Skip to main content

Table 5 Performance of radiologists without and with model assistance on the external test set

From: Automatic segmentation of fat metaplasia on sacroiliac joint MRI using deep learning

Metrics

Radiologists

Model-assisted radiologists

Radiological resident

Expert radiologist

Radiological resident

p valuea

Expert radiologist

P valueb

Accuracy (%)

70.69

(58.18–81.82)

79.31

(69.09–89.14)

77.51

(67.27–87.32)

0.572

82.55

(70.91–90.91)

0.152

Sensitivity (%)

78.38

(63.89–91.18)

89.19

(76.92–97.30)

89.22

(78.79–97.44)

0.424

94.85

(85.71–100.00)

0.508

Specificity (%)

57.14

(33.33–78.58)

61.90

(40.72–83.33)

57.20

(34.99–78.57)

0.990

61.51

(40.00–82.36)

0.302

PPV (%)

76.32

(62.16–89.29)

80.49

(68.42–92.50)

78.27

(65.00–90.24)

/

80.87

(68.18–92.50)

/

NPV (%)

60.00

(35.71–82.35)

76.47

(52.94–94.44)

75.34

(52.93–95.00)

/

87.41

(66.67–100.00)

/

Cohen’s κ

0.359

(0.082–0.599)

0.533

(0.299–0.766)

0.484

(0.238–0.725)

/

0.596

(0.369–0.802)

/

  1. All data in parentheses are 95% confidential intervals. p values less than 0.05 show statistical differences
  2. AUC an area under the receiver operating characteristic curve, PPV positive predictive value, NPV negative predictive value
  3. aData were compared between the radiological resident and model-assisted radiological resident
  4. bData were compared between the expert radiologist and model-assisted expert radiologist