Synopsis
A new fully automated
algorithm, based on structural, morphological and relaxometric information, is
proposed to segment the entire brain deep venous system from MR images. The
method is tested on brain datasets at different magnetic fields and its
inter-scan reproducibility is also assessed. The proposed segmentation
algorithm shows good accuracy and reproducibility, outperforming previous
methods and becoming a promising candidate for the characterization of venous
tree topology. Purpose
The assessment of the
intracranial venous system anatomy provides a fundamental tool to study brain
diseases such as neurodegenerative disorders or traumatic brain injury. Venous
tree can be proficiently visualized by Susceptibility Weighted Imaging (SWI)1,
which allows for detection of vascular abnormalities in different cerebral
pathological conditions.
Manual segmentation of brain vessels in a MR dataset
is, however, a complex, time-consuming task and therefore automated or
semi-automated approaches are actively sought for2-4, as they also improve
results reproducibility.
Here we present and assess an algorithm that is, to
the best of our knowledge, the first fully automated segmentation tool of the
entire brain deep venous system that implements at once several independent
criteria (structural, morphological and relaxometric information) of vein
characterization to enhance classification accuracy.
Methods
Vein voxels are
characterized by:
· low
SWI intensity (structural content);
· high
Vesselness values (morphology);
· high
R2* values (relaxometry).
The tri-parametric segmentation
(tPS) algorithm iteratively refines the vein mask by adding newly detected
vessel voxels.
The initial condition
is provided by a highly selective global thresholding that distinguishes a large
number of reliable vein voxels from the surrounding parenchyma. Each of the following
criteria has to be satisfied:
· SWI<mean(SWI)–2.5*std_dev(SWI)
· Vesselness>mean(Vesselness)+std_dev(Vesselness)
· R2*>mean(R2*)
Then, a spherical moving
window filter (radius of 5mm) is iteratively applied: each voxel that satisfies
the above conditions, locally computed on the neighborhood excluding previously
marked voxels, is added to the vein mask.
The exit condition
for the iteration is verified when no more than 1‰ of the voxels are added in
the last step. The result of the last iteration is further refined by applying an
additional condition on the size of each classified cluster, as described
previously2.
The algorithm has
been tested on brain datasets from 6 Healthy Controls (HCs) obtained on different
MR scanners at B0 of 1.5, 3 and 7T. The acquired sequences were 3D
double-echo spoiled gradient echo (GRE) with acquisition parameters largely
dependent from B0 and flip angle close to the parenchyma Ernst angle.
To assess inter-scan
reproducibility, one HC was acquired twice at 3T, with
head repositioning between the scans (GREpre and GREpost).
SWI images were derived from the second echo of
each dataset5; Vesselness maps were computed from these images using
an optimized version of the Vessel Enhancing Diffusion filter6 and
the R2*-maps were obtained as previously reported6.
Results
The segmentation
results obtained at 1.5, 3 and 7T are shown in Fig.1, 2 and 3, respectively,
where the obtained voxel classifications are placed side by side with the
corresponding SWI and their fusion.
MIPped segmentation
maps (thickness: 20mm) were presented to two experienced neuroradiologists that,
blindly and in consensus, graded the accuracy of the vascular tree depiction on
a 0-5 scale (0 corresponding to the lowest reliability of the voxel
classification; 5 reflecting an optimal compromise between sensitivity and
specificity) and compared the performance of tPS to previous mono/bi-parametric
segmentation (m/bPS) approaches6, which neglect some classification
criteria. At all B0, m/bPS maps were never preferred to the corresponding tPS
maps in 36 comparisons. The mean accuracy scores were 4.6±0.2 for tPS vs 3.4±0.2
for m/bPS.
To evaluate inter-scan reproducibility, the
segmentations (Spre and Spost) were computed,
respectively from GREpre
and GREpost.
The second echo of GREpost was co-registered to the second echo of GREpre
and the obtained affine transformation was used to map Spost to Spre
coordinate system (Fig.4). A Modified Hausdorff Distance7 (MHD) was
used to grade the segmentation matching, resulting in a MHD of 0.33mm.
Discussion
In this work, a fully
automated method based on structural, morphological and relaxometric
information is proposed to segment the entire brain deep venous system, its
performance is evaluated at different B0s, and inter-scan reproducibility is
assessed at 3T.
This approach,
combining SWI, Vesselness and R2* values, led to a reduction of false positives
coupled to an improved detection of true positives, and achieved great accuracy
in vessel display, outperforming the m/bPS, which, lacking the structural/relaxometric
reference, erroneously identified many voxels as veins.
The segmentation
algorithm showed comparable accuracy at all the B0 explored, provided that
visible vessel density increases with B0. Moreover, it exhibited excellent
inter-scan reproducibility since the measured MHD of 0.33mm was well below the
resolution of the used dataset (0.5x0.5x1mm3).
Conclusion
In our opinion, this
algorithm represents a promising approach to characterize the topology of the
venous tree. However, further studies are required to extend the present method
to obtain a more comprehensive tool capable of quantifying parameters that can
be of clinical relevance for detection of venous change in neurovascular
diseases.
Acknowledgements
No acknowledgement found.References
[1] Liu J, Xia S, Hanks R et al. Susceptibility
Weighted Imaging and Mapping of Micro-hemorrhages and Major Deep Veins after
Traumatic Brain Injury. J Neurotrauma.
2015. [Epub ahead of print].
[2] Haacke EM, Reichenbach JR. Susceptibility
weighted imaging in MRI: basic concepts and clinical applications. John Wiley
& Sons. 2011.
[3] Bériault S, Archambault-Wallenburg M, Sadikot
AF et al. Automatic Markov Random Field Segmentation of Susceptibility-Weighted
MR Venography. Lecture Notes in Computer Science. 2014; 8361:39-47.
[4] Kuijf HJ, Bouvy WH, Zwanenburg JJM et al. Automated
detection of periventricular veins on 7 T brain MRI. SPIE Medical Imaging. 2015;
9413.
[5] Haacke EM, Xu Y, Cheng YCN et al. Susceptibility
weighted imaging (SWI). Mag Reson Med. 2004; 52(3):612-618.
[6] Monti S, Palma G, Borrelli P
et al. A Multiparametric and
Multiscale Approach to Automated Segmentation of Brain Veins. Proc of 37th IEEE
EMBc. 2015; 3041-3044.
[7] Dubuisson MP, Anil KJ. A modified Hausdorff
distance for object matching. Proceedings of the 12th IAPR. 1994; 1:566-568.