A Connectome-Based Comparison of Diffusion MRI Schemes

Abstract

Diffusion MRI has evolved towards an important clinical diagnostic and research tool. Though clinical routine is using mainly diffusion weighted and tensor imaging approaches, Q-ball imaging and diffusion spectrum imaging techniques have become more widely available. They are frequently used in research-oriented investigations in particular those aiming at measuring brain network connectivity. In this work, we aim at assessing the dependency of connectivity measurements on various diffusion encoding schemes in combination with appropriate data modeling. We process and compare the structural connection matrices computed from several diffusion encoding schemes, including diffusion tensor imaging, q-ball imaging and high angular resolution schemes, such as diffusion spectrum imaging with a publically available processing pipeline for data reconstruction, tracking and visualization of diffusion MR imaging. The results indicate that the high angular resolution schemes maximize the number of obtained connections when applying identical processing strategies to the different diffusion schemes. Compared to the conventional diffusion tensor imaging, the added connectivity is mainly found for pathways in the 50–100mm range, corresponding to neighboring association fibers and long-range associative, striatal and commissural fiber pathways. The analysis of the major associative fiber tracts of the brain reveals striking differences between the applied diffusion schemes. More complex data modeling techniques (beyond tensor model) are recommended 1) if the tracts of interest run through large fiber crossings such as the centrum semi-ovale, or 2) if non-dominant fiber populations, e.g. the neighboring association fibers are the subject of investigation. An important finding of the study is that since the ground truth sensitivity and specificity is not known, the comparability between results arising from different strategies in data reconstruction and/or tracking becomes implausible to understand.

The influence of construction methodology on structural brain network measures: A review

“…However, the difference between 12 and 32 directions is not significant. It is suggested that the effect of an increasing number of gradient directions is limited (Vaessen et al, 2010;Gigandet et al, 2013). Similar trends are observed by Gigandet et al (2013) for DSI.…”

“…The number of nodes in structural brain networks can vary in the range of 100 to 100,000, with various brain templates and refinement schemes (Hagmann et al, 2008;Zalesky et al, 2010;Bassett et al, 2011;Gigandet et al, 2013). The brain parcellation templates can be divided into two categories: predefined anatomical templates and random templates.…”

“…Table 1 The most prevalent brain parcellation templates and associated number of nodes adopted by the studies of structural brain networks. Gigandet et al (2013) FreeSurfer gyral-based template and refinement 66; 1015 Lemkaddem et al (2014) FreeSurfer gyral-based template 84 Hagmann et al (2008) White matter/gray matter boundary surface 66; 998 Bastiani et al (2012) White matter/gray matter boundary surface 4000 (2) Random templates Skudlarski et al (2008) Voxel-based template 5000 Glasser et al (2013) Subcortical and cerebellar voxels, and cortical surface triangles 90,000 Besson et al (2014b) Cortical and subcortical surface model 58,880 triangles…”

“…(3) DSI (Wedeen et al, 2005;Wedeen et al, 2008) is the most complex acquisition scheme, in which hundreds of diffusion sensitive gradients with different amplitudes (max. b-value even as high as 8000 s/mm 2 (Gigandet et al, 2013)) and directions (258 or more) are applied, typically in a Cartesian grid. DSI has a long acquisition time (>35 min).…”

“…This result is not in agreement with that of Zalesky et al (2010), probably because the smaller number of gradients and smaller b in HARDI (82 directions and b = 1500 s/mm 2 ) are used in this study. Gigandet et al (2013) reported that DSI leads to the highest connectivity density, HARDI leads to an intermediate one, while DTI generates the lowest one. This trend is in accordance with the streamline (geodesic) length distributions of various schemes: DSI shows heavier tails in the distribution and provides more streamlines with lengths of 50-100 mm.…”