Analysis of preprocessing for Generative Adversarial Networks: A case study on color fundoscopy to fluorescein angiography image-to-image translation

Abstract

Generative Adversarial Networks (GANs) are capturing the attention of peer researchers in paired or unpaired image-to-image translation applications, particularly in the domain of retinal image processing. Additionally, there are several effective image preprocessing techniques available that can significantly improve the performance of GANs. This study examines the impact of five different image preprocessing techniques - Green Channel, CLAHE on Green Channel, CLAHE on RGB channels, Green Channel Gaussian Convolution, and RGB Gaussian Convolution - on five different GAN variants: CycleGAN, Pix2Pix GAN, CUT GAN, FastCut GAN, and NICE GAN. The study involved conducting 30 experiments to assess the performances of these GAN variants in the image-to-image translation of dual-mode retinal images. The evaluation utilized Frechet Inception Distance (FID) and Kernel Inception Distance (KID) metric scores to measure the performance of the GAN variants. The results demonstrated that the CycleGAN model achieved the best performance with CLAHE on RGB preprocessed images, achieving the lowest FID and KID scores of 103.49 and 0.038, respectively. This investigation underscores the significant potential of image preprocessing techniques in enhancing the performance of GANs in image translation applications.