A Comparison of Metrics and Algorithms for Fiber Clustering

Abstract

Diffusion-weighted Magnetic Resonance Imaging (dMRI) can unveil the microstructure of the brain white matter. The analysis of the anisotropy observed in the dMRI contrast with tractography methods can help to understand the pattern of connections between brain regions and characterize neurological diseases. Because of the amount of information produced by such analyses and the errors carried by the reconstruction step, it is necessary to simplify this output. Clustering algorithms can be used to group samples that are similar according to a given metric. We propose to explore the well-known clustering algorithm k-means and a recently available one, Quick Bundles [1]. We propose an efficient procedure to associate k-means with Point Density Model, a recently proposed metric to analyze geometric structures. We analyze the performance and usability of these algorithms on manually labeled data and a database a 10 subjects.