Feature Extraction for Visual Analysis of DW-MRI Data

Abstract

Diffusion Weighted Magnetic Resonance Imaging (DW-MRI) is a recentmodality to investigate the major neuronal pathways of the humanbrain. However, the rich DW-MRI datasets cannot be interpreted withoutproper preprocessing. In order to achieve understandablevisualizations, this dissertation reduces the complex data to relevantfeatures.<br>The first part is inspired by topological features in flow data. Novelfeatures reconstruct fuzzy fiber bundle geometry from probabilistictractography results. The topological properties of existing featuresthat extract the skeleton of white matter tracts are clarified, andthe core of regions with planar diffusion is visualized.<br>The second part builds on methods from computer vision. Relevantboundaries in the data are identified via regularized eigenvaluederivatives, and boundary information is used to segment anisotropyisosurfaces into meaningful regions. A higher-order structure tensoris shown to be an accurate descriptor of local structure in diffusiondata.<br>The third part is concerned with fiber tracking. Streamlinevisualizations are improved by adding features from structural MRI ina way that emphasizes the relation between the two types of data, andthe accuracy of streamlines in high angular resolution data isincreased by modeling the estimation of crossing fiber bundles as alow-rank tensor approximation problem.