波谱学杂志 ›› 2023, Vol. 40 ›› Issue (1): 39-51.doi: 10.11938/cjmr20222992

• 研究论文 • 上一篇    下一篇

基于PCAU-Net的快速多通道磁共振成像方法

施伟成,金朝阳*(),叶铮   

  1. 人机混合智能与智慧健康研究中心,自动化学院,杭州电子科技大学,浙江 杭州 310018
  • 收稿日期:2022-03-29 出版日期:2023-03-05 在线发表日期:2022-08-10
  • 通讯作者: 金朝阳 E-mail:jinzhaoyang@hdu.edu.cn.
  • 基金资助:
    国家自然科学基金面上项目(61372024)

Fast Multi-channel Magnetic Resonance Imaging Based on PCAU-Net

SHI Weicheng,JIN Zhaoyang*(),YE Zheng   

  1. Machine Learning and I-health International Cooperation Base of Zhejiang Province, School of Automation, Hangzhou Dianzi University, Hangzhou 310018, China
  • Received:2022-03-29 Published:2023-03-05 Online:2022-08-10
  • Contact: JIN Zhaoyang E-mail:jinzhaoyang@hdu.edu.cn.

摘要:

多通道磁共振成像方法采用多个接收线圈同时欠采样k空间以加快成像速度,并基于后处理算法重建图像,但在较高加速因子时,其图像重建质量仍然较差.本文提出了一种基于PCAU-Net的快速多通道磁共振成像方法,将单通道实数U型卷积神经网络拓展到多通道复数卷积神经网络,设计了一种结构不对称的U型网络结构,通过在解码部分减小网络规模以降低模型的复杂度.PCAU-Net网络在跳跃连接前增加了1×1卷积,以实现跨通道信息交互.输入和输出之间利用残差连接为误差的反向传播提供捷径.实验结果表明,使用规则和随机采样模板,在不同加速因子时,相比常规的GRAPPA重建算法和SPIRiT重建方法,本文提出的PCAU-Net方法可高质量重建出磁共振复数图像,并且相比于PCU-Net方法,PCAU-Net减少了模型参数、缩短了训练时间.

关键词: 多线圈, 快速磁共振成像, 深度学习, 复数卷积, 不对称神经网络

Abstract:

Multi-channel magnetic resonance imaging methods utilize multiple receiving coils to simultaneously under-sample k-space for fast magnetic resonance imaging. Post-processing algorithms were often used to restore the missing information for under-sampled images. However, the quality of reconstructed images is far from satisfying, especially at higher acceleration factor. In this backdrop, a fast multi-channel magnetic resonance imaging method based on parallel complex asymmetric U-Net (PCAU-Net) was proposed, which extends single-channel real-valued U-shaped convolutional neural network to multi-channel complex convolutional neural network. A U-shaped network with an asymmetric structure was designed to reduce the model complexity by reducing the network size in decoding. A 1×1 convolution block was added before skip-connection for PCAU-Net to share cross-channel information. The residual connection between the input and output of the network facilitates loss backpropagation. The experimental results show that by using regular and random sampling templates, the proposed PCAU-Net method can reconstruct the complex magnetic resonance images with high quality compared with the conventional generalized auto-calibrating partially parallel acquisitions (GRAPPA) reconstruction algorithm and iterative self-consistent parallel imaging reconstruction (SPIRiT) algorithm at different acceleration factors. Moreover, compared with the parallel complex U-Net (PCU-Net) method, PCAU-Net reduces model parameters and shortens the training time.

Key words: multi-coils, fast magnetic resonance imaging, deep learning, complex convolution, asymmetric neural network

中图分类号: