Parcelling Subjects From the Human Connectome Project

We next applied our parcellation technique to the HCP data. First, we performed Constrained Spherical Deconvolution (CSD) based tractography [20] from a dense set of points in the cortex. Then, we used our technique to parcellate the cortex by clustering the tractograms. Specifically, since each subject has a surface representing their gray-matter/white-matter interface, we used their vertices as seeds to create tractograms. To avoid superficial cortico-cortical fibers [18], we shrank each of the 66 surfaces 3 mm into the white matter. For each subject, we fitted a CSD model [20] to their diffusion data using Dipy (version 0.11) [6] and created 15,000 streamlines per seed-voxel using the implementation of probabilistic tractography in Dipy. Later, we created a tractogram as in [Eq. (2)] by calculating the fraction of particles that visited each white-matter voxel. Then, we transformed each tractogram with the logit function [17] as in Eq. (4). We clustered the tractograms of each subject using the modified AHC algorithm while imposing a minimum cluster size of 3 mm2 in the finest granularity.

To create the groupwise parcellation, we took advantage of the vertex correspondence across subjects in the HCP data set. Since we are in a vectorial space we calculated the average tractogram of each seed. Then, we created the groupwise parcellation by clustering the average tractograms with our proposed technique (Sect. 2.3). The resulting dendrogram for the groupwise case, alongside some of the obtainable parcellations, are in Fig. 3.

Clustering performance for different levels of variability

Fig. 2 Clustering performance for different levels of variability. SNR comprises both intra-cluster and across-subject variability, at smaller SNR more variability is present. Left, best overlap between synthetic regions and clusters in the dendrogram. Right, best corrected rand coefficient for horizontal cutting. The coefficients close to 1. specially in the average case (dark blue line), show the compatibility between our parcellation technique and the proposed model

Groupwise dendrogram created by our technique

Fig. 3 Groupwise dendrogram created by our technique. We can retrieve different granularities of the same parcellation by choosing cutting height, shown as a red dotted line. Motor and Sensory cortex appear at coarse levels of granularity

< Prev   CONTENTS   Source   Next >