Application of online agglomerative hierarchical clustering on real dMRI

Abstract

Magnetic resonance imaging provides diffusion weighted images (DWI), which non-invasively reconstruct the brain white matter pathways through fiber tractography. Fiber clustering algorithms are used to identify anatomically meaningful fiber bundles. Most of the clustering schemes require a (dis)similarity matrix which contains pairwise fiber distances. Computation of the pairwise fiber distances has a quadratic complexity. Online clustering schemes do not require a computation of full pairwise fiber distances, hence the overall clustering computation time is reduced. In this experimental study, we proposed to use an online agglomerative hierarchical clustering algorithm to extract white matter fiber bundles from the whole brain fibers filtered by a spherical region of interest (ROI). This method requires an initialization of the cluster model using a relatively small set of fibers. After the initialization, cluster (re)assignment is performed using the cluster model by updating the model in certain conditions. The experiments are conducted on five different real DWI, for each a spherical ROI is located in different anatomical regions for filtering the whole brain fibers.