Skip to main content

Table 2 Final model sensitivity, specificity, and area under the receiver operating characteristic curve measures of prediction outputs of early-stage FECD, late-stage FECD, and healthy cornea at both image and eye levels. All AUC P-values < 0.001 for all parameters

From: Automated diagnosis and staging of Fuchs’ endothelial cell corneal dystrophy using deep learning

  Image level (n = 7380 images)
Healthy cornea prediction output Early-stage FECD prediction output Late-stage FECD prediction output
AUC ± SE (95% CI) 0.998 ± 0.001 (0.997, 0.999) 0.984 ± 0.003 (0.979, 0.999) 0.974 ± 0.005 (0.964, 0.985)
Sensitivity 99% 91% 92%
Specificity 98% 97% 91%
Cutoff value 0.93 0.04 0.01
  Eye level (n = 41 eyes)
Healthy cornea prediction output Early-stage FECD prediction output Late-stage FECD prediction output
AUC ± SE (95% CI) 1.0 (1, 1) 0.997 ± 0.005 (0.988, 1.0) 0.988 ± 0.017 (0.954, 1.0)
Sensitivity 100% 100% 100%
Specificity 100% 97% 98%
Cutoff value 0.49 0.03 0.67
  1. FECD = Fuchs’ endothelial cell corneal dystrophy; AUC = Area under the curve; SE = Standard error; 95% CI: 95% confidence interval on the difference between groups. Specificity, sensitivity, and cutoff values are chosen to maximize total diagnostic accuracy (minimize total number of errors).