Table 2 Detection and segmentation accuracy in internal and external testing cohorts.

From: Automatic cervical lymph nodes detection and segmentation in heterogeneous computed tomography images using deep transfer learning

Variable

Sens (%)

PPV (%)

FP/vol

DSC*

HD95* (mm)

DSC#

HD95 (mm)#

ITC A

54.6 ± 22.9

69.0 ± 19.5

3.4

0.69 ± 0.15

20.37 ± 31.24

0.72 ± 0.19

3.78 ± 4.12

External testing cohorts

 ETC B

48.5 ± 16.7

57.0 ± 17.3

5.0

0.61 ± 0.11

22.28 ± 21.18

0.72 ± 0.15

2.85 ± 3.28

 ETC C

45.7 ± 18.5

64.6 ± 18.1

3.9

0.63 ± 0.13

24.05 ± 13.88

0.73 ± 0.16

2.83 ± 3.62

 ETC D

63.5 ± 21.4

50.0 ± 18.7

5.8

0.66 ± 0.11

26.13 ± 48.61

0.74 ± 0.15

2.73 ± 2.77

  1. *The DSC and HD95 were obtained in patient level, and all model-predicted lymph nodes were involved in calculation; #, The DSC and HD95 were obtained in nodal level, and only true positive lymph nodes were involved in calculation. Data were denoted as mean with standard deviation.
  2. In ITC A and ETC C, we only calculated the accuracy of lymph nodes in contrast enhanced computed tomography images.
  3. ITC = internal testing cohort; ETC = external testing cohort; Sens = sensitivity, PPV = positive predicted value, FP/vol = false positive per volume, DSC = Dice similarity coefficient, HD95 = 95% Hausdorff distance.
Back to article page