Multiparametric Magnetic Resonance Imaging Information Fusion Using Graph Convolutional Network for Glioma Grading
Table 1
Demonstration of the graph convolution module can fuse contextual information in MRI with different imaging parameters to improve the performance of the CNN models.
Dataset
Method
GoogLeNet-Inceptionv3
EfficientNet-b3
ResNet-34
AlexNet
VGGNet-16
AUC
ACC
AUC
ACC
AUC
ACC
AUC
ACC
AUC
ACC
BraTS2020
T1CE
0.911
0.893
0.921
0.893
0.932
0.907
0.910
0.880
0.915
0.867
G-TICE
0.931
0.907
0.936
0.920
0.946
0.920
0.943
0.907
0.947
0.893
T1
0.851
0.813
0.884
0.867
0.896
0.840
0.804
0.800
0.854
0.827
G-T1
0.860
0.853
0.896
0.880
0.906
0.853
0.816
0.827
0.863
0.853
T2
0.893
0.867
0.892
0.853
0.849
0.840
0.867
0.853
0.876
0.840
G-T2
0.904
0.893
0.904
0.893
0.873
0.880
0.886
0.880
0.903
0.880
FLAIR
0.897
0.827
0.885
0.853
0.894
0.853
0.802
0.827
0.889
0.840
G-FLAIR
0.911
0.867
0.905
0.867
0.908
0.867
0.834
0.840
0.915
0.867
MMIF-GCN
0.963
0.920
0.974
0.933
0.986
0.947
0.954
0.933
0.974
0.920
GliomaHPPH2018
T1CE
0.810
0.787
0.952
0.851
0.842
0.787
0.802
0.809
0.867
0.787
G-TICE
0.904
0.851
0.965
0.894
0.900
0.830
0.846
0.830
0.929
0.872
T1
0.898
0.872
0.923
0.872
0.921
0.894
0.729
0.766
0.933
0.872
G-T1
0.913
0.894
0.942
0.915
0.962
0.915
0.821
0.830
0.942
0.936
T2
0.692
0.766
0.798
0.745
0.729
0.745
0.629
0.660
0.856
0.766
G-T2
0.769
0.787
0.856
0.851
0.813
0.787
0.731
0.809
0.863
0.830
FLAIR
0.873
0.809
0.967
0.872
0.931
0.872
0.838
0.766
0.879
0.851
G-FLAIR
0.890
0.830
0.973
0.915
0.942
0.894
0.846
0.809
0.896
0.872
MMIF-GCN
0.960
0.936
1.000
0.979
0.996
0.979
0.881
0.851
0.979
0.957
Note. G means to fuse context information with GCN. The results of our study are shown in bold.
