Table 3.
Applications in selecting optimal therapy.
| Author | No. of patients | Magnet strength/MRI sequence | Segmentation manual vs automatic | Texture software | Type of radiomic analysis | Best discriminating features | Machine learning/statistical approach | Results |
|---|---|---|---|---|---|---|---|---|
| Beig et al.137 | 83 LGGs | T2w/FLAIR | Manual 2D | Matlab | GLCM, Gabor | 3D Gabor | SVM | Initial results indicate that radiomic features from non-enhancing regions on T2 and infiltrative edges on FLAIR can segregate the 3 subgroups. |
| Zhang et al.63 | 152 (IDH mutant = 92, wild-type = 60) | 1.5T–3 T/T1-CE, T2WI, FLAIR | Manual 3D | Matlab | GLCM GLGCM | 15 Optimal features from 168 Haralick features | SVM-RFE | AUC 0.841 and accuracy of 82.2% for non-invasively discriminating IDH mutation of patient with glioma |
| Hsieh et al.138 | 39 (IDH mutant = 7, wild-type = 32; TCGA) | 1.5T–3 T/T1-CE | Manual 2D | CAD system | Morphological intensity-GLCM | 14 GLCM | Binary logistic regression classifier | Textural features describing local patterns yielded an accuracy of 85% in detecting the IDH status |
| Han et al.64 | 42 (IDH mutant = 21, wild-type = 21) | 3 T/T1WI, T2WI, 3D-T1-CE | Manual 3D | OmniKinetics | 29 Texture features from first-order and GLCM | Inertia, cluster prominence, GLCM entropy | – | Showed joint variables derived from T1WI, T2WI, and contrast-enhanced T1WI imaging histograms and GLCM features could be used to detect IDH1-mutated gliomas. The AUC of Joint VariableT1WI+C for predicting IDH1 mutation was 0.984, and the AUC of Joint VariableT1WI for predicting the IDH1 mutation was 0.927 |
| Jakola et al.65 | 25 (IDH mutant = 20, wild-type = 5) | 3 T/3D-FLAIR | Semiautomatic 3D | ImageJ | Homogeneity, energy, entropy, correlation, inertia | Homogeneity | – | Homogeneity discriminated patients with LGG in IDH mutant and IDH wild-type (P = 0.005), AUC for combined parameters was 0.940 for predicting IDH mutation; authors could not separate IDH-mutant tumours on basis of 1p/19q-codeletion status |
| Bahrami et al.66 | 61 (IDH mutant = 43, wild-type = 11); 7 unknown) | 3 T/Pre- and post-T1-CE, FLAIR | Semiautomatic 3D | 3D-co-occurrence matrix | Histogram, GLCM | Homogeneity, pixel correlation, EC | Logistic regression with LASSO regularization | Greater signal heterogeneity and lower EC noted in IDH wild-type tumours; IDH mutant tumours with 1p/19q-codeleted status; lower EC in MGMT-methylated tumours |
| Shofty et al.139 | 47 LGGs | 1.5T–3 T/FLAIR, T2, TI-CE | Automatic using FSLa, 3D | Matlab | Histogram, contrast, correlation, energy, entropy, homogeneity | 39 of 152 textural features | 17 classifiers | Ensemble of bagged trees classifier achieved the best performance (Accuracy = 87%; AUC = 0.87) for the detection of 1p/19q codeletion; majority of differences detected for T2 and T1-CE |
| Kickingereder et al.94 | 172 (GBM) | 3 T/Pre- and post-T1-CE, FLAIR | Semiautomatic 3D | Medical imaging Toolkitb | 188 imaging features, 17 first-order features (FO), 9 volume and shape features (VSF) and 162 texture features (GLCM,) GLRLM). | – | Supervised principal component (superpc) analysis | The superpc predictor stratified patients in the validated set into a low or high-risk group for PFS (HR = 1.85, P = 0.030) and OS (HR = 2.60, P = 0.001). |
| Grossmann et al.100 | 126 (GBM) |
Baseline and follow-up MRI (1 and 6 wks), T1WI, T2WI, FLAIR, T1-CE |
Semi-automatically with Slicer 3Dd |
R version 3.1.0c |
First-order statistics of the voxel intensity histogram; tumour shape; tumour texture |
Information correlation | PCA | Radiomics provides prognostic value for survival and progression in patients with recurrent glioblastoma receiving bevacizumab treatment; features derived from postcontrast T1WI yielded higher prognostic power compared with T2WI |
| Wu et al.140 | 126 (grade II-III = 43 and grade IV = 83)/TCIA | T1, T1-CE, T2, FLAIR | Semiautomatic approach | R version 3.3.1c | GLCM texture, Volume, intensity, histogram, diffusion | 20 of 704 radiomic features | SVM, kNN, RF, NB, NN, FDA, Adaboost/tenfold cross-validation | Random Forest (RF) showed high predictive performance for identifying IDH genotype (AUC = 0.931, accuracy=88.5%) |
| Zhou et al.141 | 744 (LGG, HGG)/ TCIA, local datasets | T1-CE, T2-FLAIR | Semi-automatically with Slicer 3Dd | Matlab | Histogram, texture, age, shape | Top 15 features out of 127 | RF/Train-test model | The overall accuracy for 3 group prediction (IDH-wild type, IDH-mutant and 1p19q co-deletion, IDH-mutant and 1p19q non-codeletion) was 78.2% |
| Lee et al.142 | 123 (GBM) | T1, T2, T1-CE, FLAIR, PWI, DWI | Manual segmentation | Nordi-cICE | Volume, ADC map, CBV | Four of 31 radiological features | kNN, SVM, RF, Adaboost, decision tree, NB, LDA, gradient boosting | Prediction rate of IDH1 mutation status with 66.3–83.4% accuracy |
| Sudre et al.143 | 333 (IDH mutant = 151, wild-type = 182); | 1.5 T/T1, T2, FLAIR, DSC MRI | Olea Sphere, Version 3 | – | Shape, intensity, texture | Nine histogram features, 11 texture features | RF/cross-validation | Gliomas were correctly stratified 53% for grade classification and 71% for IDH classification |
GLCM grey level co-occurrence matrix, GLRLM grey level run-length matrix features, LDA linear discriminant analysis, PCA principal component analysis, Adaboost adaptive boosting, NB Naive Bayes, FDA flexible discriminant analysis, NN neural network.
aFSL (http://www.fmrib.ox.ac.uk/fsl).
bhttp://www.mitk.org/wiki/The_Medical_Imaging_Interaction_Toolkit_(MITK).
cR statistical and computing software (http://www.r-project.org).