Table 2 Clinicians’ assessment of the authenticity of real SDOCT and synthetically enhanced SDOCT images

From: Optical coherence tomography choroidal enhancement using generative deep learning

 

Expert Majority, %

Expert 1, %

Expert 2, %

Expert 3, %

 

Accuracy

Sensitivity

Specificity

Accuracy

Sensitivity

Specificity

Accuracy

Sensitivity

Specificity

Accuracy

Sensitivity

Specificity

All

47.5

47.6

47.4

45.0

43.7

46.4

53.0

57.3

48.5

56.5

61.2

51.5

Normal

52.1

50.0

54.2

50.0

45.8

54.2

56.2

58.3

54.2

56.2

62.5

50.0

Glaucoma

45.5

46.4

44.4

50.9

53.6

48.1

49.1

46.4

51.9

52.7

53.6

51.9

DR

46.4

47.1

45.7

39.2

37.3

41.3

53.6

62.7

43.5

58.8

64.7

52.2

  1. The table shows the results of the ability of experts to differentiate whether the presented images were real or synthetic on visual inspection and reports the overall findings as an averaged combination of Task 1 and Task 2 outcomes. Accuracy, sensitivity, and specificity discrimination scores were calculated for a total of 30 images from normal eyes, 25 glaucoma, and 50 diabetic retinopathy eyes considering 100 images from Task 1 and 100 image pairs from Task 1. (See Supplementary Table 2 for task-specific results). SDOCT Spectral-Domain Optical Coherence Tomography, SSOCT Swept-Source Optical Coherence Tomography, DR Diabetic Retinopathy.
Back to article page