无监督的磁共振图像重建方法研究进展

Research progress in unsupervised magnetic resonance image reconstruction methods

近年,深度学习技术在磁共振(magnetic resonance,MR)图像重建领域飞速发展。然而,由于有监督的MR图像重建方法所依赖的高质量配对MR数据难以获取,无监督的MR图像重建方法逐渐成为了研究者们关注的重点,并展现出巨大的应用前景。当前关于此类问题的研究层出不穷,但仍缺乏系统性的归纳和分析。为此,本文综述了无监督MR图像重建方法的研究进展。首先,本文对无监督的MR图像重建方法进行了总结,无监督的MR图像重建方法能够从图像域或K空间域数据学习先验信息,实现在缺少配对数据情况下的MR图像重建;其次,本文根据学习先验信息的作用域的不同,将这些方法分为基于K空间域、基于图像域和基于混合域的无监督MR图像重建方法,并重点对各类方法的算法模型和实现流程进行了详细的介绍。最后,本文对无监督MR图像重建领域的进展和各类方法的特点进行了较为全面的总结,并对未来的发展方向进行了展望,以期为实现无监督MR图像重建提供思路和参考,并促进MR成像的临床应用。

Recently,deep learning techniques have rapidly developed in the field of magnetic resonance(MR)image reconstruction.However,due to the difficulty of obtaining high-quality paired MR data for supervised MR image reconstruction methods,unsupervised MR image reconstruction methods have gradually become the focus of researchers and have shown great application prospects.Currently,there are numerous studies on unsupervised MR image reconstruction methods,but little systematic induction or analysis is provided.This article reviews the research progress of unsupervised MR image reconstruction methods.First,summary is conducted for unsupervised MR image reconstruction methods,which can learn prior information from image domain or K-space domain data to achieve MR image reconstruction without paired data.Second,this article divides these methods into methods based on K-space domain,image domain,and mixed domain,according to the prior learning domain,and introduces the algorithms and implementation processes of these methods in detail.Finally,this article provides a comprehensive summary of the progress in unsupervised MR image reconstruction and prospects for future directions,in order to providing ideas or references for achieving better unsupervised MR image reconstruction and promote clinical applications of MR imaging.