Table 3 Model performance comparisons for ROP detection before and after image quality enhancement.

From: DeepQuality improves infant retinopathy screening

Model architecture

Metrics (95% CI)

Model using raw images

Model using enhanced images

p value

InceptionV3

Sensitivity

0.508 (0.560, 0.600)

0.670 (0.651, 0.689)

<0.0001*

Specificity

0.966 (0.959, 0.974)

0991 (0.987, 0.995)

<0.0001*

Accuracy

0.827 (0.812, 0.842)

0.875 (0.862, 0.889)

<0.0001*

AUC

0.875 (0.861, 0.889)

0.962 (0.954, 0.970)

<0.0001*

DenseNet

Sensitivity

0.436 (0.416, 0.456)

0.540 (0.520, 0.561)

<0.0001*

Specificity

0.992 (0.988, 0.996)

0.989 (0.985, 0.993)

0.5689

Accuracy

0.791 (0.775, 0.808)

0.827 (0.812, 0.843)

0.0022*

AUC

0.939 (0.929, 0.949)

0.961 (0.953, 0.969)

<0.0001*

  1. p values for sensitivity, specificity, and accuracy were calculated using a two-proportion z test. The p value for AUC was calculated using the DeLong test.
  2. AUC area under the receiver operating characteristic curve, ROP retinopathy of prematurity.
  3. *p < 0.05.
Back to article page