波谱学杂志 ›› 2023, Vol. 40 ›› Issue (3): 258-269.doi: 10.11938/cjmr20233050

• 研究论文 • 上一篇    下一篇

基于自监督网络的肝脏磁共振R2*参数图像重建

陆琪琪,连梓锋,李嘉龙,斯文彬,麦兆华,冯衍秋*()   

  1. 南方医科大学 生物医学工程学院,广东 广州 510515
  • 收稿日期:2023-01-09 出版日期:2023-09-05 在线发表日期:2023-03-02
  • 通讯作者: *Tel: +86 20 61648271, E-mail: foree@smu.edu.cn.
  • 基金资助:
    国家自然科学基金资助项目(U21A6005)

Magnetic Resonance R2* Parameter Mapping of Liver Based on Self-supervised Deep Neural Network

LU Qiqi,LIAN Zifeng,LI Jialong,SI Wenbin,MAI Zhaohua,FENG Yanqiu*()   

  1. School of Biomedical Engineering, Southern Medical University, Guangzhou 510515, China
  • Received:2023-01-09 Published:2023-09-05 Online:2023-03-02
  • Contact: *Tel: +86 20 61648271, E-mail: foree@smu.edu.cn.

摘要:

磁共振等效横向弛豫率($R_{2}^{*}$)参数量化技术已经被广泛应用于肝脏铁含量的定量测量中. 然而铁沉积肝脏$R_{2}^{*}$参数图像的重建通常会受到噪声的严重影响. 随着深度学习的兴起,深度学习网络成为磁共振参数图像重建的重要方法. 本文提出了一种模型引导的自监督深度学习网络用于铁沉积肝脏的磁共振$R_{2}^{*}$参数图像重建. 通过利用一种融合噪声校正物理模型和改进的全变分模型的损失函数来引导深度学习网络的自监督训练. 网络的训练不需要使用真实的$R_{2}^{*}$参数图像. 同时,相较于传统的参数估计算法,本文提出的方法能够快速准确地重建出铁沉积肝脏$R_{2}^{*}$参数图像,较好地抑制图像中噪声的影响, 校正噪声引起的偏差,同时保持$R_{2}^{*}$参数图像中的结构细节.

关键词: 磁共振成像, 参数图像重建, 深度学习, 自监督网络, 肝脏铁过载

Abstract:

Magnetic resonance (MR) effective transverse relaxation rate ($R_{2}^{*}$) technique has been widely applied for assessing hepatic iron concentration. However,$R_{2}^{*}$ mapping of iron-loaded liver can be severely degraded by noise. With the development of deep learning, deep neural networks have become effective tools for MR parameter mapping. In this study, a model-guided self-supervised deep neural network was designed for MR $R_{2}^{*}$ parameter mapping of iron-loaded liver. A novel loss function that integrated a noise-corrected physical model and an improved total variation model was used to train the network, which did not require reference $R_{2}^{*}$ parameter maps. Meanwhile, compared to the conventional parameter fitting methods, model-guided self-supervised deep learning method enabled accurate and efficient $R_{2}^{*}$ mapping of iron-loaded liver, suppressed the effect of noise, corrected the bias introduced by noise, and preserved the detailed structure of $R_{2}^{*}$ map.

Key words: MR imaging, parameter mapping, deep learning, self-supervised network, liver iron overload

中图分类号: