Table 3 Effect of technical factors specifically convolutional neural networks and computational framework.

From: Technical and imaging factors influencing performance of deep learning systems for diabetic retinopathy

   

Convolutional neural networks

Computational frameworks

   

VGGNet

ResNet

DenseNet

Ensemble

Caffe

TensorFlow

SiDRP

Value (95% CI)

AUC

0.938 (0.929–0.945)

0.936 (0.927–0.944)

0.941 (0.933–0.947)

0.944 (0.938–0.950)

0.936 (0.927–0.944)

0.938 (0.929–0.945)

P value for AUC comparison

Reference

0.581

0.410

0.02

Reference

0.736

Sensitivity

92.1% (89.2–94.5%)

91.9% (88.9–94.3%)

92.8% (90.0–95.1%)

94.0% (91.3–96.0%)

90.5% (87.3–93.1%)

92.1% (89.2–94.5%)

Specificity

91.0% (90.7–91.3%)

90.9% (90.6–91.2%)

90.9% (90.6–91.2%)

90.7% (90.4–91.0%)

91.9% (91.6–92.2%)

91.0% (90.7–91.3%)

  1. P value was calculated by bootstrap method.
  2. Dataset used for evaluation of different computational frameworks and convolutional neural networks is Singapore integrated Diabetic Retinopathy Programme (SiDRP) 2014 to 2015. During the evaluation of the impact of the convolutional neural network (CNN) on the DL algorithm performance, the computational framework was controlled for by using TensorFlow for fair comparison. Similarly, during the evaluation of different computational frameworks, the convolutional neural network controlled was controlled for by using VGGNet for isolation of independent variables.
  3. AUC area under receiver operating curve, CI confidence Interval, SiDRP Singapore integrated Diabetic Retinopathy Programme.
Back to article page