计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (15): 238-245.DOI: 10.3778/j.issn.1002-8331.2012-0291

• 图形图像处理 • 上一篇    下一篇

基于生成对抗网络的磁共振图像重建

胥祯浩,沈茜,李霞   

  1. 中国计量大学 信息工程学院,杭州 310018
  • 出版日期:2022-08-01 发布日期:2022-08-01

Magnetic Resonance Images Reconstruction Based on Generative Adversarial Network

XU Zhenhao, SHEN Xi, LI Xia   

  1. School of Information Engineering, China Jiliang University, Hangzhou 310018, China
  • Online:2022-08-01 Published:2022-08-01

摘要: 磁共振成像因具有无辐射、无创伤性,成为临床中最常用的辅助诊断技术之一,但过长的扫描时间和封闭的环境,不仅导致病人产生幽闭恐惧心理,也造成医疗成本的升高。针对此问题,提出了一种以生成对抗网络为核心的磁共振图像重建算法,将U-net网络作为生成器,编码部分使用残差结构以缓解网络退化,并提出空洞金字塔结构,利用空洞卷积的不同扩张率融合不同尺度的上下文信息并添加于解码层之前。判别器中通过一系列卷积实现特征下采样,并利用sigmoid函数完成特征分类,将集成学习的思想融入其中,使重建效果进一步提升。对比已有研究成果和主流重建网络,该模型在10%、20%、30%、50%采样率的测试集中,各项重建指标均排名第一。结果表明,该模型不仅能有效提升磁共振图像重建质量,同时也具有良好的泛化性。

关键词: 磁共振成像, 图像重建, 生成对抗网络, 空洞金字塔, 集成学习

Abstract: Magnetic resonance imaging(MRI) technology has become one of the commonly used auxiliary diagnosis and treatment methods in the clinic because of its non-radiation and non-trauma characteristics. However, the long scanning time and the closed environment have led to claustrophobia and the medical costs increased. MR image reconstruction algorithm based on the generative adversarial network(GAN) is proposed to solve this problem. In the GAN network, the U-net structure is used as the generator, which utilizes the residual structure in the encoding layer of U-net to alleviate the gradient disappearance and propose the atrous pyramid structure which utilizing the atrous convolutions with different dilation rates for multi-scale information fusion. In the discriminator, a series of convolutions are utilized to achieve feature down-sampling, and complete feature classification through the sigmoid function. To further improve the quality of reconstruction, the idea of ensemble learning is introduced. Comparing existing research results and mainstream reconstruction networks, the proposed model ranks first with four different sampling rates of 10%, 20%, 30%, and 50%, which shows that this model can not only effectively improve the quality of MRI reconstruction, but also has good generalization.

Key words: magnetic resonance imaging, image reconstruction, generative adversarial network, atrous pyramid structure, ensemble learning