基于机器学习的多参数脑部定量磁化率成像研究

本项目旨在采用深度学习技术，根据脑部磁化率成像的机理，面向具体问题，设计三个专用的有特色的卷积神经网络：搭建集成相位解缠绕和背景场去除的双级递归残差网络，用于一步式相位预处理，有效避免误差累积，缩短计算时间；构造量化性优异的多尺度空间自适应网络，反演单方向磁化率标量值，解决组织对比度下降、细节纹理丢失、磁化率失真等重建难点；构建具有多通道输入和多通道输出的三维分支网络，利用磁化率成像独特的组织对比度，实现白质、灰质、深灰质核团、血管和脑脊液的同步分割和体视化。基于多方向多回波成像的相位和幅值数据，研究方向数和方向角受限的磁化率各向异性，分析不同回波时间的磁化率张量和弛豫张量的相关性，建立描述脑组织特性的多参数量化指标。在此基础上，结合分子影像学技术，研究基底节中的有髓鞘神经纤维，挖掘磁化率成像的应用潜力，探索深灰质核团纹苍黑纤维束的活体无损检测新方法。

This project aims to using the deep learning, based on the mechanism of brain susceptibility imaging, design three application-specific convolutional neural networks with sophisticated architectures. The first recursive residual network contains two subnets in a cascade style, integrating the tasks of phase unwrapping and background field removal, so that the phase preprocessing can effectively avoid the error accumulation and decrease the calculation time. The second multi-scale spatial adaptive network is designed to reconstruct the tissue susceptibility value at single head orientation with excellent quantification performance, meanwhile to protect the brain tissue contrast, fine structure details, as well as the susceptibility accuracy. Taking advantage of the unique tissue contrast in susceptibility imaging, the third one is a three-dimensional branch network with multi-channel input and multi-channel output, constructed for brain tissue segmentation and rendering, i.e. the white matter, gray matter, deep gray nucleus, vessel and cerebrospinal fluid. The limited number of head rotation data and the limited rotation angle range lead to the condition number of the system matrix to be insufficiently small to achieve a stable solution of the susceptibility anisotropy. Hence, the susceptibility tensor and the relaxation tensor at different echo times are exploited to investigate the brain tissue structure, using the phase and magnitude data acquired by multi-echo GRE sequence at multiple head orientations. To reveal the brain tissue characteristics, a quantitative evaluation system is established including multi-parameter indicators. Then, the myelinated nerve fibers in the basal ganglia are investigated to explore the great potential of susceptibility imaging in brain science research, while the molecular imaging technique is employed to provide a reference for the fiber bundle information. We intend to propose a new non-invasive in vivo detection approach for the fiber pathway in deep gray matter.