Table 2 (A) Percentage of features with significant differences in distribution before and after harmonization by the GMM groupings. Feature names indicate the feature whose distribution was used to generate the GMM scan grouping. GMM scan groupings are obtained by selecting the best GMM model from a set composed of GMM models generated from each of the features such that the final GMM scan grouping is estimated from a single feature. (B) Percentage of features with significantly different distributions attributable to batch effects in the original features and after applying standard ComBat, harmonizing by the GMM grouping alone (GMM), and harmonizing by both the GMM grouping and known imaging parameter batch effects (GMM + ComBat (CE)).

From: Generalized ComBat harmonization methods for radiomic features with multi-modal distributions and multiple batch effects

A

Original (%)

ComBat (%)

Lung3/CAPTK

T1_E_GLRLM_Short RunLowGreyLevel emphasis

88

45

Lung3/PyRadiomics

Idmn

84

26

Radiogenomics/CAPTK

T1_ED_GRLRLM_Bins-10_Radius-1_ShortRun LowGreyLevelEmphasis

78

50

Radiogenomics/PyRadiomics

Jointenergy

75

30

B

Original (%)

ComBat (%)

GMM (%)

GMM + ComBat (%)

Lung3/CAPTK

CE

10

16

4

4

Spatial resolution

18

21

28

10

Manufacturer

48

45

7

4

Lung3/PyRadiomics

CE

40

11

35

7

Spatial resolution

43

25

44

15

Manufacturer

61

28

43

23

Radiogenomics/CAPTK

CE

17

42

18

12

Spatial resolution

42

43

45

25

Manufacturer

20

51

17

25

Radiogenomics/PyRadiomics

CE

54

27

47

16

Spatial resolution

69

29

62

19

Manufacturer

44

36

40

23

  1. Tables contain the percentage of features out of the original number of features with detected significant (p < 0.05) differences in distribution for all batch effects.