Table 1 Diagnostic markers

IDH Mutation

Kasap et al. 2024 [15]

106

T1-w, CET1-w T2-w and FLAIR sequence

Semi-automatic

^a 3D slicer

A total of 107 radiomic features were extracted.

age_at_diagnosis, original_shape_Flatness

Radiomics

Results show that CE-T1W images are most optimal to predict IDH mutation status

^b PyRadiomics

original_gldm_LargeDependence, HighGrayLevelEmphasis

Hosseini et al. 2023 [8]

57

T1- MPRAGE, Axial T2 FLAIR, CE T2W

Manual

^b PyRadiomics

A total of 105 original radiomic features from categories (shape, first-order statistical, second-order texture, and higher-order statistic) were extracted

-

Radiomics

Best discriminatory performance (AUC = 0.93, ACC = 0.92) obtained from solid/contrast enhancing, and core tumor overlaid on post-contrast T1-weighted images

Machine learning

Deep Learning

Liu et al. 2023 [80]

205

T1C, T2, T1 FLAIR, and T2 FLAIR

Manual

^c ITK-SNAP

A total of 428 radiomic features (107 from each sequence) were extracted.

GLRLM and contrast features

Radiomics

The LR classifier achieved the best results in predicting the IDH mutation status with an AUC of 0.8572

Calabrese et al. 2022 [6]

199

T1 pre, T1 post, T2, T2/FLAIR, SWI, DWI, ASL, MD, AD, RD, and FA

Automated deep learning-based tumor segmentation followed by Manual correction

^b PyRadiomics

Default set of shape features (n = 26), first-order grayscale features (n = 19), and higher-order grayscale features (n = 75) were extracted yielding 5300 radiomics features per patient

-

Radiomics, Deep Learning

AUC/Sens/Spec of 0.96/1.00/0.83 for predicting IDH1 mutation status

^c ITK-SNAP v3.8.0

^d TensorFlow 2.4

He et al. 2022 [16]

108 (From TCGA-LGG dataset)

T1, T2, FLAIR, and T1Gd

Manual

^e FeAture Explorer (FAE, Version 0.5.2)

“6-Step” general radiomics model to noninvasively predict the IDH mutation status by simultaneously tuning combined multi-sequence MRI proposed

-

Radiomics

Optimal radiomics pipeline for predicting IDH mutation status was the T2 + FLAIR combined multi-sequence MRI with an AUC of 0.873 ± 0.05

Kawaguchi et al. 2021 [81]

TCIA dataset (n = 159), NCCH Japan (n = 166)

T1, T1-CE, T2, FLAIR

Manual

^b PyRadiomics

A total of 16,221 features generated for each patient, using both images and clinical records. Inception-ResNet v2 architecture used

-

Deep learning and Machine learning

AUROC of 0.867 achieved for IDH prediction on TCIA dataset, but decreased to 0.383/0.392 in the validation/test NCC cohort

S. Kihira et al. 2021 [82]

111 (Training n = 91, test n = 20)

T2, FLAIR, T1CE, DWI

Manual 2D

^f Olea Sphere software

Total of 92 radiomics features assessed

Conventional: FLAIR – GLCM Informal Measurement Correlation 2. T1c+ - First order skewness, GLCM Difference variance, GLSZM Small Area High Gray, GLCM Dependence Variance

Machine learning

Predictive model for IDH1 status based on conventional MR imaging achieved area under the curve (AUC) of 0.95. Upon incorporating diffusion data, combination of 5 conventional and 5 diffusion MR features remained as significant contributors, resulting in a perfect AUC of 1.0.

Diffusion: B1000: First Order Skewness

ADC: First Order Skewness, GLRLM Run Length Non-Uniformity, GLSZMGL Non-Uniformity Norm, GLSZM Small Area High Gray

Kim et al. 2020 [83]

155

T2W, FLAIR, T1W, DWI, CET1WI, and DSC. DWI

Manual

^g MITK software

A total of 6472 radiomic features extracted and analysed

-

Radiomics

Multipara metric MR radiomics achieved an AUC of 0.795/0.747 on the training/validation sets in predicting IDH mutation.

^h Matlab

Niu et al. 2020 [84]

182

T1CE

Manual 2D

ⁱ AK software (Analysis Kit, GE Healthcare)

396 features extracted including histogram & texture (GLCM, RLM, GLSZM) parameters, and form factor parameters.

Volume CC

Radiomics

The model exhibited good discriminatory performance in both the primary dataset (AUC = 0.87, ACC = 0.798, sensitivity = 85.5%, specificity = 75.4%) and the validation dataset (AUC = 0.86, ACC = 0.789, sensitivity = 91.3%, specificity = 69.0%).

Intensity variability

Machine learning

Short run emphasis_angle 90_offset 4

Park CJ et al. 2020 [85]

168

DTI, T1CE, T2, FLAIR

Manual 2D

^j Medical Image Processing, Analysis, and Visualization software package version 7.0

A total of 158 and 253 radiomic features were extracted from DTI and conventional MRI respectively.

-

Radiomics

The combined model incorporating DTI and conventional radiomics demonstrated significantly superior performance compared to the model comprising only DTI histogram parameters and conventional radiomics (AUC: 0.900 vs 0.869, p = 0.040).

^b PyRadiomics

Machine learning

Peng et al. 2020 [86]

105 (Training n = 73, test n = 32)

T1CE, T2, ASL

Manual 2D

^b PyRadiomics

A total of 851 radiomics features extracted from each VOI

-

Radiomics

The classifier, which integrated features from all three sequences, achieved an accuracy of 0.823 and an AUC of 0.770 (P < 0.05).

Machine learning

Sakai et al. 2020 [87]

100 (Training n = 60, validation n = 20, test n = 20)

DWI, FLAIR

Manual 2D

^f Olea Sphere software

A total of 184 radiomic features extracted. The DWI model utilized 71 out of the 92 features, while FLAIR model utilized 33 out of the initial 92 features.

-

Radiomics

The best performance achieved was with an AUC of 95%, Accuracy of 90%, Precision/Recall/f1-score of 94%/94%/94% for IDH1 wildtype, and 75%/75%/75% for IDH1 mutants

Machine learning

Han et al. 2020 [88]

59

Conventional MRI, contrast-enhanced MRI, and APTW imaging

Manual

^c ITK-SNAP, Non-commercial Analysis-Kit software (GE Healthcare, China)

A total of 1038 features including 8 first-order histograms, 6 GLCM and 4 GLRLM extracted and analyzed

Run Length Nonuniformity angle0 offset1

Machine learning

SVM model achieved an AUC of 0.952 and 0.84 in the training set and test set, respectively. Efficacy achieved by SVM model superior to that of univariate analysis.

Correlation All Direction offset4 SD

Fukuma et al. 2019 [89]

164

T1W, T2W, FLAIR, and GdT1W

Manual

^h MATLAB-based image analysis software

Combination of CNN (AlexNet) and Conventional radiomics based approach used for analysis

-

Deep learning

Using the combination of conventional radiomic features and/or patient age, CNN an accuracy of 73.1% achieved in predicting IDH status.

Ren et al. 2019 [90]

57

3D-ASL, T2, T2 FLAIR, DWI

Manual

Custom developed software in Matlab

265 high-throughput radiomic features were extracted on each tumor volume of interest

eADC: short run emphasis (GLRLM), energy (GLGCM), long-run emphasis (GLRLM), energy (GLCM)

Radiomics

The accuracies/AUCs/sensitivity/specificity/PPV/NPV of predicting IDH1(+) in LGG were 94.74%/0.931/100%/85.71%/92.31%/100%

^k Advantage Workstation 4.6, GE Medical Systems

Machine learning

Wu et al. 2019 [91]

126

T1, T1CE, T2, T2 FLAIR

Manual 2D

^l R Software

A total of 704 radiomic features extracted

-

Radiomics

Random Forest (RF) exhibited excellent predictive performance, with accuracy of 0.885 ± 0.041 and AUC of 0.931 ± 0.036. In contrast, neural network (NN) (accuracy 0.829 ± 0.064, AUC 0.878 ± 0.052) and flexible discriminant analysis (FDA) (accuracy 0.851 ± 0.049, AUC 0.875 ± 0.057) displayed comparatively lower predictive performance.

Machine learning

Arita et al. 2018 [92]

199

Conventional MRI, ADC, normalized blood volume

Manual

In-house-developed image analyzing software (Developed in Matlab)

A total of 109 radiomic features quantified and collected

Frontal lobe tumor involvement (MNI_str_loc.04) for IDH Mutant

Machine learning

IDH mutation predicted with an accuracy of 0.85 to 0.87, which improved by implementing lesion location information.

^m JMP Pro ver.13

Magnitude of contrast enhancement (Gdzscore_ara.of.Gd.) for IDH wildtype

Chang et al. 2018 [93]

496 (Divided in 3 cohorts)

MRI: T1, T1-CE, T2, FLAIR

Manual

ⁿ Matrix User v2.2

-

-

Deep learning

IDH prediction accuracies of 82.8% (AUC = 0.90), 83.0% (AUC = 0.93), and 85.7% (AUC = 0.94) achieved on the training, validation, and testing sets, respectively.

^a 3D Slicer software (v4.6)

Liang et al. 2018 [94]

167

MRI: T1,T1-CE,T2, FLAIR

Manual

^o MXNet (version 1.0, Apache Software Foundation)

Multimodal Three-Dimensional DenseNet constructed for the analysis

-

Deep learning

Accuracy of 84.6% achieved on the validation dataset. To evaluate the generalizability, transfer learning techniques applied to predict WHO grade status, yielding a high accuracy of 91.4%on validation dataset.

Li et al. 2017 [95]

151

MRI (T1c, T2 FLAIR)

Auto segmentation Manual

^p Brainsuite

A total of 671 image features extracted and replaced with 16, 384 CNN features

-

Deep learning based Radiomics

The radiomics method achieved an area under the operating characteristic curve (AUC) of 86% for estimating IDH1, while the AUC for DLR was higher at 92%. Integrating multiple-modality MR images and utilizing DLR further improved the AUC for IDH1 estimation to 95%.

Yu et al. 2017 [96]

110

T2 FLAIR

Auto Segmentation (Using CNN)

Total of 671 high-throughput features were extracted and quantized.

Shape; Sphericity

Radiomics

In the Leave-One-Out Cross Validation (LOOCV) analysis, IDH1 status achieved an estimation accuracy of 0.80, with a sensitivity of 0.83 and specificity of 0.74. AUC reached 0.86, indicating the promising discriminatory capability of the approach in accurately predicting IDH1 status.

Texture: Large zone high grey-level emphasis

Machine learning

Wavelet feature: small zone high gray-level emphasis

Deep learning

TERT promoter

Chen et al. 2023 [97]

143

Axial T2W, DWI, and ADC

Manual

^a 3D slicer software

A total of 2553 features extracted

ADC entropy

Radiomics

Model constructed based on the RFE and LDA achieved the best diagnostic performance (AUC, accuracy, sensitivity, and specificity: 0.964, 0.940, 0.891, and 0.982, respectively) in predicting TERT p mutation

Huo et al. 2023 [98]

109

Axial T2W, CE-T1W, T1W and CE-T1W

Manual

^q MRIcron

A Total of 2608 radiomic features extracted for each patient

Wavelet-HHH_glcm_Idmn wavelet-HHH_glcm_Idn exponential_glszm_GrayLevelNonUniformityNormalized wavelet-HLL_glszm_LowGrayLevlZoneEmphasis.

Radiomics

Fusion radiomic model with 4 radiomic features achieved an AUC value of 0.876 and 0.845 in the training and validation set respectively for predicting TERT p mutation in IDH wildtype gliomas.

^b Pyradiomics

Wang et al. 2023 [99]

140 (With independent validation on 34 separate cases)

T1W, T2W, T1CE, FLAIR, and ADC maps

Auto segmentation and Manual correction

^r Niftynet

A total of 3654 radiomic features extracted and analysed

FLAIR_wavelet-LLL_gldm_DependenceVariance

Radiomics

AUROC/sensitivity/specificity 0.952/0.714 /0.963 achieved in the independent validation set in identifying IDHmut pTERTmut tumors

^c ITK-snap

ADC_log-sigma-5-0-mm-3D_firstorder_RootMeanSquared

^b Pyradiomics version 2.2.0

Visul feature: IDHmut pTERTmut gliomas showed homogenous low-complexity texture

Zhang et al. 2023 [19]

274 (training n = 156, validation n = 118)

T1CE, T1WI, T2WI

-

-

-

-

Deep learning

DLR signature showed the best discriminative power for predicting TERT promoter mutations, yielding an AUC of 0.990 and 0.890 in the training and external validation cohorts.

Calabrese et al. 2022 [6]

199

T1 pre, T1 post, T2, T2/FLAIR, SWI, DWI, ASL, MD, AD, RD, and FA

Automated deep learning-based tumor segmentation followed by Manual correction

^b Pyradiomics ver2.2

Default set of shape features (n = 26), first-order grayscale features (n = 19), and higher-order grayscale features (n = 75) were extracted yielding 5300 radiomics features per patient

-

Radiomics, Deep Learning

AUC/Sens/Spec of 0.75/0.71/0.68 for predicting TERT p mutation status

^c ITK-SNAP v3.8.0

^d TensorFlow 2.4

Lu et al., 2022 [100]

176 (training n = 123, validation n = 53)

CE-MRI

-

-

A total of 851 radiomic features extracted.

-

Radiomics

AUC = 0.873 (Validation set) achieved for predicting TERT p mutation

Fang et al. 2021 [18]

164

T2W, CE-T1W

Manual

^h Matlab

A total of 1,293 radiomics features from multi-parametric magnetic resonance extracted and analysed

CE-T1WI_Cluster Tendency

Machine learning

An overall accuracy of 0.7988 achieved in predicting TERT P mutation,

T1WI_Contrast

T1WI_Long Run Low Gray Level Emphasis_1

T1WI_Low Gray Level Run Emphasis

T2WI_Long Run High Gray Level Emphasis_1

Z. Li et al., 2021 [101]

159

Dynamic [18F] FET PET

Manual

^s PMOD view tool

107 radiomic features including first-order statistics, shape-based features, and texture features.

The TTP (Time to peak) model showed the strongest predictive power

Radiomics

0.921/NA/0.82

^b PyRadiomics (v 3.0.1)

Machine learning

Yan et al. 2021 [102]

357 (training, n = 238 and validation, n = 119)

T1WI, cT1WI, T2WI, T2-FLAIR, and DWI

Manual

^c ITK-SNAP

A total of 8730 and 4365 radiomic features extracted for gliomas with peritumoral edema, and without edema respectively.

Tumor_log-sigma-5-0-mm-3D_firstorder_10Percentile

Radiomics

Image fusion model integrating radiomic signatures from contrast-enhanced cT1WI and ADC achieved an AUC of 0.884 and 0.669 for predicting IDH and TERT status, respectively.

^b Pyradiomics 2.0.0

Tumor_wavelet-LL_firstorder_Skewness

Tumor_gradient_glszm_LargeAreaHighGrayLevelEmphasis

Tumor_original_shape_Sphericity

Tumor_log-sigma-1-0-mm-3D_firstorder_Median

Jiang et al. 2020 [103]

116 (training n = 83, validation n = 33)

CE-T1W, T2W

Manual

^c ITK-SNAP

A total of 107 radiomic features extracted, including 14 shape features, 18 first order features, and 75 texture features for each ROI at each modality

Correlation

Radiomics

The tumoral signature model yielded the best performance, with area under the ROC curves (AUC) of 0.948 the training cohort and 0.827 in the validation cohort.

^b PyRadiomics (2.1.0)

Gray-level non-uniformity normalized

^t Scikitlearn (v0.20.0)

Large dependence high gray-level emphasis

Large dependence low gray-level emphasis

Fukuma et al. 2019 [89]

164

T1W, T2W, FLAIR, and GdT1W

Manual

MATLAB-based image analysis software

Combination of CNN (AlexNet) and Conventional radiomics based approach used for analysis

-

Deep learning

CNN features succeeded in capturing characteristics of TERT p mutation, not identified by conventional radiomic features and patient age. Accuracy of 84.0% achieved using CNN features.

EGFR amplification

Calabrese et al. 2022 [6]

199

T1 pre, T1 post, T2, T2/FLAIR, SWI, DWI, ASL, MD, AD, RD, and FA

Automated deep learning-based tumor segmentation

^b Pyradiomics ver2.2

Shape features (n = 26), first order grayscale features (n = 19), and higher order grayscale features (n = 75).

-

Radiomics, Deep Learning

AUC/Sens/Spec of 0.70/0.66/0.68

S. Kihira et al. 2021 [82]

111 (Training n = 91, test n = 20)

T2, FLAIR, T1CE, DWI

Manual 2D

^h Matlab

Total of 92 radiomics features assessed

FLAIR – First order skewness

Machine learning

AUC/Sensitivity/Specificity: 0.65/0.68/0.83

FLAIR – GLSZM Small area emphasis

T1c + GLDM Small dependence low gray

Pasquini L. et al. 2021 [21]

156

MPRAGE, T1w, T2w, T2-FLAIR, DWI, DSC MRI

Manual 2D

^h Matlab

Radiomic set included 14 shape features, 18 intensity features, and 75 texture features

rCBV

Machine learning

Accuracy 81%; ROC 74.3%.

B. Sohn et al. 2021 [22]

418 (Training n = 292, test n = 126)

CE-T1w, T1w, T2w, T2-FLAIR

Auto-Segmentation

Python 3 with ScikitLearn library v0.21.2 and the R software

A total of 660 radiomic features were extracted

Run entropy (T1WI, CE mask), high gray-level cone emphasis (T2WI, CE mask), and inverse variance (T1WI, T2 mask).

ML (Binary relevance and Ensemble classifier chain)

AUC/Sensitivity/Specificity: 0.812/0.585/0.743

Li et al. 2018 [104]

270 (Training n = 200, test n = 70)

T2w

Manual 2D

^h Matlab

431 texture features (Divided in 4 groups, first order statistics, shape and size features, texture features, wavelet features)

Radiomic signature, comprising 25 first-order statistics or related wavelet features, one shape- and size-based feature, and 15 textural features or related wavelet features

Radiomics, Machine learning

41 MRI features achieved accuracies of 82.5% (AUC = 0.90) in the training set (n = 200) and 90.0% (AUC = 0.95) in the validation set (n = 70)

Hu LS et al. 2017 [105]

25 (GBM)

T1CE, DTI, DSC, PWI

Manual 2D

^c ITK SNAP

A total of 336 features extracted composed of 56 features across 6 MR contrasts.

T2.Information.Measure.of.Correlation.2_Avg_1

Radiogenomics, Machine learning

On validation set model achieved 78% accuracy amongst the sample predictions with lowest uncertainty

T2.Angular.Second.Moment_Avg_1

T2.Kurtosis

rCBV.Contrast_Avg_1

Kickingereder et al. 2016 [106]

152

MP-RAGE, T2FLAIR

Manual

^h Matlab

A total of 31 features extracted including mutiparametric and multiregional information with histogram quantification of tumor volumes, volume ratios, apparent diffusion coefficients, cerebral blood flow, cerebral blood volume, and intratumoral susceptibility signals.

Increased Gaussian-normalized relative cerebral blood volume and Gaussian-normalized relative cerebral blood flow values

Radiomics, Machine learning

Accuracy of 63% achieved in predicting EGFR amplification status using ML.

+ 7/-10 copy number change

Calabrese et al. 2022 [6]

199

T1 pre, T1 post, T2, T2/FLAIR, SWI, DWI, ASL, MD, AD, RD, and FA

Automated deep learning-based tumor segmentation followed by Manual correction

^b Pyradiomics ver2.2

Default set of shape features (n = 26), first-order grayscale features (n = 19), and higher-order grayscale features (n = 75) were extracted yielding 5300 radiomics features per patient

-

Radiomics, Deep Learning

AUC/Sens/Spec of 0.79/0.69/0.76 for predicting + 7/-10 copy number change

^c ITK-SNAP v3.8.0

^d TensorFlow 2.4

1p/19q Co-deletion

Kihira et al. 2023 [28]

103

T2-FLAIR

Manual

^u Oleasphere

92 radiomic texture features extracted from each VOI in a patient

Radiomics model outperformed the T2-FLAIR sign

Radiomics, Machine learning

AUC/Sensitivity/Specificity/Accuracy: 0.80/87.5%/89.9%/88.8%.

Casale et al. 2021 [107]

209 (training n = 159, validation n = 50)

T2 FLAIR, T1CEN

Manual

^v Weka software version 3.8.3

5352 radiomics features per patient extracted from both T1- and T2- weighted images. After correlation-based feature subset selection, 48 features remained for cubic interpolation and 51 features for linear interpolation.

-

Radiomics, Machine learning

Cubic interpolation (AUC/Sensitivity/Specificity/Accuracy: 0.81/0.77/0.85/0.86)

^h Python 3.7.6 version

^w MIM software version 6.9.0

Linear interpolation (AUC/Sensitivity/Specificity/Accuracy: 0.76/0.72/0.81/0.82)

Yan et al. 2021 [102]

357 (training, n = 238 and validation, n = 119)

T1WI, cT1WI, T2WI, T2-FLAIR, and DWI

Manual

^c ITK-SNAP

A total of 8730 and 4365 radiomic features extracted for gliomas with peritumoral edema, and without edema respectively.

Edema_log-sigma-5-0-mm-3D_firstorder_RootMeanSquared

Radiomics

CT1WI based radiomic signature yielded an AUC value of 0.815 in predicting 1p/19q status.

^b Pyradiomics 2.0.0

Tumor_log-sigma-1-0-mm-3D_firstorder_Mean

Edema_gradient_firstorder_Kurtosis

Tumor_log-sigma-3-0-mm-3D_glcm_ClusterShade

Tumor_log-sigma-5-0-mm-3D_firstorder_90Percentile

Kong et al. 2020 [108]

96 (training n = 78, validation n = 18)

T1CE, T2

Manual

^c ITK-SNAP

A total of 107 radiomics features were extracted from the region of interest (ROI) of each imaging modality using PyRadiomics

Informational Measure of Correlation 2

Radiomics, Machine learning

The 3D-radiomics signature displayed an accuracy of 0.897 and AUC of 0.940 in the training dataset, and an accuracy of 0.833 and AUC of 0.889 in the validation dataset.

Correlation

^b Pyradiomics

Dependence Entropy

Major Axis Length

Kocak et al. 2020 [109]

107

CET1W, T2W

Semi-automatic

^x LIFEx

A total of 84 radiomic features extracted

HISTO-Skewness (T2W)

Radiomics, Machine learning

AUC and accuracy values of five tested algorithms ranged from 0.769 to 0.869 and from 80.1 to 84%, respectively. Neural network had the highest mean rank with mean AUC and accuracy values of 0.869 and 83.8%, respectively.

GLZLM-HGZE (T2W)

HISTO-Entropy-log10 (T2W)

Batchala et al. 2019 [13]

102 (Data obtained from TCIA)

Manual

-

-

texture

-

Accuracy = 86.3% achieved for predicting the 1p/19q codeletion status

T2* susceptibility blooming

T2-FLAIR mismatch sign

Location

Midline shift

Van der Voort et al. 2019 [110]

284 (With validation on 129 cases from TCIA)

T1W, T2W, T2FLAIR

Manual

^c ITK-SNAP

A total of 78 image features (such as image intensity, tumor texture, tumor shape, and tumor location) extracted Age and sex added to yield as total of 80.

Cranial/caudal location of the tumor, skewness of the T2-weighted SI histogram, and one of the texture features, together with age and sex

Radiomics, Machine learning

An AUC of 0.72 yielded in the external validation dataset. Higher predictive performance than the average of the neurosurgeons (AUC 0.52) but lower than that of the neuroradiologists (AUC of 0.81)

Han et al. 2018 [111]

277 (training n = 184, validation n = 93)

T2

Manual

^c ITK-SNAP

647 radiomic features consisting of shape and size features (8), first order statistics features (17), textural features (54) were extracted for the original image set, and first order statistics and textural features for 8 wavelet filtered image sets.

ori_ fos_skewness (degree of distortion for the image)

Radiomics, Machine learning

The radiomics signature displayed good performance on both the training and validation cohorts with areas under the curve (AUCs) of 0.887 and 0.760, respectively. Results outperformed the clinical model, which demonstrated AUCs of 0.580 and 0.627 on the training and validation cohorts, respectively.

Coif5_glcm_covariance (measurement of heterogeneity in an image filtered with low-pass in the x-direction and high-pass in the y- and z-direction)

Coif2_glcm_ sum_variance (the change frequency and period of the texture in an image filtered with low-pass in the x- and y-direction and high-pass in the z-direction)

Lu et al. 2018 [112]

214 (with independent validation on a set of 70 patients)

T1W, CET1W, T2W, FLAIR, DWI

Semi-automatic with Manual correction

Home-made software, MR Radiomics Platform (MRP) Graphic interface built in Matlab

A maximum of 39,212 MR radiomic features generated for each subject.

Texture measurements describing spatial variations of tumor intensity found to be the most illustrative for the IDH and 1p/19q genotypes

Radiomics, Machine learning

An AUC of 0.922 achieved on the training dataset while and Accuracy of 80% yielded in predicting the 1p/19q co deletion status.

Akkus et al. 2017 [113]

159 Low grade gliomas patients

T2 and post-contrast T1-weighted MR

Semi-automatic

Semi-automatic LGG segmentation software (^p STAPLE software)

-

-

CNN (Deep learning)

Best performing configuration of CNN-architecture outperformed the classical ML model (SVM) with 93.3% (sensitivity), 82.22% (specificity), and 87.7% (accuracy)

Shofty et al. 2017 [114]

47

T2W, T1W, FLAIR

-

-

A total of 152 features, including size, location and texture, extracted.

-

Radiomics, Machine learning

Ensemble Bagged Trees classifier achieved best results with an AUC/Sens/Spec of 0.87/92/83 in predicting 1p/19q co deletion status.

H3 K27

Kun et al. 2023 [115]

103

CE-T1W

Manual

^g Medical Imaging Interaction Toolkit (MITK)

A total of 1781 radiomic features extracted and combined with conventional MR and clinical features to construct an integrated model

Age, 2 radiomics features, and 3 conventional MRI features were the 6 most significant features

Radiomics, Machine learning

Integrated model (Radiomic + Clinical features) achieved an optimal AUC/accuracy of 0.98/0.903 in the testing cohort.

^b Pyradiomics (version 3.0.1)

Yang et al. 2023 [116]

126

T1W, dMRI

-

21 radiomics features and 52 topological properties of brain structural connectivity network selected to construct a machine learning-based H3K27M mutation prediction model

-

Radiomics, Machine learning, Connectomics

AUC/Accuracy of 0.9246/92.11% achieved in predicting H3 K27 alteration. Combined multivariate logistic model built using T1 and dMRI based signature achieved an AUC of 0.8783 in the validation cohort

Li et al. 2023 [117]

418 (Diffused Midline Glioma patients) 133 (Spinal Cord Glioma Patients)

T2W

-

-

-

-

Deep Learning

Predictive accuracies, sensitivities, and specificities of H3 K27M mutation status were 92.1%, 98.2%, 82.9% in the testing cohort (In Diffuse Midline Glioma patients).

Guo et al. 2022 [54]

102 (training n = 72, test n = 30)

T2WI, T1WI, FLAIR, CE-T1WI, SWI, DWI, and DSC-PWI

Manual

^c ITK SNAP

Set of 851 features was extracted from each sequence image

-

Radiomics, Machine learning

The radiomics models based on multiparametric MRI demonstrated high accuracy in predicting the H3 K27M mutant status in diffuse midline gliomas (DMG), with AUC values ranging from 0.807 to 0.969 for different sequences or sequence combinations. The study identified the optimal model as one that utilized a combination of all sequences, achieving an AUC of 0.969.

^e FeAture Explore (V 0.4.2) using PyRadiomics

^l R software 3.5.3

Wu et al. 2022 [55]

107

CE-T1W, FLAIR, DWI, ADC

Manual

^b Pyradiomics

A total of 4520 radiomics features were extracted. Nine radiomics features selected among all extracted features to construct the radiomics signature

-

Radiomics, Machine learning

Ring enhancement was found to be a significant and independent clinical predictor (p < 0.01). Constructed nomogram, incorporating radiomics signature and ring enhancement, yielded an area under the curve (AUC) values of 0.95 and 0.90 in the training and testing sets, respectively.

^c ITK-SNAP

Kandermirli et al. 2021 [30]

109

Nonenhanced T1-W, T2-W, T2-FLAIR, postcontrast T1-W, Apparent diffusion coefficient (ADC) maps

Manual

^y SimpleITK

A total of 651 radiomic features per each sequence extracted

LoG-sigma 2 mm three dimensional first-order maximum on ADC maps

Radiomics, Machine learning

The study evaluated two models, and the XGBoost model with additional feature selection yielded superior results. The AUC for this model was 0.791 and 0.737 in the training and test set respectively.

^b PyRadiomics

Wavelet-LH first-order range on T2-WI

Wavelet-LH gray-level dependence matrix large dependence high gray-level emphasis on ADC maps

Original first-order mean absolute deviation on ADC maps

LoGsigma 4 mm three-dimensional first-order maximum on FLAIR

Li et al. 2021 [118]

30

T1WI, T2WI, CE-T1WI

Manual

^c ITK SNAP

272 radiomic features were extracted from MR images of each tumor, grouped into seven categories: shape features, first-order features, GLCM features, GLRLM features, GLSZM features, and GLDM features.

T2WI

Radiomics, Machine learning?

Cyst formation exhibited a significant difference between DGM-M and DGM-W tumors (p = 0.024) among the visually accessible features. However, there were no significant differences observed between DGM-M and DGM-W tumors for necrosis (p = 0.191) hemorrhage (p = 0.657), and the T1/T2 ratio (p = 0.689).

^b PyRadiomics

Zhuo et al. 2021 [119]

81 (training n = 64, test n = 17)

T1W, T2W, T2-FLAIR, DWI, CE-T1W, APTw

Manual

^a 3D Slicer

A total of 1316 radiomic features were obtained from 3D tumor masks.

-

Radiomics, Machine learning

Utilizing support vector machine (SVM) to identify radiomic features derived from amide proton transfer-weighted (APTw) imaging, a high accuracy rate of 0.99 (63/64) was achieved for the retrospective cohort's prediction of H3K27M-mutant tumors in the training set. In the test set, the accuracy was 0.88 (15/17).

^e Feature Analysis Explorer v0.3.6

Su et al. 2020 [120]

100 (training n = 75, testing n = 25)

T1W, CE-T1W, T2W, T2 FLAIR

Manual

^c ITK SNAP

A total of 18 first-order features, 13 shape features, 22 GLCM features, 16 GLRLM) features, and 16 GLSZM features were extracted

original_glszm_GrayLevelVariance

original_firstorder_10percentile

Radiomics, Machine learning

Out of the 10 models evaluated, the highest-performing one achieved an AUC of 0.903 in the training cohort and 0.85 in the validation set.

^b PyRadiomics

original_shape_Maximum2DDiameterSlice

^l R package ver 3.6

original_shape_SurfaceVolumeRatio

original_shape_Volume

Pan et al. 2019 [121]

151

T1W, CE-T1W, T2W, CE-T2W

-

-

A total of 1697 features, including 6 clinical parameters and 1691 imaging features extracted

-

Machine learning

Machine learning-based model achieved an accuracy of 84.44% (AUC of 0.8298) in the test cohort for predicting H3 K27 mutation. The simplified model achieved an AUC of 0.7839 in the test cohort.

Liu et al 2018 [122]

55 (training n = 38, validation n = 4, testing n = 13)

T1w-MPRAGE

Auto segmentation (Deep learning)

^a 3D slicer version 4.1

Cascaded two-task framework for segmentation and H3 K27 status prediction using CNN

CNN30 + SVM

Deep learning

The CNN-feature based method outperformed the traditional hand-crafted feature based method by at least 17% and 0.30, with a prediction accuracy (ACC) of 96.52% and an AUC of 0.953

H3 G34

Shao et al. 2024 [123]

53

-

-

^b PyRadiomics

Visually Accessible Rembrandt Images (VASARI) features and radiomic features extracted

-

Radiomics

FAE-generated model, based on radiomics features (AUC 0.925), displayed better discriminatory performance between G34m-DHG and IDH-WT-GBM than VASARI feature analysis (AUC 0.843).

Machine learning