks " bior1. 3 Figure 2: Flowchart of the process of wavelet radiomics. The tumors were segmented on all MRI images to form the volume of interest (VOI). Different filter banks are applied to them and the machine learning algorithm was used to fit the best predictive model. 4-Dimensional signals. These signals were then considered as input signals to multidimensional discrete wavelets. The approximate and detail coefficients were substantially generated and saved in a python array. Approximate coefficients were the most similar signal to the primary signal, and the detail coefficients consisted of horizontal, vertical, and diagonal details (cH1, cV1, cD1). These approximate coefficients and detail coefficients were generated by low and high pass filters, respectively. We calculated eleven different criteria for approximation and detail coefficient matrixes. Seven of these criteria were: maximum, minimum, average, median, standard deviation, Shannon entropy, and signal energy applied to entire approximate coefficient and detail coefficients matrices and led to 28 different feature vectors (7 * 4). The other two were the standard error and slope between approximate coefficient and detail coefficient matrixes generating six other features (2 * 3). Eventually, two signal energy and wavelength cri-teria were used for the whole signal (original input signal), which led to 2 other features and a total of 36 features for every single 3D Slicer output sample. These features were saved in the text file and used as the input file for the machine learning model. These procedures have been done for 31 different wavelet filter banks mentioned earlier, and for each filter bank, we have made a separated feature vectors file.
To be more precise, each filter bank had a unique profile; hence, we had 31 different profiles for each sample. Eight ML used these profiles separately as an input file. Finally, we have reported our best result as our proposed model.

Evaluation Method.
Regarding the dataset size, we used the 5-fold cross-validation. To avoid bias, we repeated the 5fold cross-validation test 100 times. Eventually, the iteration average was considered the model's outputs. The classification influence can be evaluated using three indicators: area under the receiver operating characteristic curve (AUC), accuracy (ACC), and F1-score. Finally, a pairwise test was applied to compare the obtained ROC curves, and the ROC curves of the wide variety of classifiers were then investigated. According to the analysis, p < 0:05 illustrates that the two ROC curves were statistically significant differences [20,24,25,[38][39][40]. Workflows are shown in Figures 1 and 2.

First Radiomics Model Result.
We have reported the result of our proposed model by a set of conventional parameters, including F1-score, accuracy, and AUC. These parameters could best express the model's performance in machine learning research. As mentioned in the first model, we have used the selected features obtained from 3D Slicer and used SVM, NB, MLP, DT, ADA, KNN, LR, and RF to establish our predictive model. The highest accuracy for these models is achieved by random forest algorithms, shown in Table 2.

Second Radiomics Model Result.
We used the selected feature vector from the 3D Slicer as the multidimensional DWT input for the second model and extracted the approximate and detailed coefficients. After concatenating the approximate coefficient and detail coefficient matrices, particular criteria have been calculated. The same eight ML were applied to establish the predictive model. We calculated accuracy, F1-score, and AUC of ROC in 31 different filter banks, as presented in Tables 3-5. Our result proved that the second model could perform better than the first model (sig ≤0.05). This model's DB5 filter bank and Logistic Regression achieved the highest result. We have gained 0.98, 0.99, and 0.98 percent for accuracy, AUC-roc, and F1-score, respectively. We introduce the DB5 wavelet and Logistic regression as our proposed model

Discussion
We explored the diagnostic performance of radiomics using traditional machine learning classifiers for differentiating GBM from single MET. The wavelet radiomics performed better than the best-performing traditional radiomics and demonstrated good generalizability in the testing data. Four imaging modes were used, 107 features were extracted from each sequence, and all 428 parameters were used as LASSO inputs to select the best features. The selected features are used in two different ways. First, used in ML input, then transferred to 31 wavelet filter banks, and lastly imputed to ML. For the best combination of classification procedures, extensive comparative studies were performed for 31 types of wavelet features and eight classification algo-rithms (Adaboost, K.N.N., Gaussian NB, DT, LR, SVM, RF, and MLP). Moreover, the multidimensional discrete wavelet could reveal the hidden feature of the data. Multidimensional discrete wavelets provide us with approximate and detailed coefficient matrices. The approximate coefficient is the most similar wave to the original signal, while the detailed coefficient includes horizontal, vertical, and diagonal details. We found that wavelet feature extraction, a critical classification component, enables us to distinguish different tumor characteristics using different feature types.
According to our findings, wavelet radiomics-based ML can successfully discriminate GBM and MET. LR was judged to be the most effective model. Another important finding from our research was the diagnostic performance of the top models with wavelet features outperformed models without wavelet features (Tables 6-8).
Our study's most important outcome was identifying appropriate discriminative models for lesions in the brain. A number of previous studies have used multiparametric MRI data to discriminate between GBM tumors and MET, including advanced imaging methods such as diffusion, perfusion, and MR spectroscopy [1][2][3][4][5][6]. It should be noted that advanced imaging is not incorporated into all MRI protocols across all sites and is highly dependent on the acquisition and analysis method [14,34,41]. Therefore, it is important to be able to classify brain tumors based on common sequences [14,34,41]. Some previous studies attempted to distinguish between different types of brain tumors based on a single contrast; however, these studies were carried out on relatively small populations and were limited to data obtained from a specific MRI system [14,34].
Cho1 et al. [42] studied radiomics and ML in glioma grading classification, which showed that RF and LR were high-performing models. In our investigation, the LR model with db5 wavelet feature performed better (AUC 0.99) than the LR model in a previous study by Cho et al., which had an AUC of 0.95. These findings demonstrate that the proposed model has an impact on model performance. Our results were better than the previous study that looked at various ML. Priya et al. [22] used the LASOO and Elastic Net model to classify brain tumors; we have higher results in wavelet base feature (0.99 vs. 095) and worse in part without wavelet base features (0.95 vs. 0.93).
Ning et al. [43] examined seven ML classifiers and five feature reduction techniques using radiomics features produced from T1-CE and T2W pictures and obtained an AUC of 0.890 with an accuracy of 83 percent. The second part of their work used deep neural networks, which had an AUC of 0.95 and an accuracy of 89%. Even though we did not analyze deep neural networks (DNN) due to their computational cost, our findings were comparable to Ning et al.
Su et al. [17] extracted features from the T1-CE sequence, then evaluating 30 model combinations and feature reduction for their radiomics obtained an AUC of 0.80. Using 248 combinations of wavelet feature series and classifier, we achieved a higher AUC of 0.99 and an accuracy of 98 percent. The fact that we extracted wavelet features from MRI sequences improved our results.
Wavelet feature is an essential element of this model performance. The improved effectiveness of 31 types of wavelet features compared to a priori feature (without wavelet) was a key conclusion of our research. LR and RF models provide a more significant generalizability benefit than other ML models. We have found the essential type of wavelets for Ada, Knn, Nb, Dt, LR, SVM, RF, and MLP were db8, coif2, bior2.6, bior1.5, db5, db2, bior1.5, and coif3 consequently.
Adding wavelets to the analysis confirms that it can better distinguish MET and GBM than the radiomics model based solely on c MRI features. Therefore, our study provides a good solution to the problem of poor model performance in radiomics research. We propose a unique, helpful prediction model based on MRI sequences (T1-W, T2-W, T1C-W, and Flair-W) that can be used in clinical practice.

Conclusion
We found that radiomics-based ML can accurately classify GBM and MET. Also, in this study, the best results were in the LR algorithm and wavelet db5, which can be considered acceptable in data with few samples. The performance of a model might vary based on the mix of classifier and feature types used. Thus a complete model selection method should be used. Also, the result of the models applied to the extracted features' composition is very suitable compared to their separate modes.

Data Availability
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Consent
Informed consent was received from all participants.

Conflicts of Interest
The authors declare that they have no competing interests.