Table 3.
Summary of most relevant works in classification and segmentation tasks, where OCT images are taken as input.
| Year | Reference | Topic | Model | Dataset (patients) | Output: classes | AUC | Sens. | Spec. | Better than experts? (N) |
|---|---|---|---|---|---|---|---|---|---|
| 2019 | Motozawa et al. [46] | AMD | CNN (Custom, VGG-like) | Private (271) | AMD | 0.995 | 1,000 | 0.918 | |
| AMD: Fluid, no fluid | 0.991 | 0.984 | 0.883 | ||||||
| 2019 | Antony et al. [26] | AMD | CNN (VGG16) | Private (384) | AMD (+ relevant B-scans) | 0.967 | 0.910 | 0.870 | |
| 2019 | Hassan et al. [41] | DME: Fluid seg., Vessel seg. (Fundus + OCT) | CNN (Custom) + Heavy processing | Rabani and Zhang (683) | DME | 0.970 | 0.920 | Yes (3) | |
| Segmentation: Hard exudates, Blood vessel, Retinal fluid | (DSCs: 0.707, | ||||||||
| 0.820, 0.902) | |||||||||
| 2019 | Kuwayama et al. [76] | AMD, DR, ERM + Others | CNN (Custom) | Private (1200) | Normal | 0.850 | 0.970 | ||
| Wet AMD | 1,000 | 0.770 | |||||||
| DR | 0.780 | 1,000 | |||||||
| ERM | 0.750 | 0.750 | |||||||
| 2018 | De Fauw et al. [47] | AMD, DR, ERM + Others | CNN (Segmentation, Custom) + CNN (Classification, Custom) | UK National Heatlh Service (NHS) (14884) | Referral: urgent, non urgent | 0.992 | Yes (8) | ||
| Normal | 0.995 | Same | |||||||
| MRE | 0.990 | Yes | |||||||
| CNV | 0.993 | Yes | |||||||
| Drusen | 0.974 | Same | |||||||
| ERM | 0.966 | Same | |||||||
| Others (GA, CSR, Full/partial thickness macular hole, VMT) | 0.980 | Same | |||||||
| 2018 | Shigueoka et al. [57] | Glaucoma (OCT + SAP) | Feature extraction + Several ML classifiers | University of Campinas, Brazil (124) | Glaucoma: Early or moderate, none | 0.931 | 0.900 | 0.800 | Same (3) |
| 2017 | Schlegl et al. [77] | AMD, DME: Fluid seg. | CNN (Segmentation, custom) | Private (1200) | Intraretinal fluid | 0.940 | |||
| Subretinal fluid | 0.933 |