Table 3 Summary of literature review for data augmentation task using GAN in ophthalmology imaging domains

Diaz-Pinto et al. [39]

DCGAN

Peripapillary fundus photography (optic disc photo)

DCGAN was able to generate high-quality synthetic optic disc images

Burlina et al. [40]

PGGAN

Fundus photography

The GAN technique was used to synthesize high-resolution realistic fundus images serving as proxy data sets for use by retinal specialists and deep learning models

Zheng et al. [43]

PGGAN

Retinal OCT (spectral domain)

The image quality of real images vs. synthetic OCT images generated by GAN was similar; the synthetic OCT images were able to serve as augmentation of training datasets for deep learning models

Zhou et al. [17]

Conditional GAN

Fundus photography

To generate a large amount of balanced training data, the GAN model synthesized high-resolution diabetic retinopathy fundus images which can be manipulated with arbitrary grading and lesion information

Wang et al. [41]

Multi-channel GAN (modified vanilla GAN)

Fundus photography

The model generated a series of sub-fundus images corresponding to the scattering diabetic retinopathy features and made full use of both labeled and unlabeled data

He et al. [42]

Label smoothing GAN (modified vanilla GAN)

Retinal OCT

The GAN model generated the synthetic unlabeled images from limited OCT training samples, and the mixing of the synthetic images and real images can be used as training data to improve the classification performance

Yoo et al. [45]

CycleGAN

Retinal OCT

GAN generated OCT images of rare diseases from normal OCT images and increased the accuracy of diagnosing rare retinal diseases with few-shot classification

Kugelman et al. [44]

Conditional GAN

Retinal OCT (patch level)

GAN was feasible to generate patches that are visually indistinguishable from their real variants and improved the segmentation performance

Zheng et al. [82]

PGGAN

Anterior Segment OCT

The synthetic OCT images generated by GAN appeared to be of good quality, according to the glaucoma specialists, and the deep learning model for angle-closure detection was improved using both synthetic and real images

Yoo et al. [46]

CycleGAN, PGGAN

Ocular surface image

To improve the diagnostic accuracy, GAN was adopted to perform data augmentation of ocular surface images with conjunctival melanoma

Abdelmotaal et al. [47]

Pix2pix

Corneal topography (Scheimpflug images)

The synthesized images showed plausible subjectively- and objectively-assessed quality. Training deep learning with a combination of real and synthesized images showed better classification performance to detect keratoconus