Table 1.
Diagnostic applications.
| Author | No. of patients | Magnet strength/MRI Sequences | Segmentation manual vs automatic |
Software | Type of radiomic analysis | Best discriminating features | Machine learning/statistical approach | Results |
|---|---|---|---|---|---|---|---|---|
| Ismail et al.36 | 105 (Training n = 59, test n = 46; GBM) | 1.5 T/T1-CE, T2WI/FLAIR | Manual 2D | Matlaba | 30 shape features; 14 “global” contour characteristics and 16 “local” curvature |
Top two most discriminative features including the SD of S and the mean of KT |
SVM | 3D shape attributes from the lesion habitat can be differentially expressed across pseudoprogression and tumour progression and could be used to distinguish these radiographically similar pathologies |
| Larroza et al.74 | 115 (RN = 32, radiation treated mets = 23, untreated mets = 60) | T1-CE | Manual 2D | Mazda | 179 features; histogram, gradient, GLCM, GRLM, wavelets | Intensity 90th percentile | SVM | High classification accuracy (AUC > 0.9) was obtained using texture features and a support vector machine classifier to differentiate between brain metastasis and RN |
| Prasanna et al.81 | 75 (different grades of RN) | T1-CE | Manual 2D | Matlab | Four CoLlAGe entropy | Collage entropy skewness | RF | COLLAGE features exhibited decreased skewness for patients with pure and predominant RN and were statistically significantly different from those in patients with predominant recurrent tumours. |
| Hu et al.80 | 31 (RT = 15, RN = 16) | T1-CE, T2, FLAIR, PD, ADC, rCBF, rCBV and MTT | Manual 2D | – | Eight parameters derived from the multiple MR sequences: CE-T1, T2, FLAIR, PD, ADC, rCBF, rCBV and MTT. | rCBV | OC-SVMs | Greater value of advanced MRI DWI and PWI derived measures as compared to conventional imaging for discrimination of RN from viable tumour |
| Tiwari et al.26 | 58 (training n = 43, test n = 15; GBM) | T1-CE, T2, FLAIR | Manual 2D | Matlab | 119 texture maps: Haralick, Laws, Laplacian pyramid, histogram of gradient orientations | GLCM and Laws features in the lower Laplacian scale | SVM | Laplacian pyramid features were identified to be most discriminative, possibly because these emphasize edge-related differences between RT and RN at lower resolutions. |
| Skogen et al.58 | 95 (HGG = 68, LGG = 27) | 3 T/T1-CE | Manual 2D | TexRADb | Histogram metric | Fine textures scale | – | LGGs and HGGs were best discriminated using SD at fine-texture scale, with a sensitivity and specificity of 93% and 81% (AUC = 0.910, P < 0.0001) |
| Xie et al.60 | 42 (HGG = 27, LGG = 15) | 3 T/dynamic contrast-enhanced | Manual 2D | OmniKineticsc | Five GLCM features - energy, entropy, inertia, correlation, and inverse difference moment (IDM) | Entropy and IDM | – | Evaluated five GLCM features from (DCE)-MRI of 42 patients with gliomas. They reported that entropy (AUC = 0.885) and IDM (AUC = 0.901) were able to differentiate grade III from grade IV and grade II from grade III gliomas, respectively.; no feature was able to distinguish subtypes of grade II and grade III gliomas. |
| Qi et al[.126 | 39 (HGG = 26, LGG = 13) | 3 T/DWI/DKI | Manual 2D | ImageJd | Histogram metric | Mean kurtosis (MK) | – | Histogram parameters on DKI were significant in differentiating high- (grade III and IV) from low-grade (II) gliomas (P < .05); mean kurtosis was the best independent predictor of differentiating glioma grades with AUC = 0.925 |
| Tian et al.59 | 153 (Grade II = 42, III = 33, IV = 78) |
3 T/Multiparametric (T1WI, T1-CE, T2WI, DWI, ASL) |
Manual 2D-VOI | Matlab |
GLCM, GLGCM histogram mean |
30 and 28 Optimal features of 420 texture and 90 histogram features | SVM-RFE | Texture features were statistically significant over histogram parameters for glioma grading; AUC for classifying LGGs versus HGGs was 0.987, while it was 0.992 for grade III versus IV gliomas |
| Zacharaki et al.96 | 102 (mets = 24, meningiomas = 4, grade II gliomas = 22, grade III gliomas = 18, GBMs = 34) | 3 T (T1, T2, FLAIR, DTI, perfusion) | Manual 2D | – | 161: tumour shape, image intensity, Gabor texture | Different features for each pair-wise classification task, mainly comprising intensity from T1, T2, rCBV statistics and Gabor texture from FLAIR | LDA, kNN, SVM | The binary SVM classification achieved via a leave-one-out cross-validation reported accuracy, sensitivity, and specificity of 85%, 87%, and 97% for discrimination of metastases from gliomas and 88%, 85%, and 96% for discrimination of high-grade (grade III and IV) from low-grade (grade II) neoplasms. |
| Suh et al.61 | 77 (GBM = 23, PCNSLs = 54) | 3 T/T1-CE, T2, FLAIR | Manual 2D | Python package, PyRadiomics 1.2.0e | Shape, volume, 1st order, GLCM, GLRLM, mGLSZM, and wavelet transform | A total of 6366 radiomics features subjected to recursive feature elimination and random forest analysis with nested cross-validation | SVM | In comparing diagnostic performances, the AUC (0.921) and accuracy (90%) of the radiomics classifier was significantly higher than those of the 3 radiologists (P < 0.001) |
| Beig et al.127 | medulloblastomas (n = 22), ependymomas (n = 12), and gliomas (n = 25) | T1, T2, FLAIR | Manual 2D | Matlab | 52 CoLlAGe features | sum variance and entropy of CoLlAGe on T2 | RF | Medulloblastomas exhibited higher CoLlAGe entropy values than ependymomas and Gliomas for the paediatric brain tumour cases. |
OC one class, SVM support vector machine, RFE recursive feature elimination, RF Random Forest, rCBV relative cerebral blood volume, RLM run-length matrix, ADC apparent diffusion coefficient, EC edge contrast, – not available, GLSZM grey-level size-zone matrix, GLCM grey level co-occurrence matrix, GLRLM grey level run-length matrix features, LDA linear discriminant analysis, S sharpness, KT measure of the total curvature, VOI volume of interest, kNN k-nearest neighbours.
aMathWorks, Natick, Massachusetts.
bhttps://imagingendpoints.com/texrad-software/.
chttp://www.omnikinetics.com/.
dNational Institutes of Health, Bethesda, Maryland.
ehttps://pyradiomics.readthedocs.io/en/latest/modules/radiomics/ngtdm.html.