Doorly Robert, Ong Joshua, Waisberg Ethan, Sarker Prithul, Zaman Nasif, Tavakkoli Alireza, Lee Andrew G
University of Cambridge, Cambridge, UK.
Department of Ophthalmology and Visual Sciences, University of Michigan Kellogg Eye Center, Ann Arbor, MI, USA.
Graefes Arch Clin Exp Ophthalmol. 2025 Apr 22. doi: 10.1007/s00417-025-06830-9.
Generative adversarial networks (GANs) are key components of many artificial intelligence (AI) systems that are applied to image-informed bioengineering and medicine. GANs combat key limitations facing deep learning models: small, unbalanced datasets containing few images of severe disease. The predictive capacity of conditional GANs may also be extremely useful in managing disease on an individual basis. This narrative review focusses on the application of GANs in ophthalmology, in order to provide a critical account of the current state and ongoing challenges for healthcare professionals and allied scientists who are interested in this rapidly evolving field.
We performed a search of studies that apply generative adversarial networks (GANs) in diagnosis, therapy and prognosis of eight eye diseases. These disparate tasks were selected to highlight developments in GAN techniques, differences and common features to aid practitioners and future adopters in the field of ophthalmology.
The studies we identified show that GANs have demonstrated capacity to: generate realistic and useful synthetic images, convert image modality, improve image quality, enhance extraction of relevant features, and provide prognostic predictions based on input images and other relevant data.
The broad range of architectures considered describe how GAN technology is evolving to meet different challenges (including segmentation and multi-modal imaging) that are of particular relevance to ophthalmology. The wide availability of datasets now facilitates the entry of new researchers to the field. However mainstream adoption of GAN technology for clinical use remains contingent on larger public datasets for widespread validation and necessary regulatory oversight.
生成对抗网络（GANs）是许多应用于图像辅助生物工程和医学的人工智能（AI）系统的关键组成部分。GANs克服了深度学习模型面临的关键限制：包含少量严重疾病图像的小而不平衡的数据集。条件GANs的预测能力在个体疾病管理中可能也极为有用。这篇叙述性综述聚焦于GANs在眼科中的应用，以便为对这一快速发展领域感兴趣的医疗保健专业人员和相关科学家提供对当前状况和持续挑战的批判性描述。
我们检索了在八种眼病的诊断、治疗和预后中应用生成对抗网络（GANs）的研究。选择这些不同的任务是为了突出GAN技术的发展、差异和共同特征，以帮助眼科领域的从业者和未来的采用者。
我们确定的研究表明，GANs已展现出以下能力：生成逼真且有用的合成图像、转换图像模态、提高图像质量、增强相关特征的提取，以及基于输入图像和其他相关数据提供预后预测。
所考虑的广泛架构描述了GAN技术如何不断发展以应对与眼科特别相关的不同挑战（包括分割和多模态成像）。数据集的广泛可用性现在便于新研究人员进入该领域。然而，GAN技术在临床中的主流应用仍取决于用于广泛验证的更大公共数据集以及必要的监管监督。
