背景:与其他方法相比,生成对抗网络(GAN)已显示出优越的数据生成能力,使它们在医学图像应用中流行。这些特征引起了医学成像领域的研究人员的兴趣,导致这些技术在各种传统和新颖的应用中的快速实现,如图像重建,分割,检测,分类,和跨模态合成。对最近医学成像突破的全面回顾将使对该领域感兴趣的研究人员受益。在这次审查中,我们的目的是介绍起源,原则,和GAN的扩展形式,并总结了基于GAN的医学图像处理方法的最新进展。
方法:我们使用关键词“分割,\"\"分类,医学图像,\"和\"生成对抗网络。\"具体来说,在删除了重复和不可访问的全文出版物后,最初的搜索显示了5423种出版物。然后,经过标题和摘要筛选,680人进行了全文筛选。最后,在全文筛选后,121项研究被纳入我们的最终分析。
结果:本综述涵盖的研究日期范围为2017年1月1日至今。经过彻底的筛选和资格评估,最终的方法学综述中包括了121项涉及基于GAN的七个医学图像领域应用的研究。这些领域包括综合,分类,分割,转换,重建,去噪,和病变检测。我们将这些论文进一步分类和总结为临床应用,分类方法,和成像模式。
结论:我们深入研究了基于GAN的医学图像增强的最新研究进展。这些技术有效地缓解了医学图像诊断和治疗模型的有限训练样本的挑战。此外,与GAN相关的几个关键问题,比如模式崩溃,不稳定性,缺乏可解释性,在未来的研究中需要注意。
BACKGROUND: Generative adversarial networks (GANs) have demonstrated superior data generation capabilities compared to other methods, making them popular for use in medical image applications. These features have intrigued researchers in the medical imaging field, resulting in a swift implementation of these techniques in various conventional and novel applications such as image reconstruction, segmentation, detection, classification, and cross-modality synthesis. A comprehensive
review of recent medical imaging breakthroughs will benefit researchers interested in this field. In this
review, we aimed to introduce the origin, principle, and extended forms of GANs and summarize the state-of-the-art progress of GAN-based medical image processing methods.
METHODS: We searched the literature for studies on Google Scholar and PubMed using the keywords \"Segmentation,\" \"Classification,\" \"medical image,\" and \"generative adversarial network.\" Specifically, the initial search revealed 5423 publications after the removal of duplicated and non-accessible fulltext publications. Then, after the title and abstract screening, 680 underwent full-text screening. Finally, 121 studies were included in our final analysis after full-text screening.
RESULTS: The date range of the studies covered in this
review is from January 1, 2017, to the present. After a thorough screening and qualification assessment, 121 studies involving GAN-based applications in seven areas of medical images were included in the final methodological
review. These areas included synthesis, classification, segmentation, conversion, reconstruction, denoising, and lesion detection. We further classified and summarized these papers into clinical applications, classification methods, and imaging modalities.
CONCLUSIONS: We thoroughly examined the latest research progress of GAN-based medical image augmentation. These techniques effectively alleviate the challenge of limited training samples for medical image diagnosis and treatment models. Furthermore, several critical issues associated with GANs, such as pattern collapse, instability, and lack of interpretability, require attention in future research.