Retinal Image Synthesis for Glaucoma Assessment Using DCGAN and VAE Models

Abstract

The performance of a Glaucoma assessment system is highly affected by the number of labelled images used during the training stage. However, labelled images are often scarce or costly to obtain. In this paper, we address the problem of synthesising retinal fundus images by training a Variational Autoencoder and an adversarial model on 2357 retinal images. The innovation of this approach is in synthesising retinal images without using previous vessel segmentation from a separate method, which makes this system completely independent. The obtained models are image synthesizers capable of generating any amount of cropped retinal images from a simple normal distribution. Furthermore, more images were used for training than any other work in the literature. Synthetic images were qualitatively evaluated by 10 clinical experts and their consistency were estimated by measuring the proportion of pixels corresponding to the anatomical structures around the optic disc. Moreover, we calculated the mean-squared error between the average 2D-histogram of synthetic and real images, obtaining a small difference of 3 × 10 −4. Further analysis of the latent space and cup size of the images was performed by measuring the Cup/Disc ratio of synthetic images using a state-of-the-art method. The results obtained from this analysis and the qualitative and quantitative evaluation demonstrate that the synthesised images are anatomically consistent and the system is a promising step towards a model capable of generating labelled images.

Accuracy of Using Generative Adversarial Networks for Glaucoma Detection: Systematic Review and Bibliometric Analysis

“…Papers [ 42 , 46 , 49 , 51 , 57 , 73 , 75 , 76 , 80 - 87 ] represented about 50% of the studies (n=16) and were U-Net-based architectures. However, the other 50% of the papers [ 43 , 46 , 50 , 51 , 58 , 74 , 77 - 79 , 88 - 94 ] were CNN-based generators (n=16).…”

“…These references contributed to 20/30 (67%) of the total papers. Seventeen of them were BV-based methods [ 42 , 43 , 49 - 51 , 58 , 75 , 76 , 78 , 81 - 85 , 88 , 91 , 92 , 94 ]. Only 2 studies [ 57 , 81 ] were OD-based detection approaches, and 1 [ 82 ] utilized RNFL-based detection ( Figure 6 ).…”

“…Studies [ 80 , 93 ] worked on BV and OD segmentation, and only Wang et al [ 73 ] used RNFL and RL. The rest of the papers used triple landmarks in their work, such as [ 58 ] and [ 90 ], which involved work on BV, OD, and background, and Yu et al [ 46 ] used BV, OD, and OC.…”

“…As illustrated in Figure 5 , references with P, H, or G letters represent refer to PixelGAN, PatchGAN, or ImageGAN, respectively. ImageGAN papers were [ 42 , 51 , 58 , 80 , 86 , 88 , 90 , 93 , 94 ], while PixelGAN papers were [ 49 , 73 , 74 , 76 - 78 , 82 , 91 , 92 ]. In addition, PatchGAN papers were [ 43 , 46 , 50 , 57 , 75 , 79 , 81 , 83 - 85 , 87 , 89 ].…”

“…To win the game, both G and D need to continuously improve their generation and discrimination capabilities, respectively. The training process lasts until both G and D reach a convergence point (Nash equilibrium), where G generates an output distribution very close to the real data distribution [ 42 , 57 , 58 ].…”