Table 4 The clinical validation by grading actual and predicted visual field (VF)

From: Interpretable longitudinal glaucoma visual field estimation deep learning system from fundus images and clinical narratives

HPA method

Mild

Moderate

Severe

Total

The number of each category

 Original VFs

169

64

65

  

298

 VFs predicted by ROI

183

55

60

  

298

 VFs predicted by ROI + Clin

184

57

57

  

298

Accuracy

 VFs predicted by ROI

0.92 (0.88-0.96)

0.53 (0.41-0.66)

0.83 (0.74–0.92)

  

0.82 (0.77-0.86)

 VFs predicted by ROI + Clin

0.93 (0.90–0.97)

0.56 (0.44–0.69)

0.83 (0.74–0.92)

  

0.83 (0.79–0.87)

Voronoi method

Clear

Mild

Moderate

Severe

Diffuse

Total

The number of each category

 Original VFs

50

77

83

68

20

298

 VFs predicted by ROI

67

71

70

66

24

298

 VFs predicted by ROI + Clin

86

56

72

60

24

298

Accuracy

 VFs predicted by ROI

0.80 (0.69-0.91)

0.55 (0.43–0.66)

0.61 (0.51-0.72)

0.75 (0.64-0.86)

0.65 (0.42-0.88)

0.66 (0.61-0.72)

 VFs predicted by ROI + Clin

0.86 (0.76-0.96)

0.52 (0.41-0.63)

0.72 (0.62-0.82)

0.75 (0.64-0.86)

0.75 (0.54-0.96)

0.70 (0.65-0.75)

  1. VF visual field, CFP color fundus photograph, ROI region of interest, Clin clinical information