Comparison of machine learning classifiers for differentiation of grade 1 from higher gradings in meningioma: A multicenter radiomics study

The Current State of Radiomics for Meningiomas: Promises and Challenges

“…In the initial stages of radiomics analysis experimentation, multiple studies have explored the feasibility of various radiomic features in the prediction of pathological grade of meningiomas. The results showed that both conventional radiomic features, including the shape, histogram, texture, gray-level run length matrix, wavelet transform, and other higher-order statistics (19,(34)(35)(36)(37)(38)40), and the DLR features (39) could predict the tumor grades. Yan et al have identified two textural features based on the run length matrix and two shape-based features significantly related with the WHO grade II meningiomas; Similarly, in terms of the low grade meningiomas (WHO grade I), one textural feature based on run length matrix and one shape-based feature were selected (35).…”

“…Because the fluctuance of parameters from second or higher order statistics revealed irregular changes in the gray pixels in aggressive meningiomas due to the intratumoral nonuniform structure tissue (49). Furthermore, it seems that diversified combinations of these features, such as a combination of radiomic features from different feature categories, multiple imaging sequences, heterogenous raw data or combined with qualitative imaging features or clinical data, could improve the performance of the classification models even if those improvements may not always be significant (19,34,37,39).…”

“…Since there has been no standardized guidance of algorithms selection yet, the selection usually depends on the preference and experience of analysts (32). At present, the classification methods presently applied for grade prediction include the random forest (RF) (19,34,37,40), logistic regression (LR) (35,36), naïve Bayes (NB) (35), support vector machine (SVM) (35,37,38,40), eXtreme gradient boosting (XGBOOST) (37), multilayer perceptron (MLP) (37), and linear discriminant analysis (LDA) (38,39) (Table 3). Among these various algorithms, numerically, the best performance of prediction was achieved in a tree-based classification algorithm, XGBOOST, which based on a combination of features derived from multiple MRI sequences and yielded a high AUC of 0.97, a sensitivity of 1.0 and a specificity of 0.97 (37).…”

“…At present, the classification methods presently applied for grade prediction include the random forest (RF) (19,34,37,40), logistic regression (LR) (35,36), naïve Bayes (NB) (35), support vector machine (SVM) (35,37,38,40), eXtreme gradient boosting (XGBOOST) (37), multilayer perceptron (MLP) (37), and linear discriminant analysis (LDA) (38,39) (Table 3). Among these various algorithms, numerically, the best performance of prediction was achieved in a tree-based classification algorithm, XGBOOST, which based on a combination of features derived from multiple MRI sequences and yielded a high AUC of 0.97, a sensitivity of 1.0 and a specificity of 0.97 (37). While the most widely used algorithms are the RF and SVM, the RF is an ensemble method that calculates multiple decision tree-based classifiers containing several identically distributed random independent vectors (37, 51), whereas the SVM is a non-linear classifier that iteratively constructs a hyperplane or highdimensional feature space consisting of a series of hyperplanes that separates different classes (52,53).…”

“…While the most widely used algorithms are the RF and SVM, the RF is an ensemble method that calculates multiple decision tree-based classifiers containing several identically distributed random independent vectors (37, 51), whereas the SVM is a non-linear classifier that iteratively constructs a hyperplane or highdimensional feature space consisting of a series of hyperplanes that separates different classes (52,53). There have been two studies comparing different performances of the RF and SVM, Hamerla et al have built four different classification models, including the RF, XGBOOST, SVM, and MLP, based on same radiomic features; Their results demonstrated both RF and SVM had same AUC of 0.93 (37). However, in the study of Park et al, the SVM have shown to have a better predicting performance with an AUC of 0.86 comparing to 0.84 in the RF (40).…”