Table 2 Summary of literature review for image segmentation task using GAN in ophthalmology imaging domains

Iqbal et al. [27]

Conditional GAN

Fundus photography

Retinal vessels

The framework achieved retinal vessels image segmentation from a small training set

Wang et al. [33]

Conditional GAN (using PatchGAN)

Fundus photography

Optic disc and optic cup

Unsupervised domain adaptation for joint optic disc and cup segmentation using a patch-based discriminator

Son et al. [28]

Conditional GAN (using U-Net as a generator)

Fundus photography

Retinal vessels

The GAN model segmented retinal vasculature and the optic disc using a dataset, which consisted of fundoscopic images and manual segmentation of the vessels

Rammy et al. [29]

Conditional GAN

Fundus photography

Retinal vessels

Patch-based GAN with additional loss function to learn thin and thick vessel segmented retinal vessels with the enhanced performance

Park et al. [30]

Conditional GAN

Fundus photography

Retinal vessels

The proposed GAN model with a multi-kernel pooling and false negative loss could segment retinal blood vessels more robustly

Heisler et al. [36]

Pix2Pix (conditional GAN)

Peripapillary retinal OCT

Nerve fiber layer, Bruch’s membrane, choroid-sclera boundary

The use of a generative adversarial network and unlabeled data can improve the performance of segmentation for the 3D morphometric analysis of glaucomatous optic nerve head volumes

Yang et al. [31]

Conditional GAN (topological structure-constrained)

Fundus photography

Retinal vessels

The topological structure-constrained proposed GAN model identified retinal arteries and veins via segmentation from the complex background of retinal images

Yang et al. [106]

Conditional GAN

Fundus photography

Retinal vessels

To separate blood vessels from fundus image, the model could detect more tiny vessels and locate the edge of blood vessels more accurately

Zhao et al. [32]

Conditional GAN (with a large receptive field)

Fundus photography

Retinal vessels

The proposed retinal vessel segmentation algorithm using GAN with a large receptive field could capture large-scale high-level semantic vessel features

Kadambi et al. [34]

Wasserstein GAN

Fundus photography

Optic disc and optic cup

WGAN-based domain adaptation showed a better performance than baseline models for the joint optic disc-and-cup segmentation in fundus images

Bian et al. [35]

Conditional GAN (using U-Net as a generator)

Fundus photography

Optic disc and optic cup

The proposed method successfully performed optic disc and cup segmentation. The cup-to-disc ratio was automatically calculated with a good performance

Khan et al. [38]

Conditional GAN

Meibography infrared images

Meibomian gland

The proposed GAN automatically identified the area of the meibomian glands and outperformed the state-of-art methods

Yildiz et al. [37]

Conditional GAN

In vivo corneal confocal microscopy images

Sub-basal nerves

Automatic segmentation of sub-basal nerves in in vivo confocal microscopy images was performed using the U-Net and GAN-based techniques as a diagnostic tool for corneal diseases

Zhou et al. [107]

Conditional GAN (using U-Net as a baseline architecture)

Fundus photography

Retinal vessels

The GAN model strengthened retinal vessel segmentation in the low-contrast background using a symmetric equilibrium GAN (U-Net-based), multi-scale features refine blocks and attention mechanism