生成对抗式网络在心脏磁共振中的应用  

Application of generative adversarial networks in cardiac magnetic resonance

作　　者：刘维肖 方进[1] 王莹 莫笑开 张水兴 LIU Weixiao;FANG Jin;WANG Ying;MO Xiaokai;ZHANG Shuixing(Department of Radiology,the First Affiliated Hospital of Jinan University,Guangzhou 510000,China)

机构地区：[1]暨南大学附属第一医院影像科,广州510000

出　　处：《磁共振成像》2023年第6期139-144,共6页Chinese Journal of Magnetic Resonance Imaging

基　　金：国家自然科学基金(编号:81871323);广东省医学科学技术研究基金(编号:A2022267);暨南大学中央高校青年基金(编号:21621050)。

摘　　要：心脏磁共振(cardiac magnetic resonance, CMR)成像是评估心血管结构与功能的重要检查方法。深度学习已经被广泛应用于CMR图像处理中。生成对抗式网络(generative adversarial network, GAN)作为深度学习中的一种新兴网络,利用负责图像生成的生成器与负责判断图像真假的判别器之间的对抗博弈,可创建具有强大生成和泛化能力的图像处理模型。本文总结GAN在CMR图像分割与合成、图像重建、超分辨率重建、虚拟增强方面的主要应用,结合目前CMR的临床应用需求,分析GAN模型面临的挑战和未来的展望,以期尽快提升其实际应用价值。Cardiac magnetic resonance(CMR)is an important examination method for evaluating cardiovascular structure and function.Deep learning has been widely used in CMR image processing.Generative adversarial network(GAN),as a new type of network in deep learning,uses the adversarial game between the generator responsible for image generation and the discriminator responsible for judging the authenticity of images to create image processing models with powerful generation and generalization capabilities.We summarized the main applications of GAN in CMR image segmentation and synthesis,image reconstruction,super-resolution reconstruction,and virtual native enhancement in this paper.Combined with the current clinical application requirements of CMR,we analyzed the challenges and future prospects faced by GAN models in order to improve their practical application value as soon as possible.

关 键 词：心脏磁共振 图像处理 生成对抗式网络 深度学习 人工智能 磁共振成像 

分 类 号：R445.2[医药卫生—影像医学与核医学] R816.2[医药卫生—诊断学]