GAN review: Models and medical image fusion applications

Abstract

Generative Adversarial Network (GAN) is a research hotspot in deep generative models, which has been widely used in the field of medical image fusion. This paper summarizes GAN models from the following four aspects: firstly, the basic principles of GAN are expounded from two aspects: basic model and training process; secondly, variant GAN models are summarized into three directions (Probability Distribution Distance, Overall Network Architecture, Neural Network Structure), from the methods based on f-divergence, the methods based on IPM, Single-Generator and Dual-Discriminators GAN, Multi-Generators and Single-Discriminator GAN, Multi-Generators and Multi-Discriminators GAN, Conditional Constraint GAN, Convolutional Neural Network structure GAN and Auto-Encoder Neural Network structure GAN are eight dimensions to summarize the typical models in recent years; thirdly, the advantages and application of GAN models in the field of medical image fusion are explored from three aspects; fourthly, the main challenges faced by GAN and the challenges faced by GAN models in medical image fusion field are discussed and the future prospects are given. This paper systematically summarizes various models of GAN, advantages and challenges of GAN models in medical image fusion field, which is very important for the future research of GAN.