Conference PaperPDF Available

Investigating the mechanisms of structural connectome construction on HCP data: implications for structural network analysis

Conference Paper

Investigating the mechanisms of structural connectome construction on HCP data: implications for structural network analysis

Abstract

Diffusion MRI streamlines tractography is the main in vivo technique for inferring structural brain connectivity. Ideally, during structural connectome construction, each streamline should connect exactly 2 brain regions-of-interest (i.e. network nodes); the efficacy of identifying such 'pairwise' connectivity has however been shown to be heavily affected by the mechanisms used for streamline termination, brain parcellation, and assigning streamlines to nodes. In particular, the ill-posed nature of track termination criteria in many tractography algorithms can cause streamlines apparently being associated with either zero or only one grey matter (GM) regions (up to ~50–80%), or non-GM regions within nodes. Such apparent connectivity is not biologically meaningful, and could result therefore in a misleading connectome in non-trivial ways; however, the actual influence of node assignment strategies, and their interactions with track terminations and parcellation schemes, on subsequent connectomic measures has not yet been determined. This study extends our previous findings and investigates: (a) whether problems associated with the mechanisms for constructing pairwise connectivity can be rectified by using high quality MRI datasets such as those from the Human Connectome Project (HCP); and (b) the practical consequences of these construction mechanisms on graph theory metrics using state-of-the-art methods: the multi-shell multi-tissue (MSMT) constrained spherical deconvolution (CSD) approach for improved estimates of fibre orientation distribution functions (FODs); anatomically-constrained tractography (ACT) for meaningful track terminations; and the spherical-deconvolution informed filtering of tractograms (SIFT2) technique for quantitative tractogram reconstructions.
Investigating the mechanisms of structural connectome construction on HCP data: implications for structural network analysis
Chun-Hung Yeh, Robert E. Smith, Thijs Dhollander, Fernando Calamante, Alan Connelly
The Florey Institute of Neuroscience and Mental Health, University of Melbourne, Melbourne, Victoria, Australia
Tar g et Au d ie nc e: This study is intended for those interested in quantification of tractogram-based structural connectivity and connectomics.
Purpose: Diffusion MRI streamlines tractography is the main in vivo technique for inferring structural brain connectivity. Ideally, during structural connectome
construction, each streamline should connect exactly 2 brain regions-of-interest (i.e. network nodes); the efficacy of identifying such pairwiseconnectivity has however
been shown to be heavily affected by the mechanisms used for streamline termination, brain parcellation, and assigning streamlines to nodes [1]. In particular, the ill-
posed nature of track termination criteria in many tractography algorithms can cause streamlines apparently being associated with either zero or only one grey matter (GM)
regions (up to ~5080% [2,3]), or non-GM regions within nodes. Such apparent connectivity is not biologically meaningful, and could result therefore in a misleading
connectome in non-trivial ways; however, the actual influence of node assignment strategies, and their interactions with track terminations and parcellation schemes, on
subsequent connectomic measures has not yet been determined. This study extends the finding of [1] and investigates: (a) whether problems associated with the
mechanisms for constructing pairwise connectivity can be rectified by using high quality MRI datasets such as those from the Human Connectome Project (HCP) [4]; and
(b) the practical consequences of these construction mechanisms on graph theory metrics using state-of-the-art methods: the multi-shell multi-tissue (MSMT) constrained
spherical deconvolution (CSD) [5] approach for improved estimates of fibre orientation distribution functions (FODs); anatomically-constrained tractography (ACT) [6]
for meaningful track terminations; and the spherical-deconvolution informed filtering of tractograms (SIFT2) [7] technique for quantitative tractogram reconstructions.
Methods: (A) HCP data: Pre-processed T1s (0.7×0.7×0.7 mm3) and DWIs (1.25×1.25×1.25 mm3, 18/90/90/90 directions at b=0/1000/2000/3000 s/mm2) of 10 healthy
subjects from a Siemens 3T Connectome Skyra system were downloaded from the ConnectomeDB (http://db.humanconnectome.org) [4]. (B) Tractogram reconstruction:
FODs were computed using MSMT-CSD in which tissue response functions for GM, white matter (WM), and cerebrospinal fluid (CSF) were estimated from the multi-
shell DWIs based on T1 image segmentation [5]. For each subject, tractograms of 10 million streamlines were generated through seeding from a WM mask, tracking either
with or without ACT using the probabilistic iFOD2 algorithm [8]. The FOD threshold for terminating tracking was reduced to 0.06 (default=0.1) due to the benefit of
MSMT-CSD [5]. (C) Connectome construction: Two popular parcellation schemes were used to define network nodes: (i) the T1-based FreeSurfer (FS) parcellation [9]
pre-processed by the HCP analysis pipeline; (ii) the AAL atlas [10] transformed onto individual’s T1 space. Streamlines were assigned to nodes by: (i)end voxels =
voxels at streamline endpoints; or (ii) local search= a search from track endpoints to locate the nearest nodes within a 2-mm radius [11]. In order to make the quantification
of streamline connectivity biologically meaningful, SIFT2 [7] was used to modulate the contribution weight of each streamline to the relevant connectome edge. (D)
Network analysis: Brain Connectivity Toolkit [12] and two-sample t-test were used to statistically compare the outcomes of some key weighted connectomic metrics.
Results: Fig. 1 shows the frequency of identifying two nodes per streamline for the eight connectome construction strategies as described in Methods, demonstrating that
even on the HCP data, the construction of pairwise connectivity was strongly dependent on the connectome construction mechanisms. This highlights that advanced
methodologies are necessary, and that improved data quality is not in itself sufficient to achieve this end. Tabl e 1 shows the outcomes of global network metrics across 10
subjects. The majority of network metrics were altered significantly by a change in the mechanism used to assign streamlines to nodes for both parcellation schemes,
whether with or without ACT. Additionally, the majority of network metrics were altered significantly between non-ACT and ACT for both parcellation schemes, regardless
of streamline assignment mechanism used.
Discussion: Our results demonstrate that fundamental connectome characteristics can be heavily influenced by both the mechanism of streamlines termination, and the
mechanism used to assign streamlines to nodes associated with the brain parcellation. These processes are frequently overlooked, but they are in practice non-trivial and
can alter connectomic measures significantly, even on the state-of-the-art HCP datasets with the application of advanced data processing techniques. The connectome
constructed based on a heuristic mechanism that discards the majority of streamlines (~5080% [2,3]) is unlikely to provide an appropriate representation of the underlying
fibre connectivity. By contrast, the advanced mechanisms used here are specifically tailored to correct for inaccuracies of connectome reconstruction, i.e. ACT to forbid
erroneous streamline terminations and local search to account for small inconsistencies between terminations and parcellations; this is supported by the increased
frequency of successfully assigning each streamline to exactly two nodes when these methods are used in conjunction, independently of parcellation scheme used (Fig. 1:
~85% for both FS+ACT & AAL+ACT usinglocal search’). Following this logic, the connectomic metric values derived using these advanced mechanisms should be in
principle the most trustworthy results among all strategies being evaluated (Table 1, last column), and therefore any significant differences from these values imply potential
biases introduced by the use of inadequate connectome construction techniques.
Conclusion: The strategies by which individual streamlines are terminated, and subsequently contribute to a particular edge in the connectome, have significant influence
on the results of structural network analysis. Wit hout ap prop riat e d esig n o f these processing steps, the frequency of successful assignment of each streamline to exactly
two nodes is unacceptably low even when using high quality MRI data; the results presented here demonstrate that this has non-negligible concomitant effects on the
outcomes of structural connectome analysis. We th ere fo re str on gl y a dvo ca te th e application of advanced techniques for any studies investigating properties of the structural
connectome using diffusion MRI.
References: [1] Yeh et a l. , P r oc . ISMRM 2016; p.118. [2] Hagmann et al., PLoS Biol 2008; 6(7): e159. [3] Zalesky et al., NeuroImage 2010; 50(3): 970-83. [4] Van Essen et al.,
NeuroImage 2013; 80: 62-79. [5] Jeurissen et al., NeuroImage 2014; 103: 411-26. [6] Smith et al., NeuroImage 2012; 62(3): 1924-38. [7] Smith et al., NeuroImage 2015; 119 : 338-
51. [8] Tou rn ier et a l., Proc. ISMRM 2010; p.1670. [9] Desikan et al., NeuroImage 2006; 31: 968-80. [10] Tzourio-Mazoyer et al., NeuroImage 2002; 15(1): 273-289. [11] Smith et
al., NeuroImage 2015; 104: 253-65. [12] Rubinov & Sporns, NeuroImage 2010; 52: 1059-69.
Fig. 1. Frequency of 2-nod e connecti ons
obtained from eight differen t strat egies for
structural connectome constructions.
Table 1. Weighted connectomic metrics (mean±std) derived from 10 HCP data
Parcellation
FS parcellations
Metrics
Withou t ACT
With ACT
End vo xels
Local search
Local search
KW
( ×104 )
5.45 ± 0.42
10.44 ± 0.37Ω
13.15 ± 0.41Ω*
BW
( ×10-2 )
4.78 ± 0.39
4.44 ± 0.42
3.91 ± 0.17Ω*
CW
( ×10-2 )
8.58 ± 0.98
7.88 ± 0.82
9.05 ± 0.70Ω*
LW
( ×10-4 )
11.70 ± 1.41
4.17 ± 0.28Ω
2.40 ± 0.09Ω*
𝐸"
#
( ×10-2 )
3.81 ± 0.44
4.60 ± 0.47Ω
6.13 ± 0.44Ω*
𝐸$
#
( ×10-2 )
0.58 ± 0.07
0.71 ± 0.06Ω
1.21 ± 0.11Ω*
Parcellation
AAL at las
Metrics
Withou t ACT
With ACT
End vo xels
Local search
Local search
KW
( ×104 )
8.99 ± 0.39
11.50 ± 0.44Ω
13.15 ± 0.44Ω*
BW
( ×10-2 )
4.30 ± 0.31
4.13 ± 0.24
3.67 ± 0.22Ω*
CW
( ×10-2 )
7.03 ± 0.83
6.84 ± 0.83
8.92 ± 0.90 *
LW
( ×10-4 )
5.64 ± 0.43
4.03 ± 0.20Ω
2.82 ± 0.05Ω*
𝐸"
#
( ×10-2 )
4.65 ± 0.58
5.01 ± 0.62
6.59 ± 0.67Ω*
𝐸$
#
( ×10-2 )
0.82 ± 0.10
0.97 ± 0.12Ω
1.46 ± 0.17Ω*
Significant differences (p-value < 0.05) are denoted by: (i) Ω for end voxels versus local searc h; (ii)
* for non-ACT versu s ACT using the sam e assignment m echanism. (KW: strength; BW: betweenness
centrality; CW: clustering coefficient; LW: shortest path length; 𝐸"
#: global efficiency; 𝐸$
#: local
efficiency. W indi cates weighted version.)
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.