Generative Adversarial Networks in Medicine: A Scoping Review

Background. The effectiveness of artificial intelligence (AI) in medicine can increase significantly if the ways of using AI are studied, and a large amount of training data is available. Generative adversarial networks (GAN) have demonstrated a huge potential for data augmentation by generating synthetic data that helps replace missing data quite well. In addition, this type of neural networks, namely its specific architectures, are also popular for segmentation and prediction. Purpose. This review aims to explore how GAN methods are used in medicine. The review describes various GAN applications for various fields of medicine, presents a variety of GAN architectures, including very specific ones, at the same time presents the most popular ones, and the review also presents the most popular metrics used to evaluate models. This is the first review in which an attempt is made to cover quite generally the field of application in medicine. It attempts to answer questions related to GAN applications, popular GAN architectures and metrics, as well as to find out how the frequency of use of the COVID topic in articles in this area has changed. Materials and methods. This review was conducted in accordance with the PRISMA-ScR recommendations for the search and selection of studies. The search was conducted on the popular scientific database Scopus. Results. This review included 49 studies from 228 search results. The most common use case of GANs was data synthesis to augment data. GANs were also used for image segmentation and for predicting the necessary information, as well as converting objects from one structure to another. The included studies have shown that GANS can improve the performance of artificial intelligence methods used to process data from various fields of medicine. However, additional efforts are needed to transform GAN-based methods into clinical applications. Conclusion. Studies have shown that GANS have great potential to solve the problem of lack of research data in various fields of medicine. The data synthesized using GAN was useful for improving convolutional neural network (CNN) learning. Models trained to diagnose various diseases or diagnose covid. In addition, GANS have also contributed to performance enhancement (CNN) through image over-resolution and segmentation. This review also identified key limitations of the potential transformation of GAN-based methods in clinical application.