Andrey Zhylka1, Alexander Leemans2, Josien Pluim1, and Alberto De Luca2
1Biomedical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands, 2Image Sciences Institute, University Medical Center Utrecht, Utrecht, Netherlands
Synopsis
Neurosurgery planning is an important
application of fiber tractography which requires the results to be consistent
and accurate. Deterministic tractography methods are generally characterized by
high specificity and limited sensitivity, whereas the opposite typically holds
for probabilistic methods. Here, we propose a multi-level fiber tractography
strategy that takes fiber branching into account and incorporates an anatomical
prior to provide a balance between true and false positive reconstructions. We
evaluated our approach on the MASSIVE dataset and compared its performance to the
existing state of art.
Introduction
Fiber tractography (FT) with diffusion MRI is a powerful tool for neurosurgery
planning as it can reconstruct brain pathways in vivo and noninvasively1.
However, it is challenging to obtain a reliable and consistent reconstruction
of fiber trajectories2.
In particular, probabilistic FT approaches tend to produce more false positive
reconstructions than deterministic ones, while deterministic approaches typically
produce more false negative pathways. Additionally, neither of the approaches
addresses the notion of fiber bundles taking very sharp turns. Moreover, given
the angular constraints and resolution that is too coarse for such purpose,
branching pathways are quite likely to be pruned leading to an increased
false negative rate3. We developed a novel way to reconstruct fiber
bundles as multi-level structures, where each consecutive level is a set of
pathways branching from the previous one. We investigate whether this strategy
in combination with the prior knowledge of the location of seed and target
regions of interest (ROI) can improve the sensitivity of the reconstructions, while
maintaining properties of deterministic FT.Methods
The proposed strategy was implemented for tractography
methods based on constrained spherical deconvolution4 (CSD). It iteratively complements the results
with the new levels of branches (Figure 1). At
each iteration CSD-based deterministic tractography is performed as the first
step (with an angular-deviation threshold of 45° and a fiber orientation distribution (FOD) threshold of 0.1). Secondly, points
along a pathway with unused crossing directions are used as new seed points,
and the unprocessed FOD peaks as initial directions. This procedure runs either for a pre-defined
number of iterations or until a convergence criterion is met. Finally, using
prior anatomical knowledge of the tract topology, a terminal ROI is defined to select the pathways that are to remain.
The MASSIVE5 dataset
was used with isotropic voxel resolution of 2.5mm.
The acquisition consisted of 430 volumes at b = 0s/mm2,
250 volumes at b = 500 s/mm2,
500 volumes at 1000s/mm2, 2000s/mm2 and 3000s/mm2 each, 600 volumes at 4000s/mm2.
The data was corrected for signal drift, motion and Eddy current using ExploreDTI6. Target areas were selected from the cortical
parcellation obtained with FreeSurfer7.
Subsequently, the fiber orientation distributions were estimated with
multi-shell CSD8.
To evaluate the performances of the proposed
method, we reconstructed the corticospinal tract (CST) by selecting the precentral
and paracentral gyri as target regions, while seeding below the internal
capsule. Each seed voxel was subsampled evenly on 3x3 grid at single slice level.
The results of the proposed procedure were compared to those obtained with CSD-based
deterministic FT and probabilistic FT given identical tracking parameters and
conditions. Probabilistic FT was performed with iFOD29. Pathway coherence was analyzed using TSNE10 as well as functional topology preservation for
both algorithms. Generalizability of the method was tested on the cingulum bundle
in comparison to deterministic CSD-based FT. For this case, the target region
was chosen based on the WMQL-like11
cingulum query, including only the regions containing endpoints of the bundle.
The seed ROI for the cingulum was placed on the edge of Brodmann areas 23 and
24.
Results
Figure 2 compares
the proposed approach to conventional CSD-based tractography for a left branch
of the CST. As compared to CSD, our method provides a more
extensive reconstruction of the CST, including branches that constitute more
than a half of the well-known “fanning” of the tract into the motor cortex. The iFOD2
algorithm can also reconstruct the CST branches and provides tracts with a good
coverage of the motor cortex (Figure
3).
However, our approach allows reconstruction of pathways with higher coherence, as
supported by the TSNE clustering shown in Figure
3.
This aspect is also represented in better preservation of functional topology
by the proposed approach, which is shown in Figure
4. Improved
tractography results were also obtained for the case of cingulum delineation. Figure
5 compares
the cingulum tracking results obtained with our method to those of deterministic
fiber tractography from the same seed point regions. The bundle obtained with the
proposed approach additionally captures pathways going to the superior lobes in
the anterior and posterior parts, whereas deterministic FT reconstructs a more
limited part of the bundle.Discussion
In this work, we have shown that the proposed
approach is capable of reconstructing complex fiber bundles, such as the corticospinal
tract and the cingulum, with improvement in false-negative rate while maintaining
properties of deterministic tractography. The produced pathways are more
coherent than those produced by probabilistic FT, which are characterized by
higher tortuosity and an erratic nature. Consequently, our approach preserves
functional topology better than iFOD2. The accuracy of our approach is dependent
on the terminal region, which should be defined carefully. An
adequate seeding area and a sufficient seed density are also important for an
accurate reconstruction. Sparse seeding might lead to a deficit of the expected
branching structures in complex cases.Conclusion
We developed a novel way to perform FT by reconstructing multi-level structures that integrates the orientations with a higher angular deviation and
that incorporates prior knowledge of fiber tract anatomy. In doing so, we
obtain a more complete representation of brain pathways.Acknowledgements
This project has received funding from the European Union's Horizon 2020
research and innovation program under the Marie Sklodowska-Curie grant
agreement No 765148.References
-
Jeurissen B., Descoteaux M., Mori S., Leemans A. Diffusion MRI
tractography of the brain. NMR Biomed 32(4) (2019) e3785.
-
Maier-Hein
K. et al. The challenge of mapping the human connectome based on diffusion
tractography, Nature Communications 8
(1) (2017) 1349.
-
Mortazavi F., Oblak A. L., Morrison W. Z.,
Schmachmann J. D., Stanley H. E., Wedeen V. J.,
Rosene D. L. Geometric navigation of axons in a cerebral pathway: comparing
dMRI with tract tracing and immunohistochemistry. Cerebral Cortex 28(4) (2018) 1219-1232.
-
Jeurissen B., Leemans A., Jones D. K., Tournier J.-D., Sijbers J.
Probabilistic fiber tracking using the residual bootstrap with constrained
spherical deconvolution, Hum Brain Mapp
32 (3) (2011) 461–479.
-
Froeling M., Tax C. M., Vos S. B., Luijten P.
R., Leemans A. MASSIVE brain dataset: Multiple acquisitions for standardization
of structural imaging validation and evaluation, Magn Reson in Med 77 (5) (2017) 1797–1809.
-
Leemans, A., Jeurissen, B., Sijbers, J., Jones,
D. K. ExploreDTI: a graphical toolbox for processing, analyzing, and
visualizing diffusion MR data. ISMRM; 3537 (2009).
- Fischl B. FreeSurfer, NeuroImage 62 (2) (2012) 774–781.
- Jeurissen B., Tournier J.-D., Dhollander T.,
Connelly A., Sijbers J. Multi-tissue constrained spherical deconvolution for
improved analysis of multi-shell diffusion MRI data, NeuroImage 103 (2014) 411–426.
-
Tournier J.-D., Calamante F., Connelly A.
Improved probabilistic streamlines tractography by 2nd order integration over
fiber orientation distributions, ISMSRM;
1670 (2010).
-
Oliveira
F.H.M., Machado A.R.P., Andrade A.O. On the Use of t-Distributed
Stochastic Neighbor Embedding for Data Visualization and Classification of
Individuals with Parkinson's Disease. Comput Math Methods Med 8019232
(2018).
-
Wassermann
D., Makris N., Rathi Y., Shenton M., Kikinis R., Kubicki M., Westin C.F. The
white matter query language: a novel approach for describing human white matter
anatomy. Brain Struct and Func 221
(9) (2016) 4705-4721.