Conference PaperPDF Available

The influence of node assignment strategies and track termination criteria on diffusion MRI-based structural connectomics


Abstract and Figures

This study highlights the issue of using the common strategy for assigning individual streamlines to an atlas-based brain parcellation. This process is non-trivial and can introduce ambiguity into connectome quantification. In many fibre-tracking algorithms, track termination criteria can cause premature termination of streamlines within WM or CSF, which can result in up to ~50–80% of streamlines failing in identifying pairwise connections between nodes from streamline endpoints. Our results demonstrate that such issue can be largely ameliorated through the combination of biologically meaningful track terminations and an appropriate node assignment mechanism. This could therefore be advantageous to structural connectome construction.
Content may be subject to copyright.
7/15/16, 5:21 PM
Page 1 of 3
The inuence of node assignment strategies and track termination
criteria on diusion MRI-based structural connectomics
Chun-Hung Yeh , Robert Elton Smith , Thijs Dhollander , Fernando Calamante , and Alan Connelly
The Florey Institute of Neuroscience and Mental Health, Melbourne, Australia
This study highlights the issue of using the common strategyThis study highlights the issue of using the common strategy
for assigning individual streamlines to an atlas-based brainfor assigning individual streamlines to an atlas-based brain
parcellation. This process is non-trivial and can introduceparcellation. This process is non-trivial and can introduce
ambiguity into connectome quantication. In many bre-ambiguity into connectome quantication. In many bre-
tracking algorithms, track termination criteria can causetracking algorithms, track termination criteria can cause
premature termination of streamlines within WM or CSF,premature termination of streamlines within WM or CSF,
which can result in up to ~50–80% of streamlines failing inwhich can result in up to ~50–80% of streamlines failing in
identifying pairwise connections between nodes fromidentifying pairwise connections between nodes from
streamline endpoints. Our results demonstrate that suchstreamline endpoints. Our results demonstrate that such
issue can be largely ameliorated through the combination ofissue can be largely ameliorated through the combination of
biologically meaningful track terminations and anbiologically meaningful track terminations and an
appropriate node assignment mechanism. This couldappropriate node assignment mechanism. This could
therefore be advantageous to structural connectometherefore be advantageous to structural connectome
Diusion MRI streamlines tractography has become the main technique
for inferring structural brain connectivity. This is typically achieved by
constructing a connectome by identifying streamlines that link regions-
of-interest dened by an atlas-based parcellation scheme to provide a
summary of white matter (WM) connections between pairs of grey
matter (GM) regions (i.e. network nodes) . The mechanism used to
assign streamlines to nodes is potentially crucial to providing meaningful
characterisation of the connectome. Ideally, each streamline should
connect exactly 2 nodes. However, track termination criteria in many
bre-tracking algorithms result in premature termination of streamlines
within WM or CSF. This causes streamlines apparently being associated
with either zero or only one GM regions (up to ~50–80% ), or even
non-GM regions within nodes. Such apparent ‘connectivity’ is not
biologically meaningful, and could result therefore in a misleading
connectome in non-trivial ways. The biological plausibility of streamline
terminations in principle can be improved by using methods such as
anatomically-constrained tractography (ACT), which accepts a
streamline only if it terminates within either cortical, sub-cortical GM, or
brainstem. This study investigates the inuence of track assignment
strategies and their interactions with some popular parcellation schemes
–with or without the application of ACT – for identifying pairwise
connectivity between nodes, revealing important implications for
connectome quantication.
MRI acquisition: T1s and DWIs (2.5-mm isotropic resolution, 60
directions, b=3000 s/mm ) of 22 healthy volunteers were acquired using
a Siemens 3T Tim Trio MRI scanner.
1 1 1 1 1
Fig. 1.Fig. 1. Compared to the FreeSurfer
parcellation (left), the AAL parcels (right)
extend far into the WM; the spatial extent of
these GM parcels may inuence the frequency
of successful streamlines assignment.
Fig. 2.Fig. 2. Histograms of node count per
streamline. Tractograms, generated either with
or without ACT, were assigned to network
nodes dened by either FreeSurfer or AAL
parcellation through the (a) end voxels and (b)
local search.
Fig. 3.Fig. 3. (a) The segmented anatomical
reference map overlaid with track endpoints
(yellow) and the FreeSurfer parcellation. (b-c)
With ACT, track endpoints (see purple arrows)
are located at the GM-WM interface, which
could however be inconsistent with the GM
parcellation due to factors such as
discretisation of GM labels.
7/15/16, 5:21 PM
Page 2 of 3
Tractogram reconstruction: Fibre orientation distributions (FODs) were
computed using constrained spherical deconvolution . For each scan,
tractograms of 10 million streamlines were generated through seeding
from WM mask, tracking either with or without ACT using the iFOD2
algorithm .
GM parcellation: Two parcellation schemes were used (Fig. 1): (i) the T1-
based FreeSurfer parcellation ; (ii) the AAL atlas transformed onto each
individual’s T1 space.
Connectome construction: Streamlines were assigned to nodes by: (i)
end voxels = voxels at streamline endpoints; (ii) local search = a search
from each streamline endpoint to locate the nearest node within a 2-mm
Fig. 2 shows the histogram of node count per streamline. The results are
summarised as follows:
Eect of assignment mechanism: Compared to using end voxels (Fig. 2(a)),
using the local search (Fig. 2(b)) increased the proportion of identifying 2
nodes, particularly when FreeSurfer parcellations were used.
Eect of ACT: Compared to non-ACT, the use of ACT improved the overall
ability of identifying 2 nodes. Combining ACT with the local search (Fig.
2(b)), the number of tracks identifying 2 nodes reached ~90% for both
parcellation schemes.
Eect of parcellation scheme: Due to the presence of some WM voxels in
the AAL parcels, the probability of streamlines reaching nodes was
higher than using the FreeSurfer parcellation. When using end voxels
(Fig. 2(a)), the parcellation scheme had a dominant eect; e.g. when
using the FreeSurfer parcellation, <25% of streamlines were assigned to
2 nodes.
Our data demonstrate that the connectome construction can be
strongly inuenced by many factors during the process of assigning
streamlines to nodes; this process is non-trivial, and a number of
confounds can introduce ambiguity into connectome quantication. For
non-ACT, using a parcellation scheme with parcels not constrained to the
GM ribbon, such as the AAL atlas, does not provide a meaningful
compensation for the defect in both tracking terminations and
assignment mechanism: the increase in pairwise connections using AAL
probably largely results from those tracks terminating within WM but
which nevertheless reach the AAL parcels at both endpoint voxels. By
contrast, using anatomical information to constrain track terminations to
the interface of GM and WM, such as ACT , is more benecial for
identication of connectivity between GM parcels; however, the ecacy
of connectome construction also relies on the consistency between
tissue segmentation and GM parcellation (Fig. 3). This explains why a
7 8
7/15/16, 5:21 PM
Page 3 of 3
short local search (~2 voxels) is helpful in overcoming small
inconsistencies, although the biological plausibility of such a ‘searching’
process requires further investigation.
The commonly-used mechanism by which individual streamlines
contribute to the connectome may be ill-dened in many instances. In
practice, it is dicult to develop such a mechanism that can eectively
deal with every potential scenario, and connectomes constructed based
on such methods may actually be heavily inuenced by these
mechanisms rather than realistic biology. In this context, both robust
terminations of streamlines respecting anatomical information at the
bre-tracking stage and an appropriate mechanism at the node
assignment stage should be benecial to identifying pairwise GM
connections from meaningful endpoints for connectome construction.
No acknowledgement found.
1. Bullmore E & Sporns O. Complex brain networks: graph theoretical
analysis of structural and functional systems. Nat Rev Neurosci.
2. Hagmann P, et al. Mapping the structural core of human cerebral
cortex. PLoS Biol. 2008;6(7):e159.
3. Zalesky A, et al. Whole-brain anatomical networks: does the choice of
nodes matter? NeuroImage. 2010;50(3):970-83.
4. Smith RE, et al. Anatomically-constrained tractography: improved
diusion MRI streamlines tractography through eective use of
anatomical information. NeuroImage. 2012;62(3):1924-38.
5. Tournier JD, et al. Robust determination of the bre orientation
distribution in diusion MRI: non-negativity constrained super-resolved
spherical deconvolution. NeuroImage. 2007;35(4):1459-72.
6. Tournier JD, et al. Improved probabilistic streamlines tractography by
2nd order integration over bre orientation distributions. Proc. ISMRM.
7. Desikan RS, et al. An automated labeling system for subdividing the
human cerebral cortex on MRI scans into gyral based regions of interest.
NeuroImage. 2006;31:968-80.
8. Tzourio-Mazoyer N, et al. Automated anatomical labeling of activations
in SPM using a macroscopic anatomical parcellation of the MNI MRI
single-subject brain. NeuroImage. 2002;15(1):273-289.
Proc. Intl. Soc. Mag. Reson. Med. 24 (2016) 0118
... In the ideal case, each fiber contributes to exactly two ROIs. However, the node assignment is far from perfect (Daducci et al., 2016;Yeh et al., 2016). We tried to improve the mapping by allowing the extension of each fiber start-/endpoint in the direction of the first/last fiber segment for an additional stretch of 2 mm. ...
... We tried to improve the mapping by allowing the extension of each fiber start-/endpoint in the direction of the first/last fiber segment for an additional stretch of 2 mm. This approach is more restrictive compared to the radial search proposed in Smith et al. (2015a) and further evaluated in Yeh et al. (2016). Nevertheless, we were still unable to perfectly map the tractogram onto the cortical parcellation. ...
Full-text available
Diffusion magnetic resonance imaging (dMRI) is a compelling tool for investigating the structure and geometry of brain tissue based on indirect measurement of the diffusion anisotropy of water. Recent developments in global top-down tractogram optimizations enable the estimation of streamline weights, which characterize the connection between gray matter areas. In this work, the intra-axonal cross-sectional area coverage of the gray-to-white matter interface was examined by intersecting tractography streamlines with cortical regions of interest. The area coverage is the ratio of streamline weights divided by the surface area at the gray-to-white matter interface and assesses the estimated percentage which is covered by intra-axonal space. A high correlation (r = 0.935) between streamline weights and the cortical surface area was found across all regions of interest in all subjects. The variance across different cortical regions exhibits similarities to myelin maps. Additionally, we examined the effect of different diffusion gradient subsets at a lower, clinically feasible spatial resolution. Subsampling of the initial high-resolution diffusion dataset did not alter the tendency of the area coverage at the gray-to-white matter interface across cortical areas and subjects. However, single-shell acquisition schemes with lower b-values lead to a steady increase in area coverage in comparison to the full acquisition scheme at high resolution.
... An important step in providing such is ensuring that streamline terminations (see Section above) are consistent with the anatomical information used to drive the bundle classification process [19,52,[147][148][149]. 3. The number of streamlines generated for reconstruction of a particular pathway/whole-brain structural connectivity must be sufficiently large to capture the extent of both the complexity of the underlying structure, and the stochastic nature of the tractography experiment. ...
“Tractography” refers to the process of digitally reconstructing the spatial trajectories of elongated thin structures based on a continuous field of local orientation data; most commonly in reference to white matter axons within the central nervous system based on a diffusion Magnetic Resonance Imaging (MRI) model. This technology provides a unique capacity to investigate the long-distance connectional structure of such biology both in vivo and noninvasively. This chapter describes the basic principles behind the near-ubiquitous “streamlines” approach to tractography (which has remained relatively unchanged for 20 years) as well as some modern advancements to such, with a particular emphasis on the quantitative capabilities of this technology.
... This argues against the possibility that only a small, possibly random, sub-set of all recon- structed streamlines was selected for analyses. Yet, it also has to be borne in mind that some AAL seeds extend farther into white matter than others, so that more streamlines ending prematurely in white matter could have been selected nevertheless ( Yeh et al., 2016). Thus, our results cannot inform about the anatomical plausibility and reliability of reconstructed streamlines, but only about their test-retest reproduc- ibility. ...
As quantitative measures derived from fiber tractography are increasingly being used to characterize the structural connectivity of the brain, it is important to establish their reproducibility. However, no such information is as yet available for global tractography. Here we provide the first comprehensive analysis of the reproducibility of streamline counts derived from global tractography as quantitative estimates of structural connectivity. In a sample of healthy young adults scanned twice within one week, within-session and between-session test-retest reproducibility was estimated for streamline counts of connections based on regions of the AAL atlas using the intraclass correlation coefficient (ICC) for absolute agreement. We further evaluated the influence of the type of head-coil (12 versus 32 channels) and the number of reconstruction repetitions (reconstructing streamlines once or aggregated over ten repetitions). Factorial analyses demonstrated that reproducibility was significantly greater for within-than between-session reproducibility and significantly increased by aggregating streamline counts over ten reconstruction repetitions. Using a high-resolution head-coil incurred only small beneficial effects. Overall, ICC values were positively correlated with the streamline count of a connection. Additional analyses assessed the influence of different selection variants (defining fuzzy versus no fuzzy borders of the seed mask; selecting streamlines that end in versus pass through a seed) showing that an endpoint-based variant using fuzzy selection provides the best compromise between reproducibility and anatomical specificity. In sum, aggregating quantitative indices over repeated estimations and higher numbers of streamlines are important determinants of test-retest reproducibility. If these factors are taken into account, streamline counts derived from global tractography provide an adequately reproducible quantitative measure that can be used to gauge the structural connectivity of the brain in health and disease.
... A 3 mm dilation was used to ensure a robust overlap between streamlines end-points (e.g. GM/WM interface) and anatomical labels [14]. Finally, streamlines and brain labels were loaded in FiberNavigator 1 [3]. ...
Conference Paper
Full-text available
Structural networks contain high dimensional data that raise huge computational and visualization problems, especially when attempting to characterise them using graph theory. As a result, it can be non-intuitive to grasp the contribution of each edge within a graph, both at a local and global scale. Here, we introduce a new platform that enables tractography-based networks to be explored in a highly interactive real-time fashion. The framework allows one to interactively tune graph-related parameters on the fly, as opposed to conventional visualization softwares that rely on pre-computed connectivity matrices. From a neurosurgical perspective, the method also provides enhanced understanding regarding the potential removal of a specific node or transection of an edge from the network, allowing surgeons and clinicians to discern the value of each node.
Full-text available
Recent developments in the quantitative analysis of complex networks, based largely on graph theory, have been rapidly translated to studies of brain network organization. The brain's structural and functional systems have features of complex networks--such as small-world topology, highly connected hubs and modularity--both at the whole-brain scale of human neuroimaging and at a cellular scale in non-human animals. In this article, we review studies investigating complex brain networks in diverse experimental modalities (including structural and functional MRI, diffusion tensor imaging, magnetoencephalography and electroencephalography in humans) and provide an accessible introduction to the basic principles of graph theory. We also highlight some of the technical challenges and key questions to be addressed by future developments in this rapidly moving field.
Full-text available
Author Summary In the human brain, neural activation patterns are shaped by the underlying structural connections that form a dense network of fiber pathways linking all regions of the cerebral cortex. Using diffusion imaging techniques, which allow the noninvasive mapping of fiber pathways, we constructed connection maps covering the entire cortical surface. Computational analyses of the resulting complex brain network reveal regions of cortex that are highly connected and highly central, forming a structural core of the human brain. Key components of the core are portions of posterior medial cortex that are known to be highly activated at rest, when the brain is not engaged in a cognitively demanding task. Because we were interested in how brain structure relates to brain function, we also recorded brain activation patterns from the same participant group. We found that structural connection patterns and functional interactions between regions of cortex were significantly correlated. Based on our findings, we suggest that the structural core of the brain may have a central role in integrating information across functionally segregated brain regions.
Whole-brain anatomical connectivity in living humans can be modeled as a network with diffusion-MRI and tractography. Network nodes are associated with distinct grey-matter regions, while white-matter fiber bundles serve as interconnecting network links. However, the lack of a gold standard for regional parcellation in brain MRI makes the definition of nodes arbitrary, meaning that network nodes are defined using templates employing either random or anatomical parcellation criteria. Consequently, the number of nodes included in networks studied by different authors has varied considerably, from less than 100 up to more than 10(4). Here, we systematically and quantitatively assess the behavior, structure and topological attributes of whole-brain anatomical networks over a wide range of nodal scales, a variety of grey-matter parcellations as well as different diffusion-MRI acquisition protocols. We show that simple binary decisions about network organization, such as whether small-worldness or scale-freeness is evident, are unaffected by spatial scale, and that the estimates of various organizational parameters (e.g. small-worldness, clustering, path length, and efficiency) are consistent across different parcellation scales at the same resolution (i.e. the same number of nodes). However, these parameters vary considerably as a function of spatial scale; for example small-worldness exhibited a difference of 95% between the widely-used automated anatomical labeling (AAL) template (approximately 100 nodes) and a 4000-node random parcellation (sigma(AAL)=1.9 vs. sigma(4000)=53.6+/-2.2). These findings indicate that any comparison of network parameters across studies must be made with reference to the spatial scale of the nodal parcellation.
An anatomical parcellation of the spatially normalized single-subject high-resolution T1 volume provided by the Montreal Neurological Institute (MNI) (D. L. Collins et al., 1998, Trans. Med. Imag. 17, 463-468) was performed. The MNI single-subject main sulci were first delineated and further used as landmarks for the 3D definition of 45 anatomical volumes of interest (AVOI) in each hemisphere. This procedure was performed using a dedicated software which allowed a 3D following of the sulci course on the edited brain. Regions of interest were then drawn manually with the same software every 2 mm on the axial slices of the high-resolution MNI single subject. The 90 AVOI were reconstructed and assigned a label. Using this parcellation method, three procedures to perform the automated anatomical labeling of functional studies are proposed: (1) labeling of an extremum defined by a set of coordinates, (2) percentage of voxels belonging to each of the AVOI intersected by a sphere centered by a set of coordinates, and (3) percentage of voxels belonging to each of the AVOI intersected by an activated cluster. An interface with the Statistical Parametric Mapping package (SPM, J. Ashburner and K. J. Friston, 1999, Hum. Brain Mapp. 7, 254-266) is provided as a freeware to researchers of the neuroimaging community. We believe that this tool is an improvement for the macroscopical labeling of activated area compared to labeling assessed using the Talairach atlas brain in which deformations are well known. However, this tool does not alleviate the need for more sophisticated labeling strategies based on anatomical or cytoarchitectonic probabilistic maps.
In this study, we have assessed the validity and reliability of an automated labeling system that we have developed for subdividing the human cerebral cortex on magnetic resonance images into gyral based regions of interest (ROIs). Using a dataset of 40 MRI scans we manually identified 34 cortical ROIs in each of the individual hemispheres. This information was then encoded in the form of an atlas that was utilized to automatically label ROIs. To examine the validity, as well as the intra- and inter-rater reliability of the automated system, we used both intraclass correlation coefficients (ICC), and a new method known as mean distance maps, to assess the degree of mismatch between the manual and the automated sets of ROIs. When compared with the manual ROIs, the automated ROIs were highly accurate, with an average ICC of 0.835 across all of the ROIs, and a mean distance error of less than 1 mm. Intra- and inter-rater comparisons yielded little to no difference between the sets of ROIs. These findings suggest that the automated method we have developed for subdividing the human cerebral cortex into standard gyral-based neuroanatomical regions is both anatomically valid and reliable. This method may be useful for both morphometric and functional studies of the cerebral cortex as well as for clinical investigations aimed at tracking the evolution of disease-induced changes over time, including clinical trials in which MRI-based measures are used to examine response to treatment.
Diffusion-weighted (DW) MR images contain information about the orientation of brain white matter fibres that potentially can be used to study human brain connectivity in vivo using tractography techniques. Currently, the diffusion tensor model is widely used to extract fibre directions from DW-MRI data, but fails in regions containing multiple fibre orientations. The spherical deconvolution technique has recently been proposed to address this limitation. It provides an estimate of the fibre orientation distribution (FOD) by assuming the DW signal measured from any fibre bundle is adequately described by a single response function. However, the deconvolution is ill-conditioned and susceptible to noise contamination. This tends to introduce artefactual negative regions in the FOD, which are clearly physically impossible. In this study, the introduction of a constraint on such negative regions is proposed to improve the conditioning of the spherical deconvolution. This approach is shown to provide FOD estimates that are robust to noise whilst preserving angular resolution. The approach also permits the use of super-resolution, whereby more FOD parameters are estimated than were actually measured, improving the angular resolution of the results. The method provides much better defined fibre orientation estimates, and allows orientations to be resolved that are separated by smaller angles than previously possible. This should allow tractography algorithms to be designed that are able to track reliably through crossing fibre regions.