Skip to main content
. 2022 Jun 21;12:10455. doi: 10.1038/s41598-022-13919-2

Table 1.

Review of studies using CNN method to assess retinal image quality.

Author Year Database Method (architecture) Category of image quality Definition of classification Performance
Mahapatra et al.18 2016 A DR screening initiative CNN, and local saliency map Gradable, ungradable NA Se: 98.2%, Sp: 97.8%, Acc: 97.9%
Yu et al.19 2017 Kaggle CNN(Alexnet), and saliency map Good, poor NA Se: 96.63%, Sp: 93.10%, Acc: 95.42%
Saha et al.20 2017 EyePACS CNN (Alexnet) Accept, reject, ambiguous Yes Acc: 100%
Zago et al.21 2018 DRIMDB and ELSA-Brasil CNN (Inception-v3) Good, poor, outlier NA

DRIMDB: Se: 97.10%, Sp:100.0%, AUC: 99.98%

ELSA-Brasil: Se: 92.00%, Sp: 96.00%,AUC: 98.56%

Chalakkala et al.22 2019 DRIMDB, DR1–DR2, HRF, MESSIDOR, UoA-DR, Kaggle, IDRiD, Six pre-trained CNN (AlexNet, GoogLeNet, ResNet50, ResNet101, Inception-v3, SqueezeNet) MSRI high quality, low quality Some databases Se: 98.38%; Sp: 95.19%; Acc: 97.47%
Shen et al.23 2020 Shanghai Diabetic Retinopathy Screening Program Multi-task deep learning framework (VGG16) Gradable, ungradable Yes Se: 83.62%, Sp: 91.72%, AUC: 94.55%
Yuen et al.24 2021

Primary dataset: CUHK Eye Center, National University Hospital

External dataset: Hong Kong Children Eye Study, Queen’s University Belfast

Two CNN (EfficientNet-B0, MobileNetV2) Gradable, ungradable Yes Se: 92.1%, Sp: 98.3%, Acc: 92.5%, AUC: 97.5%

DL deep learning, CNN convolutional neural network, MSRI medically suitable retinal image, NA not available, Se sensitivity, Sp specificity, Acc accuracy, AUC area under the curve.