Quantitative Magnetic Resonance Imaging of Anisotropic Biological Tissues

磁共振成像(MRI)已成为研究中枢神经系统的有力工具。目前正在紧锣密鼓地开发针对生物组织特定成分的定量核磁共振方法。在过去的五年里，我和我的学生们研究了定量MRI方法，如髓鞘水成像或灌注图像，是如何依赖于组织组成、组织结构及其相对于主磁场的方向的。我们发现，血流灌注测量受到组织取向的强烈影响，因为大脑中大约一半的血液是在与大脑神经纤维束平行运行的血管中。根据MRI技术的不同，平行于磁场的白质组织和垂直于磁场的组织之间的效果差异高达100%。我们开发了数值方法，使我们能够确定大脑白质中的血液体积，正电子发射断层扫描(黄金标准)与我们的MRI方法之间具有高度的一致性。对于多回波自旋回波扫描测定的髓鞘含水率，我们还发现在最低值和最高值之间存在30%的系统取向效应。造成这种效果的原因尚不清楚。偶极-偶极相互作用以及髓鞘的磁化率可能对这种效应起到作用。在我接下来的五年研究计划中，我将继续研究这些定向效应，以帮助我们理解MRI信号的产生。我将分析在固定前后从无髓鞘的人类新生儿脑以及有髓鞘和脱髓鞘的死后组织中获得的髓鞘水数据，并开发数值组织模型。通过这项工作，我们将深入了解髓鞘在定向效应中的作用，以及在以前的尸检研究中，组织固定可能如何掩盖定向效应。此外，我们已经开始采用机器学习的方法，从核磁共振扫描中恢复有意义的信息。正常情况下，机器学习依赖于大量的训练数据。随着我们对磁共振图像生成的理解，我们能够计算任意数量的合成训练数据。在这种合成数据上训练的神经网络能够解决在真实数据中遇到的问题，通常不需要或最少地对真实数据进行额外训练。我们将训练两个神经网络，帮助我们克服障碍，使我们的定量方法获得更广泛的适用性。其中一个网络将根据传统的解剖扫描确定组织方向，这几乎是每个大脑MRI协议的一部分。另一个网络将学习数据的噪声分布，以便执行更好的参数推断。

Magnetic resonance imaging (MRI) has become a powerful tool for the investigation of the central nervous system. Intense efforts are underway to develop quantitative MRI methods specific for particular components of biological tissue. Over the past five years, my students and I have investigated how quantitative MRI methods, such as myelin water imaging or perfusion image depend on tissue composition and tissue architecture and its orientation relative to the main magnetic field. We found that perfusion measurements are strongly influenced by tissue orientation since about half of the blood in the brain is in blood vessels that run in parallel with the brain's nerve fibre tracts. Depending on the MRI technique, the effect is up to 100% difference between white matter tissue that runs parallel to the magnetic field and tissue that runs perpendicular to the magnetic field. We developed numerical methods that allowed us to determine the blood volume in the brain's white matter, with high agreement between positron emission tomography (the gold standard) and our MRI approaches. For the myelin water fraction determined with a multi echo spin echo scan, we also found a systematic orientation effect of 30% between the lowest values and the highest values. The reasons for this effect are not clear. Dipole-dipole interactions as well as the magnetic susceptibility of the myelin sheath could contribute to the effect. During the next five years of my research program I will continue to investigate these orientation effects in order to contribute to our understanding of MRI signal generation. I will analyse myelin water data acquired in the non-myelinated human newborn brain and in myelinated and demyelinated post mortem tissue before and after fixation and develope numerical tissue models. From this work we will gain insights into the role of myelin on the orientation effects and how tissue fixation may have obscured orientation effects in previous post mortem studies. Furthermore, we have embarked on machine learning approaches to recover meaningful information from MRI scans. Normally, machine learning relies on large amounts of training data. With our understanding of magnetic resonance image generation, we are in the position to compute arbitrary amounts of synthetic training data. Neural networks trained on such synthetic data are able to solve problems encountered in real data, often with no or minimal additional training on real data. We will train two neural networks that will help us overcome barriers to the broader applicability of our quantitative methods. One network will determine tissue orientation from a conventional anatomical scan that is part of almost every brain MRI protocol. The other network will learn the noise distribution of the data in order to perform much improved parameter inference.