Table 2 Classification performance across different cancer types.

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

Organ

Number of cases

Number of positive cases

Number of negative cases

AUC (Area under curve)

TP/FP/TN/FN

Sensitivity

Specificity

Accuracy

Lung

207

153

54

0.8895(0.8466–0.9324)

153/48/6/0

100%

11.1%

76.8%

Breast

173

35

138

0.9986(0.9958-1.000)

35/10/128/0

100%

92.8%

94.2%

Lymph node

121

16

105

0.8774(0.7570–0.9978)

11/3/102/5

68.8%

97.1%

93.4%

Female adnexal tumors

50

12

38

0.9930(0.9762-1.000)

11/6/32/1

91.7%

84.2%

86.0%

Thyroid

42

23

19

0.9144(0.8266-1.000)

17/2/17/6

73.9%

89.5%

81.0%

Other organs

76

22

54

0.8242(0.6816–0.9668)

15/3/51/7

68.2%

94.4%

86.8%

  1. The metrics in Table 2 represent the raw diagnostic potential of the model per organ site. Note that the overall metrics in Table 1/3 are derived from a multi-stage MIL aggregation (including a top-10% instance selection cutoff and a synchronized slide-level threshold), which optimizes global specificity for clinical safety. Consequently, the summation of individual organ counts in Table 2 may slightly differ from the global optimized metrics due to this hierarchical thresholding logic.
Back to article page