Reconstruction of robust, high SNR functional brain images using machine learning and oscillatory steady state MRI

项目摘要/摘要： 功能磁共振成像(FMRI)是临床和科学研究的重要工具， 具有从认知神经科学到术前规划的广泛应用。FMRI可以生成完整的 脑神经元激活图，但它是一种固有的低信噪比(SNR)方法，因为 激活信号相对于基线信号的相对较小的变化。这个项目的首要目标是 是通过使用一种新的采集方法开发和验证健壮的、高信噪比的功能磁共振成像，振荡稳态 (OSS)功能磁共振成像，以及机器学习(ML)技术，以重建OSS功能磁共振成像的高质量图像 数据。与当前最先进的fMRI采集相比，OSS fMRI可以将SNR提高2倍 方法，这种提升大致相当于从3T磁共振扫描仪到7T磁共振扫描仪的SNR增益 扫描仪。信噪比的增加是稳态方法的直接结果。然而，对于一个更复杂的， 振荡采集，OSS fMRI可能比传统方法更容易受到常见MRI伪影的影响，如果 原封不动。MRI伪影的两个最常见来源是主磁场(B0)的变化 由于生理噪音(如呼吸)和病人的运动。该项目将开发强大的，高信噪比 通过将(1)B0变化和(2)患者运动的影响合并到OSS信号模型和 图像重建算法。一种结合了神经网络的新图像重建将纠正 B0波动并消除B0引起的伪影。我们将训练神经网络来生成B0场地图，使用 来自传统的、基于物理的双回波场映射技术的数据，以隐含地合并先前 将物理信息转化为重建，同时还提供了一种快速的、基于ML的场映射方法。 为了校正对象的运动，我们将开发一个神经网络来估计刚体运动参数 来自fMRI扫描中的连续图像帧。这些运动参数将在迭代图像中使用 重建以产生高质量的静息状态和基于任务的功能磁共振成像，即使在受试者在场的情况下也是如此 动议。这项提案中开发的技术将导致改进的功能磁共振成像，具有潜在的 以显著推进人脑研究和神经疾病的治疗。 密歇根大学是美国顶尖的研究型大学之一，提供了理想的环境 和基础设施，以完成拟议的研究战略。功能磁共振 UM的实验室和电气工程和计算机科学系拥有所有必要的 该项目所需的硬件和计算资源，包括两个最先进的GE 3T MRI 扫描仪和广泛的GPU硬件用于该项目的机器学习组件。此外，戴维斯博士还指出， 杰弗里·费斯勒和道格拉斯·诺尔在fMRI图像采集和重建方面也拥有成熟的专业知识 作为广泛的指导经验，这将有助于指导这个项目和我的培训。

Project Summary/Abstract: Functional magnetic resonance imaging (fMRI) is an important tool both clinically and in scientific research, with broad applications ranging from cognitive neuroscience to presurgical planning. FMRI can generate whole brain neuronal activation maps, yet it is an inherently low signal to noise ratio (SNR) method due to the relatively small changes in activation signal relative to the baseline signal. The overarching goal of this project is to develop and validate robust, high SNR fMRI by using a novel acquisition method, oscillating steady state (OSS) fMRI, along with machine learning (ML) techniques, to reconstruct high-quality images of the OSS fMRI data. OSS fMRI can increase SNR by >2x compared to the current state-of-the-art in fMRI acquisition methods, and this boost is roughly equivalent to the SNR gain going from a 3T MRI scanner to a 7T MRI scanner. The SNR increase is a direct result of the steady state approach. However, with a more complex, oscillating acquisition, OSS fMRI can be more susceptible to common MRI artifacts than traditional methods if left uncorrected. Two of the most common sources of MRI artifacts are changes in the main magnetic field (B0) due to physiological noise (such as respiration) and patient motion. This project will develop robust, high SNR fMRI at 3T by incorporating the effects of (1) B0 changes and (2) patient motion into the OSS signal model and image reconstruction algorithms. A new image reconstruction that incorporates neural networks will correct for B0 fluctuations and remove B0 induced artifacts. We will train a neural network to generate B0 field maps using data from a conventional, physics-based two echo field mapping technique to implicitly incorporate prior information of the physics into the reconstruction, while also providing a fast, ML-based field mapping method. To correct for subject motion, we will develop a neural network that estimates rigid-body motion parameters from sequential image frames in the fMRI scan. These motion parameters will be used in an iterative image reconstruction to produce high-quality resting-state and task-based fMRI, even in the presence of subject motion. The technology developed in this proposal will result in improved functional MRI that has the potential to significantly advance both the study of the human brain and the treatment of neurological disorders. The University of Michigan is one of the top research universities in the US, and provides an ideal environment and infrastructure to complete the proposed research strategy. The Functional Magnetic Resonance Laboratory and the Electrical Engineering and Computer Science Department at UM have all the necessary hardware and computational resources needed for this project, including two state-of-the-art GE 3T MRI scanners and extensive GPU hardware for the machine learning components of the project. Furthermore, Drs. Jeffrey Fessler and Douglas Noll have proven expertise in fMRI image acquisition and reconstruction, as well as extensive mentorship experience, that will help guide this project and my training.