基于深度学习图像重建的磁共振并行成像
Deep Learning Image Reconstruction for Parallel
Magnetic Resonance Imaging
答 辩 人：路铁源
指导老师：屈小波 教授/博导
答辩委员会主席：蔡聪波 教授 博导 厦门大学电子科学系
答辩委员会成员：包立君 副教授 硕导 厦门大学电子科学系
杨 钰 助理教授 硕导 厦门大学电子科学系
答辩秘书： 王 新 高级工程师 厦门大学电子科学系
时间：2020年05月20日上午9:00
地点：厦门大学海韵园物理与机电大楼307
----------------------------------------------------------------------------------------------------------------
摘要
并行成像是加速磁共振成像数据采集的常用方法，可以适用于目前已有的序列且不会影响图像的对比度，同时也不会引入过多的伪影。但是，并行磁共振成像的加速倍数受线圈数目的限制，引入压缩感知中的随机采样和稀疏图像重建是加速并行成像和提高图像质量的重要手段。在快速成像中，欠采样的多通道 k 空间数据需要通过求解图像稀疏重建的最优化问题得到完整的磁共振图像。然而，求解这些问题的数值算法通常比较耗时，需要经过成百上千次迭代才能得到稀疏重建模型的最优解。因此，如何快速重建高质量图像依然是亟待解决的重要问题。
近来，深度学习凭借深层的网络结构和硬件的强大计算能力，被研究人员引入到快速磁共振成像的图像重建中，实现了比传统压缩感知方法更快的重建和更低的重建误差。但是对于磁共振并行成像而言，现有的深度学习磁共振重建方法的加速倍数有待提高，网络也缺乏较好的可解释性。如何在高加速成像倍数下进行解释性强的磁共振图像重建是本学位论文的研究重点。
本文首先回顾了磁共振并行成像、压缩感知和深度学习的前沿进展，然后提出了一种基于深度学习的并行磁共振图像重建网络。受快速迭代软阈值投影算法的迭代求解过程启发，所提重建网络结构由固定数目的迭代块级联而成，每一个迭代块均包含数据校验模块和网络学习模块。在网络学习模块中，还引入了残差结构来提升网络的学习能力。所提方法与最新的深度学习方法在公开的膝盖数据集中相比，能够以重建速度不超过 1 秒每张的前提下，获得最低的重建误差和临床医生盲评的最高图像质量。此外，还将所提深度学习网络扩展到多对比度磁共振图像的联合重建上，实验结果表明，重建图像误差低于单对比度图像重建误差。
关键词：深度学习；快速成像；图像重建；磁共振成像
ABSTRACT
Parallel imaging is a widely used method for accelerating magnetic resonance imaging data acquisition. It can be applied to existing sequences without affecting thecontrast of the image and introducing too many artifacts. However, the acceleration factor of parallel magnetic resonance imaging is limited by the number of receiving coils. Random sampling and sparse image reconstruction in compressed sensing are essential means to accelerate parallel magnetic resonance imaging and improve image quality. In fast imaging, obtaining a complete magnetic resonance image needs to solve the optimization problem of image sparse reconstruction with undersampled multi channel k-space data. However, numerical algorithms are usually time-consuming and require hundreds or thousands of iterations to obtain the optimal solution for the sparse reconstruction model. Therefore, how to quickly reconstruct high-quality images is still an important problem to be solved urgently.
Recently, due to the deep network structure and powerful computing capabilities of the hardware, deep learning has been introduced to the image reconstruction in fast magnetic resonance imaging, which achieves faster reconstruction speed and lower reconstruction error than traditional compressed sensing methods. However, in parallel magnetic resonance imaging, the acceleration factor in the existing deep learning magnetic resonance reconstruction method needs to be improved, and the network also lacks good interpretability. How to perform highly interpretable magnetic resonance image reconstruction at a high acceleration factor is the research focus of this thesis.
In this thesis, we first review the frontier progress of parallel magnetic resonance imaging, compressed sensing and deep learning, and then propose a parallel magnetic resonance image reconstruction network based on deep learning. Inspired by the iterative solution process of the projected fast iterative soft-thresholding algorithm, the proposed reconstruction network structure is formed by cascading a fixed number of iterative blocks, and each iterative block includes a data consistency module and a network learning module. In the network learning module, a residual structure is also introduced to improve the learning ability of the network. In the reconstruction of a public knee data set, compared with the latest deep learning methods, the proposed method can obtain the lowest reconstruction error and the highest image quality blindly evaluated by clinicians under the premise that the reconstruction speed does not exceed 1 second per slice. In addition, the proposed deep learning network is extended to the joint reconstruction of multi-contrast magnetic resonance images. The experimental results show that the reconstruction error is lower than the single-contrast image reconstruction error.
Key words:
Deep Learning; Fast Imaging; Image Reconstruction; Magnetic Resonance Imaging
欢迎广大师生莅临指导！
