Table 3 Importance score and stability of top 10 radiomic features.

From: Robust performance of deep learning for distinguishing glioblastoma from single brain metastasis using radiomic features: model development and validation

Radiomic feature

Permutation feature importance

ICC (95% CI)

AdaBoost

L-SVM

LDA

Total

First order, median (CE mask on T2WI)

15.8

100.0

45.9

161.7

0.989 (0.968–0.996)

First order, 10 percentile (CE mask on T2WI)

86.8

55.3

7.7

149.8

0.942 (0.844–0.979)

GLCM, inverse difference (CE mask on CE-T1WI)

31.6

1.3

100.0

132.9

0.995 (0.985–0.998)

Shape, maximum 2D diameter row (PT mask on T2WI)

78.9

43.4

9.0

131.4

0.969 (0.915–0.989)

First order, median (PT mask on T2WI)

60.5

65.8

NS

126.3

0.863 (0.660–0.950)

GLRLM, low gray level run emphasis (CE mask on CE-T1WI)

0.0

44.7

63.4

108.1

0.995 (0.985–0.998)

Shape, flatness (PT mask on T2WI)

60.5

34.2

5.5

100.2

0.866 (0.665–0.951)

GLSZM, gray level non uniformity (CE mask on T2WI)

100.0

NS

NS

100.0

0.972 (0.921–0.990)

GLCM, inverse variance (CE mask on CE-T1WI)

0.0

36.8

57.7

94.5

0.994 (0.980–0.998)

GLCM, autocorrelation (CE mask on CE-T1WI)

0.0

0.0

94.3

94.3

0.991 (0.972–0.997)

  1. Importance scores of all the radiomic features used for seven machine learning classifiers are presented in Supplementary Table S4.
  2. AdaBoost adaptive boosting, L-SVM linear support vector machine, LDA linear discriminant analysis, ICC intraclass correlation coefficient, CI confidence interval, CE contrast-enhancing, T2WI T2-weighted imaging, T1WI T1-weighted imaging, PT peritumoral T2 hyperintense, GLCM gray level co-occurrence matrix, GLRLM gray level run length matrix, GLSZM gray level size zone matrix, NS not selected for this classifier.
Back to article page