Differentiation of Recurrence from Radiation Necrosis in Gliomas Based on the Radiomics of Combinational Features and Multimodality MRI Images
Table 3
rs between features (portion of the handcrafted and deep features) and glioma recurrence versus necrosis (,α = 0.05, and K = 176, 4,096, and 2,048 for handcrafted, AlexNet, and Inception v3 features, respectively).
Type
Feature
Modality
rs
value
Nontexture
Volume
T1
0.0373
0.7949
T2
FLAIR
T1C
Size
T1
0.0172
0.9045
T2
FLAIR
T1C
Solidity
T1
0.0115
0.9363
T2
FLAIR
T1C
Eccentricity
T1
−0.0172
0.9045
T2
FLAIR
T1C
GLRLM
HGRE
T1
0.3273
0.0190
T2
−0.3331
0.0169
FLAIR
−0.3187
0.0226
T1C
0.4594
0.0007
GLSZM
HGZE
T1
0.3790
0.0061
T2
−0.4508
0.0009
FLAIR
−0.4852
0.0003
T1C
0.4738
0.0004
SZLGE
T1
0.3876
0.0049
T2
−0.3790
0.0061
FLAIR
−0.4652
0.0006
T1C
−0.3962
0.0040
SZHGE
T1
0.4163
0.0024
T2
−0.3446
0.0133
FLAIR
−0.4738
0.0004
T1C
0.3618
0.0091
AlexNet
F7_618
T1C
0.5656
0.00001
F7_1394
T1
0.5168
0.0001
F7_2793
FLAIR
0.4823
0.0003
F7_3501
T2
0.4421
0.0012
Inception v3
avg_pool_663
T1
0.5770
0.0000093
avg_pool__469
T1C
0.5483
0.000031
avg_pool_827
FLAIR
0.3876
0.005
avg_pool_774
T2
0.4651
0.000584
For deep feature names, the first character indicates the layer of the CNN and the second character represents the neuron. For example, F7_618 was extracted from a T1C image and taken from the 618th neuron of fully connected layer 7.