Table 3 Comparison of model performance and diagnostic metrics (Test set).

From: Pan-cancer frozen section classification using a soft mixture of experts vision transformer under weak supervision

Model

AUC

(95% CI)

PR-AUC

(95% CI)

Accuracy

Sensitivity

Specificity

LR+

(95% CI)

LR-

(95% CI)

Brier score

Soft MoE-ViT (Ours)

0.957 (0.940–0.971)

0.949 (0.932–0.965)

90.4%

81.7%

96.1%

22.27 (13.5–38.2)

0.19 (0.14–0.24)

0.167

Standard ViT

0.934 (0.914–0.952)

0.907 (0.875–0.934)

85.8%

91.2%

82.4%

5.23 (4.20–6.49)

0.11 (0.07–0.15)

0.171

ResNet

0.921 (0.900-0.941)

0.830 (0.770–0.887)

88.2%

86.6%

89.2%

8.21 (6.16–10.73)

0.15 (0.10–0.20)

0.174

  1. *Data for Soft MoE-ViT, Standard ViT, and ResNet reflect strict validation-to-test parameter selection.
Back to article page