Tractography Mapping for Dissimilarity Space across Subjects

Abstract

Structural brain connectivity can be studied with the help of diffusion magnetic resonance imaging (dMRI), through which the pathways of the neuronal axons of the white matter can be reconstructed at the millimeter scale. Such connectivity structure, called deterministic tractography, is represented as a set of polylines in 3D space, called streamlines. Streamlines have a non-homogeneous number of points and, for this reason, the dissimilarity representation (DR) has been proposed as accurate Euclidean embedding. By providing a vectorial representation of the streamlines, DR enables the use of most machine learning and pattern recognition algorithms for connectivity analysis. However, the DR is subject-specific and thus applies only to intra-subject analysis, while neuroscientific studies often address inter-subject comparisons. For this reason, in this work, we propose an algorithmic solution to build a common vectorial representation for streamlines across subjects. The core idea is based on finding a small set of corresponding streamlines, a problem known as streamline mapping. With experiments on a task of segmentation, we show that the quality of alignment of tractographies, through the common vectorial representation, is even superior to that of the traditional linear registration.