Novel Deep Learning Approaches for Analyzing Diffusion Imaging Data

Abstract Diffusion magnetic resonance imaging (dMRI) is developing into one of the most important non-invasive tools for clinical brain research. This development is supported by a project funded by the German Research Foundation, in which four major obstacles related to dMRI data were addressed: (1) the lack of transferability of dMRI data between clinical sites, (2) the lack of training and label data, (3) the potential of complex diffusion data, and (4) the integration of spherical signals in neural networks to improve accuracy. To overcome the problem of different MRI systems producing slightly varying data, the project developed a method for harmonizing MRI signals. To address the issue of limited ground truth data, a framework was developed to synthesize individual diffusion data and complete datasets based on important diffusion characteristics and statistics. The integration of complex signals, often discarded during acquisition, to improve reconstruction was also explored. Finally, new methods were developed to preserve the spherical character of the diffusion data in the DL model. The resulting methods are intended to improve the usability of diffusion imaging data and to enable the creation of processing pipelines for dMRI data in clinical studies and clinical practice.