Table 2.
Characteristics of the 10 studies reporting the highest accuracy results for their best performing models, including: glioma grade classification task, dataset source and size, ratio of high- to low-grade gliomas, validation technique, imaging sequences used in prediction, feature types used in prediction, best performing algorithm (based on accuracy results), and performance metrics.
| Paper | Glioma Grade Classification Task | Dataset | HGG : LGG Ratio | Validation Technique | Imaging Sequences | Features | Best Algorithm | Performance Metrics |
|---|---|---|---|---|---|---|---|---|
| Hedyehzadeh et al. (2020) (31) | 2/3 vs. 4 | TCIA (n=461 patients) | 1.3:1 (262 HGG, 199 LGG in total set) | Internal (4-fold cross-validation) | T1, T1CE, T2, FLAIR | Texture | Support Vector Machine | Accuracy = 1.00 Sensitivity = 1.00 Specificity = 1.00 |
| BashirGonbadi and Khotanlou (2019) (32) | 1/2 vs. 3/4 | BraTS (n=285 patients) | 2.8:1 (210 HGG, 75 LGG in total set) | Internal (Holdout, 15% of dataset) | T1, T1CE, T2, FLAIR | Deep learning extracted | Convolutional Neural Network | Accuracy = 0.9918 |
| Polly et al. (2018) (33) | HGG vs. LGG (unclear) | BraTS (n=160 images) | 1:1 (50 HGG, 50 LGG in testing set) | Unspecified | T2 | First-order, Shape, Texture | Support Vector Machine | Accuracy = 0.99 Sensitivity = 1.00 Specificity = 0.9803 |
| De Looze et al. (2018) (34) | HGG vs. LGG (unclear) | Single center hospital (n=381 patients) | Unclear | Internal (5-fold cross-validation) | T1, T1CE, T2, FLAIR, Diffusion | Qualitative | Random Forest | Accuracy = 0.99 AUC = 0.99 Sensitivity = 1.00 Specificity = 0.92 |
| Sharif et al. (2020) (35) | HGG vs. LGG (unclear) | BraTS (n=30 patients) | 2.3:1 (7 HGG, 3 LGG in testing set) | Internal (Holdout, 10-fold cross-validation) | T1, T1CE, T2, FLAIR | Deep learning extracted | Convolutional Neural Network | Accuracy = 0.987 |
| Muneer et al. (2019) (36) | 1 vs. 2 vs. 3 vs. 4 | Single center hospital (n=20 patients) | 1.3:1.6:1:1.5 (39 grade 1, 51 grade 2, 31 grade 3, 47 grade 4 images in testing set) | Internal (Holdout, 30% of dataset) |
T2 | Deep learning extracted | VGG19 (Deep Convolutional Neural Network) | Accuracy = 0.9825 Sensitivity = 0.9272 Specificity = 0.9813 Positive Predictive Value = 0.9471 F1 Score = 0.9371 |
| Dandil and Bicer (2020) (37) | 1/2 vs. 3 vs. 4 vs. meningioma | INTERPRET (n=179 patients) | Unclear | Unspecified | MR Spectroscopy (Time of Echo 20ms and 136ms) | First-order, Shape and size, Texture | Long Short-Term Memory (Neural Network) | Accuracy = 0.982 AUC = 0.9936 Sensitivity = 1.00 Specificity = 0.9753 |
| Tian et al. (2018) (38) | 2 vs. 3/4 | Single center hospital (n=153 patients) | 2.6:1 (111 HGG, 42 LGG in total set) | Internal (10-fold cross-validation) | T1, T1CE, T2, Diffusion, Perfusion (3D Arterial Spin Labeling) | Texture | Support Vector Machine | Accuracy = 0.981 AUC = 0.992 Sensitivity = 0.987 Specificity = 0.974 |
| Lo et al. (2019) (39) | 2 vs. 3 vs. 4 | TCIA (n=130 patients) | 1:1.4:1.9(30 grade 2, 43 grade 3 and 57 grade 4 in total set) | Internal (10-fold cross-validation) | T1CE | Deep learning extracted | Deep Convolutional Neural Network | Accuracy = 0.979 AUC = 0.9991 |
| Kumar et al. (2020) (40) | 1/2 vs. 3/4 | BraTS (n=285 patients) | 2.8:1 (210 HGG, 75 LGG in total set) | Internal (5-fold cross-validation) | T1, T1CE, T2, (T2W)-FLAIR | First-order, Shape, Texture | Random Forest | Accuracy = 0.9754 AUC = 0.9748 Sensitivity = 0.9762 Specificity = 0.9733 F1 Score = 0.983 |
Testing or validation metrics are reported when available, otherwise training metrics are reported. HGG, high-grade gliomas; LGG, low-grade gliomas; ML, machine learning; PRISMA-DTA, Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy; T1CE, T1-weighted contrast-enhanced; TRIPOD, Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis.