Table 2 Prognostic applications.
From: Radiomics and radiogenomics in gliomas: a contemporary update
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.87 | 115 (GBM) | T1WI, T2WI, FLAIR | Manual 2D | Matlab | 30 features: Laws, Gabor, GLCM | Laws energy (R5R5, E5E5, S5S5) from enhancing tumour and oedema | RF model | Features from the enhancing and oedematous regions were predictive of the extent of hypoxia. Features on the validation set were also found to be prognostic of overall survival |
Kickingereder et al.88 | 119 (GBM) | 3 T/T1-CE, FLAIR | Manual 3D | Medical Imaging Toolkita | Histogram volume and shape features; texture features; wavelet analysis | 11 features of 12,190 features | SPC analysis | SPC analysis performed better than clinical (age and Karnofsky Performance Score) and radiologic (rCBV and ADC) parameters |
Molina et al.128 | 79 (GBM) | 1.5–3 T/T1-CE | Manual 3D | Matlab | Five GLCM features, 11 GLRLM features | GLCM (entropy, homogeneity, contrast, dissimilarity) GLRLM (LRE, HGLRE, LRHGE, RPC) | Kaplan–Meier curves and Cox-proportional hazards analysis | Patients had better prognosis when high LRHGE, low RPC, low entropy, high homogeneity, and low dissimilarity were present (P < 0.05) |
Chaddad et al.129 | 73 (GBM)/ TCIA | T1-CE, FLAIR | Semi-automatic 2D | – | GLCM JIM | JIM (entropy, inverse-variance) in necrosis region and (entropy, contrast) in oedema region | RF model | JIM of T1-CE and FLAIR images significantly predicted survival outcomes with moderate correlation; nine features were found to be associated with glioblastoma survival, P < 0.05, accuracy of 68%-70%; AUC of 77.56% with P = 0.003 was achieved when combining JIM, GLCM, and gene expression features into a single radiogenomic signature. |
Liu et al.130 | 119 (GBM)/TCGA | T1WI, T2WI, FLAIR, T1-CE | Manual 2D | – | Histogram, GLCM, GLRLM | 13 textural features | SVM-RFE | T1-CE sequence performed best, with AUC of 0.7915 and accuracy of 80.67% |
Yang et al.131 | 82 (GBM)/TCGA | T1-CE, T2WI, FLAIR | Manual 3D | Matlab | SFTA, RLM, LBP, HOG, Haralick texture features | – | RF model | Molecular subtypes and 12-month survival was predicted by several features; SFTA features on T1-CE were most predictive of survival and proneural subtype (AUC = 0.82); RLM features on T2 FLAIR axial for neural (AUC = 0.75); Haralick features on T1-CE for classic (AUC = 0.72); HOG features on T2 FLAIR axial for mesenchymal (AUC = 0.70) |
Bahrami et al.95 | 33 (HGG) | 3 T/FLAIR | Semi-automatic 3D | – | EC | – | Kaplan–Meier curves | Reported that lower EC of the FLAIR hyperintense region was associated with poor PFS (P = 0.009) and OS (P = 0.022) status post-bevacizumab therapy |
Jain et al.89 | 45 (GBM) | 1.5/3 T, T1, T2w/perfusion, FLAIR | Manual 2D | Rb | VASARI features, rCBV statistics | rCBVNER | Random Survival Forest | Increased maximum rCBVCER found to be associated with increased risk of death |
McGarry et al.90 | 81 (GBM) | 1.5/3 T, T1, CE-T1, T2, FLAIR | Semi-automatic 3D | – | 81 radiomic profiles (5 RPs): 4-digit code assigned to each voxel representing the intensity-based segmentation | Five RPs correlated with survival when thresholded by volume | Cox Regression | Presented a method for creating radiographic profiles by combining intensity information from multiple MRI scans. Pathologically validated that voxels indicated by one of the RPs contained hypercellular tumour and necrosis. |
Lao et al.117 | 112 (GBM) | 1.5/3 T, T1, T1-CE, T2, FLAIR | Manual 2D | Python CAFFE | 1403 handcrafted (HC) features: geometry, intensity, texture and 98304 deep features using transfer learning | 150 HC and deep features | LASSO Cox Regression | The radiomics signature achieved a C-Index of 0.731 for the discovery dataset, and 0.710 for the independent validation set. |
Li et al.132 | 92 (GBM)/TCIA, local data | T1, T1-CE, T2, FLAIR | Automated using Matlab | Rb | Texture at different voxel size, quantization and grey levels | - | Cox regression model | The multiparametric signature achieved better performance for OS prediction (C-Index = 0.705) |
Sun et al.133 | 542 (training n = 285; validation =66; test=191; LGG,HGG)/BRATS 2018 | T1, T1-CE, T2, FLAIR | 3D CNN architectures | Pyradiomics toolbox | Shape, texture, first-order statistics | Age, 14 selected features | RF regression | Achieved 61% accuracy in predicting survival outcome. |
Sanghani et al.134 | 163 (GBM)/BRATS 2017 | T1, T1-CE, T2, FLAIR | Segmentation masks from BraTS 2017 dataset | – | Texture, shape, volumetric, age | Top 150 selected features out of 2200 | SVM-RFE/fivefold cross-validation | The 2-class and 3-class OS group prediction accuracy obtained were 98.7% and 88.95% respectively. |
Beig et al.135 | 460 (GBM), Male = 290, female = 170 | T1, T1-CE, T2, FLAIR | Semi-automatically with Slicer 3Dd | Matlab | 2850 features: Laws, Gabor, shape based | Five Laws energy, two Gabor wavelets; one shape based | Cox regression model | Sexually dimorphic radiomic risk score (RRS) models that are prognostic of overall survival (OS) in primary GBM |
Han et al.136 | 178 (HGG)/TCIA, local dataset | T1-CE | Manual segmentation | Matlab, Kerasc, R | 348 handcrafted - volume, size, texture, intensity, first-order statistical + 8192 deep CNN features | – | Elastic net-Cox modelling | The combined feature analysis framework classified the patients into long- and short-term survivor groups with a log-rank test P < 0.001 |
