Algorithms and methods for the study of brain connectivity based on dMRI.

Abstract

A main goal for understanding how the brain works is the description of the network of brain connections or the human brain connectome. The basic elements of the connectome are regions of gray matter (GM) formed by neuronal brain bodies (somas) and the connections (axons) between them formed by the white matter (WM) neuronal fibers. Currently, diffusion-weighted Magnetic Resonance Imaging (dMRI) techniques with High Angular Resolution Diffusion Imaging (HARDI) have improved the quality of tractography datasets concerning Diffusion Tensor Imaging (DTI). Tractography data are complex and contain noise and artifacts, so they require computational methods capable of processing them efficiently and extract useful information. Therefore, this thesis proposes the development and implementation of several methods to process the brain tractography data and through them provide analysis tools for study the brain structural connectivity. Among them, we contributed to the improvement of a clustering of white matter fibers, to the labeling of the superficial white matter bundles, to the development of a method for the parcellation of the cortical surface from short and long segmented bundles for a group of subjects, and to the development of complementary methods to carry out the individual parcellation. We collaborated in the development of an efficient clustering of white matter fibers that was evaluated in terms of quality and execution time against other state-of-theart methods, giving as a result 8.6 times more speed than the most efficient method. Moreover, we created a method which performs intra-subject labeling of superficial white matter fibers in 3.6 min, and two inter-subject labeling methods. One is based on matching and obtains good correspondence but little reproducibility with an execution time of 96 s. The other one focused on clustering that obtains good correspondence and reproducibility among subjects achieving a short execution time of 9 s for a subject. On the other hand, a method for cortical surface parcellation which creates parcel atlases was developed. Then, two generated parcellations were compared with stateof-the-art methods, finding a degree of similarity with dMRI, functional, anatomical, and multi-modal atlases. The best comparison was between our parcellation composed of 185 sub-parcels and another dMRI-based parcellation, obtaining 130 parcels in common for a Dice coefficient ≥0.5. The parcellation composed of 160 parcels achieves a reproducibility across subjects of ≈0.74, based on the average Dice’s coefficient between subject’s connectivity matrices, rather than ≈0.73 obtained for a macroanatomical parcellation of 150 parcels. In addition, two complementary methods were developed to perform individual cortical parcellations, one based on clustering and the other based on the geodesic distance, the latter obtains a parcellation of 350 parcels in 18 s for the atlas-based mode and 82 s for the whole cortex mode. Moreover, it obtains better reproducibility against two state-of-the-art methods with a difference of 0.024 and 0.043 according to the Dice coefficient. This thesis contributes to the development of efficient computational methods for the study of brain connectivity that can be applied to high-quality and large databases, capable of dealing with the noise present in the tractographies. In addition, thanks to this research, these methods can be used by neuroscientists, neuroanatomists, and neurologists to study and develop new studies of brain connectivity and to obtain more answers about the structure of the brain and its connectivity