A novel framework for the study of neural architectures in the human brain with diffusion MRI
Diffusion Magnetic Resonance Imaging (diffusion MRI) is a relatively recent imaging technique, which has allowed the study of nervous fibers and their connectivity in the white matter of the human brain in vivo. It is based on the analysis of the directions of diffusion of water molecules inside the neural axons surrounded by myelin coats, which is related to the orientations of neural fibers themselves. This phenomenon has been used in medical imaging since the mid nineties, with the appearance of the first works on Diffusion Tensor Imaging (DTI). This technique is based on the premise that diffusion may be described in terms of a Gaussian process, which, as it has been widely reported, is strictly valid only for very simple neural configurations.
With the advent of new hardware and protocols for the accelerated or noise--reduced acquisition of MRI data sets, new techniques for image analysis and processing are possible. Among them, High Angular Resolution Diffusion Imaging (HARDI), based on the dense sampling of the diffusion signal for all possible orientations, is of special interest in this dissertation, since it allows the accurate characterization of the diffusion process. At the same time, the wide variety of scenarios and features for such data sets raises a number of new challenging problems of paramount importance, since the traditional foundations of processing algorithms for diffusion MRI, such as the Rician nature of noise in scanned data, or the Gaussian model for diffusion, are no longer adequate. These problems are currently a very active research field and the motivation for the present work.
In this thesis, a new statistical and probabilistic framework is proposed to cope with this diversity, whose ultimate goal is the description of complex micro-architectures of neural fibers through the characterization of the macroscopic diffusion process. It comprises specific tools for the statistical characterization (signal modeling and noise description and estimation) and conditioning (denoising by means of statistics-based image filtering) of diffusion data sets provided by MRI scanners. The novelty of the proposed methods relies on their compatibility with standard scenarios based on Rician signals, but also with modern systems like multiple-coils scanners and parallel MRI acquisition schemes, for which each part of the image is acquired in parallel by an independent receiving coil and further combined with all remaining parts to obtain the whole data set.
The center of the work carried out (and the main contribution of the thesis), however, is the inference of probabilistic information relative to fiber populations in the white matter of the human brain. The techniques described are exclusively based on HARDI data sets, so they perfectly fit currently existing acquisition protocols without further restrictions. The proposed algorithms aim to obviate any specific parametric model for diffusion, completely avoiding the Gaussian assumption for the process of diffusion of water molecules. It is demonstrated that the use of true probabilistic information clearly benefits the inference of fiber populations and orientations, and the methods developed highly improve the previous attempts in the related literature.
The quantitative evaluation of the entire framework is carried out over synthetic and phantom data specifically designed for each part, comparing the results to related works in the recent literature. Besides, extensive experiments over real data sets, gathered from several institutions worldwide, are presented to illustrate the work developed. The features of these data are diverse enough to derive general conclusions on the performance of the new methodology.
