Table 1.
Summary of the major publications in retinal analysis using DL, grouped by standard multiple-field fundus photography, ultra-wide field photography and smartphone-based photography
| Authors and year of publication | Approach | Training dataset | Validation datasets | Performance |
|---|---|---|---|---|
| Standard view photography | ||||
| Gulshan et al. 2016 [35] | Inception-V3 network |
Public EyePACS and Messidor-2 (> 120,000 images) |
Public EyePACS-1 and Messidor-2 (> 10,000 images) |
EyePACS-1 |
| AUC: 0.99 | ||||
| Sensitivity: 90% | ||||
| Specificity: 98% | ||||
| Messidor-2 | ||||
| AUC: 0.99 | ||||
| Sensitivity: 87% | ||||
| Specificity: 99% | ||||
| Abràmoff et al. 2016 [46] | AlexNet/VGG network |
Public Messidor-2 |
Public Messidor-2 (~ 2000 images) |
AUC: 0.98 |
| Sensitivity: 97% | ||||
| Specificity: 87% | ||||
| Ting et al. 2017 [26] | VGGNet-19 network |
Proprietary SiDRP 2010–2013 (> 76,000 images) |
Proprietary SiDRP 14–15 and 10 others (> 112,000 images) |
SiDRP 2014–2015 |
| AUC: 0.93 | ||||
| Sensitivity: 91% | ||||
| Specificity: 92% | ||||
| Others | ||||
| AUC range: 0.89 to 0.98 | ||||
| Sensitivity range: 92 to 100% | ||||
| Specificity: 76 to 92% | ||||
| Gargeya et al. 2017 [47] | Customised CNN network | Public EyePACS-1 (> 75,000 images) |
Public EyePACS-1, Messidor-2,E-Ophtha (> 17,000 images) |
EyePACS-1 |
| AUC: 0.97 | ||||
| Sensitivity: 94% | ||||
| Specificity: 96% | ||||
| Messidor-2 and E-Ophtha | ||||
| AUC range: 0.83 to 0.95 | ||||
| Sensitivity range: 74 to 93% | ||||
| Specificity range: 87 to 94% | ||||
| Abràmoff et al. 2018 [34] | AlexNet/VGGNet network |
Public Messidor-2 |
Proprietary Primary care sites (~ 900 patients) |
Sensitivity: 87% |
| Specificity: 91% | ||||
| Keel et al. 2018 [48] | Inception-V3 network |
Public LabelMe (~ 59,000) |
Proprietary Endocrinology outpatient services (96 patients) |
Sensitivity: 92% |
| Specificity: 94% | ||||
| Kanagasingam et al. 2018 [49] | Inception-V3 network |
Public and proprietary DiaRetDB1, EyePACS, Australian tele-eye care (30,000 images) |
Proprietary Primary care (~ 200 patients) |
Sensitivity: 92% |
| Gulshan et al. 2019 [50] | Inception-v4 network |
Public EyePACS and Messidor-2 (> 144,000 images) |
Proprietary Two eye hospitals (~ 6000 images) |
AUC range: 0.97 to 0.98 |
| Sensitivity range: 89 to 92% | ||||
| Specificity range: 92 to 95% | ||||
| Raumviboonsuk et al. 2019 [51] | Inception-v4 network |
Public EyePACS and Messidor-2 (> 120,000 images) |
Proprietary Hospitals and health centers (~ 30,000 images) |
AUC: 0.99 |
| Sensitivity: 96.9% | ||||
| Specificity: 95.3% | ||||
| Bellemo et al. 2019 [52] | VGGNet/ResNet network |
Proprietary SiDRP 2010–2013 (> 76,000 images) |
Proprietary Mobile screening unit (> 4000 images) |
AUC: 0.97 |
| Sensitivity: 92% | ||||
| Specificity: 89% | ||||
| Ultra-wide field photography | ||||
| Wang et al. 2018 [53] | EyeArt software | – |
Proprietary Eye clinics (~ 1500 images) |
AUC: 0.85 |
| Sensitivity: 90% | ||||
| Specificity: 54% | ||||
| Nagasawa at al. 2019 [54] | VGGNet-16 network | Proprietary Hospitals (< 400 images) |
Proprietary Hospitals (< 400 images) |
AUC: 0.97 |
| Sensitivity: 95% | ||||
| Specificity: 97% | ||||
| Smartphone-based photography | ||||
| Rajalakshmi et al. 2018 [55] | EyeArt software | – |
Proprietary Tertiary care diabetes hospital (~ 300 images) |
Sensitivity: 96% |
| Specificity: 80% | ||||
| Natarajan et al. 2019 [56] |
Remidio software Inception-V3 network |
Public and proprietary EyePACS and hospitals (> 52,000 images) |
Proprietary Population-based screening (> 4000 images) |
Sensitivity range: 96 to 100% |
| Specificity range: 79 to 88% | ||||
| Rogers et al. 2019 [57] | Pegasus software | – |
Public and proprietary IDRiD and research laboratory study (> 6000 images) |
AUC range: 89 to 99% |
| Sensitivity range: 82 to 93% | ||||
| Specificity range: 82 to 94% | ||||