Table 2 Five-fold cross validation results for each of the single- and multiple-column models.

From: Deep learning models for screening of high myopia using optical coherence tomography

CNN backbone

Micro-average AUC of the single-column models

Micro-average AUC of the multiple-column models

Initialization with ImageNet-pretrained models

Initialization with the pretrained single-column models

VGG 16

Vertical

0.9859 ± 0.00

(0.9826–0.9906)

0.5801 ± 0.07

(0.5189–0.6409)

0.6827 ± 0.17

(0.5307–0.8341)

Horizontal

0.9873 ± 0.01

(0.9811–0.9934)

Resnet 50

Vertical

0.9746 ± 0.01

(0.9646–0.9850)

0.5545 ± 0.08

(0.4829–0.6261)

1.0000 ± 0.00

(1.0–1.0)

Horizontal

0.9844 ± 0.01

(0.9796–0.9896)

Inception V3

Vertical

0.8967 ± 0.04

(0.8625–0.9310)

0.8048 ± 0.07

(0.7455–0.8648)

0.9170 ± 0.03

(0.8886–0.9453)

Horizontal

0.9188 ± 0.04

(0.8809–0.9568)

  1. CNN convolutional neural network, AUC area under the receiver operating characteristic curve.
  2. Data are mean ± standard deviation (95% confidence interval).
Back to article page