Table 2. Recent studies applying machine learning to the development of neuro-oncology prognostic biomarkers.
| Author(s) | Dataset | Method | Results |
|---|---|---|---|
| Choi et al., 201560 | 61 preoperative DCE | Retrospective Multivariate Cox regression using MRI, pharmacokinetic, & clinical parameters |
C-index: 0.82 |
| Kickingereder et al., 201661 | 119 (training = 79 & testing = 40) T1, T1 C, FLAIR, DWI, DSC | Retrospective Supervised principal component analysis with Cox regression analysis |
C-index: 0.70 |
| Chang et al., 201662 | 126 (training = 84 & testing = 42) patients T1, T2, FLAIR, T1 C, DWI | Retrospective Random forest on radiomic features (including Laws, Haralick) |
Accuracy: 76% |
| Liu et al., 201663 | 147 rs-fMRI and DTI | Retrospective SVM using clinical features & network features of structural & functional network |
Accuracy: 75% |
| Nie et al., 201664 | 69 T1 C, rs-fMRI, DTI | Prospective SVM using supervised CNN-derived features |
Accuracy: 89.9% Sensitivity: 96.9% Specificity: 83.8% PPR: 84.9% NPR: 93.9% |
| Macyszyn et al., 201651 | 134 (training = 105 & testing = 29) T1, T1 C, T2, FLAIR, DTI, DSC | Prospective SVM for OS <6 months & SVM for OS <18 months |
Accuracy (<6 months): 82.76% Accuracy (<18 months): 83.33% Accuracy (combined): 79% |
| Zhou et al., 201765 | 32 TCGA T1 C, FLAIR, T2 & 22 T1 C, FLAIR, T2 | Retrospective Group difference features to quantify habitat variation Supervised forward feature ranking with SVM |
Accuracy: 87.5%, 86.4% |
| Dehkordi et al., 201766 | 33 pre-treatment DCE | Retrospective Adaptive neural network with fuzzy inference system using Ktrans, Kep and ve |
Accuracy: 84.8% |
| Lao et al., 201767 | 112 (training = 75 & testing = 37) pretreatment T1, T1 C, T2, FLAIR | Retrospective Multivariate Cox regression analysis using radiomic features as well as "deep features" from pre-trained CNN |
C-index: 0.71 |
| Liu et al., 201768 | 133 T1 C | Retrospective Recursive feature selection with SVM |
Accuracy: 78.2% AUC: 0.81 Sensitivity: 79.1% Specificity: 77.3% |
| Li et al., 201769 | 92 (training = 60, testing = 32) T1, T1 C, T2, FLAIR. TCGA data used. |
Retrospective Random forest for segmentation into 5 classes Multivariate LASSO-Cox regression model |
C-index: 0.71 |
| Chato & Latifi, 201752 | 163 T1, T1 C, T2, FLAIR. Short-, mid-, long-term survivors | Retrospective SVM, KNN, linear discriminant, tree, ensemble & logistic regression applied to volumetric, statistical & intensity texture, histograms & deep features |
Accuracy: 91% Linear discriminant using deep features |
| Ingrisch et al., 201770 | 66 T1 C | Retrospective Random survival forests using 208 global & local features from segmented tumour |
C-index: 0.67 |
| Li et al., 201771 | 92 (training = 60 & testing = 32) T1, T1 C, T2, FLAIR. TCGA data used. |
Retrospective LASSO Cox regression to define radiomics signature |
C-index: 0.71 |
| Bharath et al., 201772 | 63 TCGA preoperative: T1 C, FLAIR | Retrospective LASSO Cox regression using age, KPS, DDIT3 & 11 principal component shape coefficients |
C-index: 0.86 |
| Shboul et al., 201773 | 163 T1, T1 C, T2, FLAIR | Retrospective Recursive feature selection & random forest regression |
Accuracy: 63% |
| Peeken et al., 201874 | 189 T1, T1 C, T2, FLAIR & clinical data. | Retrospective Multivariate Cox regression using VASARI features and clinical data |
C-index: 0.69 |
| Kickingereder et al., 201875 | 181 (training = 120 & testing = 61) pretreatment MRI | Retrospective Penalised Cox model for radiomic signature construction |
C-index: 0.77 |
| Chaddad et al., 201876 | 40 (training = 20 & testing = 20) preoperative MRI, T1 & FLAIR. | Retrospective Random forest on multi-scale texture features |
AUC: 74.4% |
| Bae et al., 201877 | 217 (training = 163 & testing 54) preoperative MRI, T1 C, T2, FLAIR, DWI | Retrospective Variable hunting algorithm for selection & random forest classifier |
iAUC: 0.65 |
TCGA, The Cancer Genome Atlas; T1 C, post contrast T1-weighted; SVM, support vector machine; DCE, dynamic contrast-enhanced imaging; CNN, convolutional neural network; KNN, k-nearest neighbours/rs-fMRI, resting state functional MRI; KPS, Karnofsky performance status; DDIT3, DNA damage inducible transcript 3; DTI, diffusor tensor imaging; DSC, dynamic susceptibility weighted; OS, overall survival.