Table 2 Comparison of classification results for the TCGA and UCSF external validation sets with CNN, Visual transformer, and GlioMT

From: Interpretable multimodal transformer for prediction of molecular subtypes and grades in adult-type diffuse gliomas

Classification task

DL network

TCGA

UCSF

AUC (95% CI)

Accuracy (%)

Sensitivity (%)

Specificity (%)

AUC (95% CI)

Accuracy (%)

Sensitivity (%)

Specificity (%)

IDH mutation

CNN

0.862 (0.807–0.911)

79.5

79.1

79.8

0.950 (0.926–0.970)

89.5

83.5

91.2

Visual transformer

0.901 (0.852–0.942)

83.0

74.4

89.5

0.971 (0.952–0.985)

93.3

81.6

96.5

GlioMT

0.915 (0.869–0.955)

85.5

82.6

87.7

0.981 (0.968–0.991)

94.8

85.4

97.3

1p/19q codeletion

CNN

0.773 (0.663–0.869)

65.1

89.7

52.6

0.740 (0.580–0.875)

55.3

69.2

53.3

Visual transformer

0.831 (0.741–0.909)

72.1

34.5

91.2

0.737 (0.559–0.890)

82.5

61.5

85.6

GlioMT

0.854 (0.770–0.929)

75.6

58.6

84.2

0.806 (0.646–0.946)

76.7

76.9

76.7

Tumor grade

CNN

0.793 (0.722–0.859)

65.0

47.6

77.7

0.932 (0.907–0.956)

82.0

54.4

88.0

Visual transformer

0.840 (0.779–0.895)

65.5

49.8

81.2

0.947 (0.925–0.966)

85.7

70.6

92.7

GlioMT

0.862 (0.806–0.912)

70.0

54.1

80.4

0.960 (0.942–0.977)

90.4

66.1

91.0

  1. For the 1p/19q codeletion task, the TCGA and UCSF validation sets included only 86 and 103 cases, respectively. For the IDH mutation and tumor grade prediction tasks, the TCGA and UCSF validation sets included 200 and 477 cases, respectively. Best AUC is highlighted in bold.
  2. DL deep learning, AUC the area under the receiver operating characteristic curve, CI confidence interval, CNN convolutional neural network, CNS central nervous system, IDH isocitrate dehydrogenase, TCGA Tumor Cancer Genome Atlas, UCSF The University of California, San Francisco, WHO World Health Organization.
Back to article page