Hans Martin Kjer1,2, Mariam Andersson1,2, Yi He2, Marie Louise Elkjaer3, Alexandra Pacureanu4,5, Zsolt Illes3, Bente Pakkenberg6, Anders Bjorholm Dahl1, Vedrana Andersen Dahl1, and Tim B. Dyrby1,2
1DTU Compute, Technical University of Denmark, Kgs. Lyngby, Denmark, 2Danish Research Centre for Magnetic Resonance, Hvidovre, Denmark, 3Department of Neurology, Odense University Hospital, Odense, Denmark, 4X-ray Nanoprobe Group, ID16A, The European Synchrotron, Grenoble, France, 5University College London, London, United Kingdom, 6Research Laboratory for Stereology and Neuroscience, Bispebjerg University Hospital, Copenhagen NV, Denmark
Synopsis
We
present an efficient image analysis pipeline that enables us to reveal white
matter organization in high-resolution 3D non-MRI structural datasets, in cases
where a strict image segmentation is not required nor possible. We apply the
method to a synchrotron X-ray holographic tomography scan from a healthy mouse
sample, and show the organization of axon bundles in a region covering parts of
the corpus callosum and the cingulum. The method has a potential to improve our
general understanding of white matter organization and our ability to generate
realistic phantoms for validation of microstructure modelling from low-resolution
diffusion MRI scans.
Introduction
Diffusion-weighted
MRI (DWI) allows us to probe and model the microstructure of white matter
tissues from in-vivo scans. Validation of those models is an on-going and
crucial challenge for the community. The key is to work on a ground truth
structural data set. Phantom-based validation is typically the go-to option,
but they are in themselves simplified models of the real anatomy! Eventually,
we end up studying high-resolution volumetric datasets of microstructure. In this context, a high
resolution corresponds to images with enough detail, so that individual axons
and other microstructures can be resolved.
Making
a thorough data analysis of these large volumes is required, but too time
consuming and difficult to do manually. Semi- or fully automated segmentation
approaches are being developed1,2. However,
they are challenging to make generally applicable and work best at really high
resolutions, where all structural boundaries are clearly defined. Another downside
is that the high resolution typically is traded for a smaller field-of-view
(FOV) and this obscures the ability to study the larger organization of tissues
and the long-range behavior of white matter bundles etc.
Clearly, we
need to deal with a class of white matter structural datasets, where axons are
resolved almost as streamlines. Tracking and segmentation of an individual axon
is in such a case almost impossible to achieve reliably. However, as we will
show, it is possible to extract information about the organization in a
relatively simple manner, by the use of structure tensor analysis and tractography
as illustrated in Figure 1.Methods
Data Acquisition: In this preliminary study, we demonstrate the
method using a single healthy mouse sample. After perfusion fixation, the brain
was sliced and a biopsy (approx. 2.5 x 0.7 x 0.7 mm) extracted from a region
covering the splenium in Corpus Callosum (CC) and cingulum. The biopsy was
stained with osmium (OsO4, 0.5%), and embedded in EPON. Imaging of
the sample took place at the European Synchrotron and Radiation Facility (ESRF)
at beamline ID16A using X-ray holographic nano-tomography. The obtained volume
used in this study, see Figure 2, covers an extended FOV of 0.24 x 0.24 x 0.24
mm in a voxel resolution of 75 nm.
Structure Tensor: The primary workhorse for the data analysis is
the 3D structure tensor estimation3, here using a local Matlab
implementation. In short, the image gradients in all three axis directions are
measured in a small neighborhood around each voxel and collected in a 3x3
matrix. Using an Eigen-decomposition, we extract information about the local
orientation. In the case of a fiber-like material such as white matter, we can
estimate a clear dominant direction aligned with the main fiber orientation,
see Figure 3. The concept is very similar to DTI4, but based on
structural data content and not a diffusion MR signal.
Tractography: While the structure tensor provides an estimate
of the orientation information in all voxels, it does not reveal how structures
are connected. That we have to probe using deterministic tractography, here
using the MRTrix implementation5. The inspiration comes from
DWI-based connectivity data analysis, but the application to structural data is
still novel. Based on a seed point, a particle trajectory through the volume is
simulated using the local main orientation for direction until some stopping
mask or criteria is met, see Figure 4.
Clustering: The
output of the tractography is a large number of unorganized streamlines without
a direct biological interpretation. It is then beneficial to apply a streamline
clustering method, which collects multiple streamlines into meaningful
axonal-bundles. We use the
QuickBundles method6 for its simplicity and scalability, and
the result can be seen on Figure 1.Results
Axon-bundles
are clearly revealed both within the CC and cingulum. An immediate observation
is that bundles trajectories are non-parallel and perform subtle bends and
dispersions to move around cellular structures. An analysis on the cluster
centroid trajectories, shows that bundles turn with angles up to 28.6 and 7.8
degrees in CC and cingulum respectively. Such information is valuable in the
design of realistic white matter phantoms.Discussion
While we
demonstrate our method on a synchrotron X-ray tomography dataset, it is in
principle no hindrance to apply it to other structural and volumetric modalities.
The synchrotron is a good option, as it provides relatively large FOVs with
enough resolution to generate the streamline characteristic in white matter
that our method targets. It is further a non-destructive technique, allowing us
to cover an even larger volume with overlapping FOVs.
Extensions
to our work includes exploring more white matter regions of the brain and
comparing healthy vs. diseased samples. More advanced approaches of both
tractography and clustering can be investigated, which might be beneficial in crossing
fibers regions.Conclusion
We have demonstrated
an efficient image analysis pipeline based on structure tensor and tractography
to investigate the 3D white matter organization. The method is ideally applied
to high-resolution 3D structural datasets, in cases where a strict image segmentation
is not required nor possible. It can serve to improve our general understanding
of white matter organization and our ability to generate realistic phantoms for
validation of microstructure modelling from low-resolution diffusion MRI scans.Acknowledgements
M. Andersson and H.M. Kjer were supported by the Capital Region Research Foundation (grant number:
A5657) (PI: T. Dyrby).References
[1] Abdollahzadeh A,
Belevich I, Jokitalo E, Tohka J, Sierra A. Automated 3D Axonal
Morphometry of White Matter. Scientific reports. 2019;9(1):6084.
[2] Zaimi A,
Wabartha M, Herman V, Antonsanti PL, Perone CS, Cohen-Adad J. AxonDeepSeg:
Automatic Axon and Myelin Segmentation From Microscopy Data Using Convolutional
Neural Networks. Scientific reports. 2018;8(1):3816.
[3] Khan AR, Cornea A, Leigland LA, Kohama SG, Jespersen SN, Kroenke CD. 3D
Structure Tensor Analysis of Light Microscopy Data for Validating Diffusion MRI.
NeuroImage. 2015;111:192-302.
[4] Mori S, Zhang J. Principles of Diffusion Tensor Imaging and Its Applications to Basic Neuroscience Research. Neuron.
2006;51(5):527-539
[5] Tournier JD, Smith RE, Raffelt D, Tabbara R, Dhollander T, Pietsch M, Christiaens D, Jeurissen B, Yeh CH, Connelly A. MRtrix3: A Fast, Flexible and Open Software Framework for Medical Image Processing and Visualisation. NeuroImage. 2019;202:116-37.
[6] Garyfallidis E, Brett M, Correia MM, Williams GB, Nimmo-Smith I. QuickBundles,
a Method for Tractography Simplification. Front Neurosci. 2012;6:175.