Novel Deep Learning Approaches for Analyzing Diffusion Imaging Data

近年来，扩散成像(DI)迅速发展成为临床脑研究中最重要的无创性工具之一。然而，由于所需获取的梯度方向的数量，测量时间较长，导致DI在临床实践中很少使用。为了克服这个问题，最近的方法证明了机器学习的力量，特别是深度学习(DL)，它能够非常准确地描述和重建组织潜在的复杂功能，而只需要很少的梯度方向。因此，扫描时间可以大大减少。然而，为了使DL在临床DI中的最佳适用性，确定了四个主要障碍，这些障碍将在本项目中得到解决。最大的障碍是来自不同MRI系统的数据之间的巨大差异。为了克服这一障碍，将比较现有的不同MRI系统的协调方法，并开发出一种最优的MRI信号协调方法。其次，解决了在DI中难以获得的地面真实数据的需求，这使得DL方法的训练变得复杂。为了解决这个问题，将开发一个读入数据集的框架，以确定重要的扩散特征和统计数据。随后，可以基于该信息合成各个扩散数据，从而可以合成完整的扩散数据集。所得到的数据及其对应的地面真实数据可以在以后的训练过程中用于改进DL模型的性能，此外，由于在常规重建方法中很少使用而在采集过程中通常被丢弃的复信号被整合到使用新的DL方法的重建中。研究表明，复杂的MRI信号携带着重要的组织信息，因此可以在DL网络内重建时用作额外信息。为此，需要开发能够处理复杂信号的新的下行链路组件。在本项目结束时，重点在于每个体素与角度相关的扩散信号。以前的DL方法目前不能将这种额外的球面信息合并到处理中，这就是为什么需要新的方法来将先前的DL元素转移到球面上并将它们链接到正常的DL元素。在整个项目的前半部分，将在不同的位置采集MRI数据，包括其相位数据、高数量的梯度方向和高分辨率，以评估所描述的所有方法。这个项目和由此产生的方法的目的是在保持相同准确性的同时，显著减少临床实践中扩散成像序列的扫描次数。

Recently, diffusion imaging (DI) rapidly developed into one of the most important non-invasive tools for clinical brain research. However, long measurement times, due to the required number of acquired gradient directions, result in a rare usage of DI in clinical practice. To overcome this problem, recent methods demonstrated the strength of machine learning and in particular deep learning (DL), which is able to describe and reconstruct the tissue's underlying complex functions very accurately while only few gradient directions are required. Thus, scanning time can be greatly reduced.For an optimal applicability of DL in clinical DI, however, four major obstacles were identified which will be addressed within this project. The biggest barrier is the large variance between data from different MRI systems. To overcome this barrier, existing methods for harmonizing different MRI systems will be compared and an optimal method for harmonizing MRI signals will be developed.Next, the need of ground truth data is addressed, which is difficult to obtain in DI, complicating the training of DL methods. To solve this problem, a framework will be developed that reads in a dataset, to determine important diffusion characteristics and statistics. Subsequently, individual diffusion data and thereby a complete diffusion dataset can be synthesized based on this information. The resulting data and its corresponding ground truth can later be used during training to improve the DL model’s performance.Furthermore, complex signals, which are commonly discarded during acquisition, due to their rare usage in regular reconstruction methods, are integrated into the reconstruction utilizing novel DL methods. Studies have shown that complex MRI signals carry important tissue information, which could therefore be used as additional information during reconstruction within DL networks. For this purpose, new DL components that are capable of processing complex signals need to be developed. At the end of this project, the focus lies on the angle-related diffusion signals per voxel. Previous DL methods are currently not able to incorporate this additional spherical information into the processing, which is why new methods are needed that transfer the previous DL elements onto a sphere and link them to normal DL elements. In this way, neighboring information within the signal as well as between signals can be included to ensure optimal reconstruction.Throughout the first half of the project, MRI data, including its phase data, a high number of gradient directions and a high resolution will be acquired at various locations to evaluate all the methods described. The aim of this project and the resulting methods is to significantly reduce the scan times for diffusion imaging sequences in clinical practice while maintaining the same accuracy.