Table 1.
The characteristics of typical GAN variant techniques and examples of general tasks in general medicine and ophthalmology fields
| GAN Techniques | Dataset | Characteristics | Task examples in general medicine | Task examples in ophthalmology |
|---|---|---|---|---|
| Deep convolutional GAN (DCGAN) | Images from one domain | Improved image quality using deep convolutional layers | Augmentation of CT images [88] | Fundus photographs synthesis [39] |
| Wasserstein GAN (WGAN) | Images from one domain | Using Wasserstein distance as a loss function |
Augmentation of CT images [89] Removing artifacts in CT [90] |
Anomaly detection [25] OCT segmentation [91] |
| Progressively growing GAN (PGGAN) | Images from one domain (generally high resolution) | High resolution & realistic image generation |
X-ray image synthesis [92] Data augmentation for cytological images [93] |
Data augmentation for fundus photography, retinal OCT, and ocular images [40, 46, 82] Super-resolution of fundus photographs [55] |
| StyleGAN | Images of one domain or multiple domains (unpaired images) | Disentanglement of representations (mapping features to low dimensions) |
Augmentation of CT and MRI images in specific conditions [13, 94] Skin image synthesis [95] |
None |
| Conditional GAN (vector input models) | Images annotated by conditional variables | Image synthesis conditioned to specific variables | Super-resolution guided by a conditional variable [96] |
Data augmentation for retinal OCT [44] Post-intervention (orbital decompression) prediction [15] |
| Conditional GAN (Pix2pix and other image input models) | Paired images of two domains or classes. (Training samples should be aligned) | Supervised learning for image-to-image translation |
Super-resolution for fluorescence microscopy images [97] Domain transfer (CT → PET) [98] Segmentation of lungs from chest X-ray [99] CT image synthesis [100] |
Domain transfer (fundus photography → angiography) [62] Retinal OCT segmentation [36] Retinal vessel segmentation [28] Data augmentation for fundus photography and corneal topography [17, 47] |
| Super-resolution GAN (SRGAN) | Low- and high-resolution image pairs | Adopting perceptual loss to generate super-resolved realistic images | Super-resolution for dental X-ray [101] | Super-resolution for optic disc photography [56] |
| Cycle-consistent GAN (CycleGAN) | Unpaired images of two domains or classes | Adopting a cycle consistency for domain transfer without any paired dataset |
Manipulating breast imaging [102] Data augmentation for CT and throat images [103, 104] Segmentation for cardiac ultrasound [105] |
Denoising for fundus photography and OCT [22, 53] Domain transfer (Ultra-widefield retinal images → classic fundus photography) [63] |
| StarGAN | Unpaired images of multiple domains or classes | A single network to achieve translation of multiple domains | Domain transfer between MRI contrasts [24] | None |
CT = computed tomography; GAN = generative adversarial network; MRI = magnetic resonance imaging; OCT = optical coherence tomography; PET = positron emission tomography