A systematic review of generative adversarial networks (GANs) in plastic surgery.

ABSTRACT

INTRODUCTION: Generative adversarial networks (GANs) are a form of deep learning architecture based on the zero-sum game theory, which uses real data to generate realistic fake data. GANs use two opposing neural networks working a generator and a discriminator. They represent a powerful tool for generating realistic synthetic patient data sets and can potentially revolutionize research. This systematic literature review evaluated the scale and scope of GANs within plastic surgery, constructing a framework for its use and evaluation within subspecialties. METHODS: Following Preferred Reporting Items for Systematic Reviews and Meta-analyses guidelines, a systematic review was performed for applications of GANs in plastic surgery from 2014 to 2022. Three independent reviewers screened from databases PubMed, Embase, PsychInfo, Scopus, and Google Scholar. RESULTS: A total of 70 studies were captured by the search, of which seven studies met our criteria. The most common subspecialty was craniofacial (n = 4). Proposed uses of GANs included facial recognition, burn estimation, scar prediction, and post-breast cancer reconstruction anomaly scoring. GANs were conditional, trained on data sets averaging 54,652 ± 112,180 samples, with some sourced publicly and others being primary. CONCLUSION: GANs hold promise for advancing plastic surgery, backed by diverse applications in the literature. Studies should follow a standardized reporting structure for consistency and transparency, as outlined, especially regarding the data sets used to ensure appropriate representation from an ethnic and cultural diversity perspective. Although GANs require specialist computational expertise to create, surgeons need to understand their development by leveraging the full potential of GANs within the emerging field of computational plastic surgery and beyond.