Partitioning of the brain using graph-cut
Robert Dahnke, Rachel Aine Yotter, Christian Gaser
Structural Brain Mapping Group, Department of Psychiatry
Friedrich-Schiller-University of Jena, Germany
Skull-stripping and partitioning in major regions are important tasks for further analysis of MRI data We test our method for normal adult and child brains, but also for extremely deformed brains with
that still allows improvements (Han et al., 2007; Sadananthan et al. 2010; Zhao et al, 2010). We have large ventricles. Figure 2 shows slices of di erent subjects with the T1-data as background and the
developed a new method to re ne the SPM segmentation (Ashburner et al., 2005) by a improved personal atlas SLAB as colored overlay. The SEGPF map was used to generate surfaces with PBT
removal of non-brain tissues, and to separate the brain into major regions. The method uses atlas, (Dahnke et al. 2010). Figure 3 shows improvements of our method compared to VBM8.
distance and intensity information, and was validated for normal and abnormal brains of subjects
with di erent age.
Child Adult Old Patient
As input, the label map SEG and the anatomical bias corrected T1 image of the VBM8 toolbox
(http://dbm.neuro.uni-jena.de) of SPM8 (Ashburner et al., 2005) are required (Figure 1). Dartel
normalization (Ashburner et al., 2007) is used to project the atlas map ALAB onto subject space. It
allows identi cation of major WM regions, and creation of an individual atlas map SLAB.
After this rst alignment, a graph-cut algorithm is used to allocate GM/WM voxels around these
regions. To remove blood vessels and meninges the graph-cut of cerebral structures only allows small
increases of the image intensity for clustered of neighbors to the start voxels. Remaining voxels are
set to GM if they are close to the WM or added as blood vessel to the SLAB map. For further blood
vessel detection, morphological operations, distance, and thickness information are used.
To nd the ventricle, the bottleneck method (Mangin et al., 1996) is used. CSF voxels in the ventricular
ALAB regions get a high potential, whereas voxels that belong to the background, cerebellum or
brainstem represent low potential. The Laplace equation is solved within the CSF with Dirichlet
boundary conditions. The biggest left and right component of voxels with more than 70% of the
high potential were added as ventricle to the SLAB map.
For subcortical structures, GM and WM voxels of subcortical and ventricular ALAB regions get high
potential, whereas non-ventricular CSF is used as the negative potential. A threshold of 95% for AC-20 mm
the high potential is used to locate subcortical regions in the GM and update the SLAB map. Figure 2: Volumes of a healthy child (far left), a healthy adult (middle left), a healthy old (middle right), and a
patient Volumes of dinormal child (far left), row: AC+20 mm,(middle left), :aAC-20 mm). (middle right),
Fig. 2: (far right) at a erent axial levels (top a normal adult bottom row normal old
For surface reconstruction, ventricles and subcortical structures are lled and non-brain tissues and a patient (far right) at di erent axial levels (top row: AC+20 mm, bottom row : AC-20 mm).
are removed, resulting in a map SEGPF. Morphological operations are used in the neighborhood
The Segmentation Validation Engine (Shattuck et
of the corrected regions to allow continuous structures.
al. 2009) was used for further evaluation (Figure
Finally, the graph-cut algorithm divides each region into its left and right component, starting 4). Because the manual tracing method of blood
from the highest tissue class in SEGPF. vessels, meninges and CSF is not fully clear an
interpretation of the results is problematic. In (Rehm
Input images: T1 image, VBM8 segment et al., 2003) two experts created tracings with and
map SEG and atlas map ALAB. without blood vessels, leading to big visible results
in the volume rendering, but only small di erences
in Dice similarity (below 0.02).
T1 SEG ALAB
Calculation times are around 8 minutes on a 2.4 GHz iMac.
Creation of TSEG image as a combination
of the T1 and SEG image. Standard VBM8 segmentation (left) contains more
blood vessels (manually colored) than our method
DARTEL projection of the atlas map to
(right). Further colors on the right surface indicate the
TSEG di erent regions.
Results for Study #221:
Identi cation of main structures in the TSEG Dice coe cient 0.9706
SVE sesults for study #221: ± 0.0041
images with help of atlas information. Jaccard index 0.9430 ± 0.0076
Dice coe cient 0.9706 ± 0.0041 Figure 3: Standard VBM8 segmentation (left) contains
Fig. 3: Standard VBM8 segmentation (left) contains more blood vessels (manually colored) than our
Alignment of WM tissue with graph-cut. Sensitivity
Jaccard index 0.9542 ± 0.0105
0.9430 ± 0.0076 more blood vessels (manually colored) than our
method (right). Further colors on the right surface indicate the di erent regions.
Sensitivity 0.9979 ± 0.0009
0.9542 ± 0.0105 method (right). Further colors on the right surface
Estimation of GM tissues around the main Speci city 0.9979 ± 0.0009 indicate the di erent regions.
structures with graph-cut. Detection of
blood vessels and meninges. Average False Positive
15.8 11.1 7.4
Identi cation of the ventricle and deep GM
structures, and alignment of CSF.
Closing of the ventricle and deep GM
structures for later surface reconstruction.
Removal of blood vessels and meninges.
Morphological operations for continuous
0 0 0
Average False Negative
37.4 25.2 24.3
Bilateral alignment of all structures using
SEGC SEGPF SLAB
Figure 1: Algorithm ow diagram input and and output images and sub-steps)
Fig. 1: Algorithm ow diagram (with (with input output images and sub-steps)
0 0 0
REFERENCES Figure 3: Shown are results for 40 subjects processed using the SVE (Segmentation Validation Engine). More
Ashburner, J. (2005), ‘Uni ed segmentation’, Neuroimage, vol. 26 (3) pp. 839-51 detailedFig. 4: Shown are results for 40 subjects processed using the SVE (Segmentation Validation Engine).
results are available at http://sve.loni.ucla.edu/archive/study/?id=221.
More detailed results are available at http://sve.loni.ucla.edu/archive/study/?id=221.
Ashburner, J. (2007), ‘A fast di eomorphic image registration algorithm’, Neuroimage, vol. 38 (1) pp. 95-113
Dahnke, R. (2010), ‘Central Surface Reconstruction using a Projection Scheme’, HBM 2010
Han, X. (2007) ‘Atlas renormalization for improved brain MR image segmentation across scanner platforms’,
IEEE transactions on medical imaging, vol. 26 (4) pp. 479-86 The new introduced method is able to improve the results of the default SPM8 segmentation
Liang, L. (2007), ‘Automatic segmentation of left and right cerebral hemispheres from MRI brain volumes and allows the separation of the brain into di erent atlas-based regions. Due to the DARTEL
using the graph cut algorithm’, Neuroimage, vol. 34 (3) pp. 1160-1170 normalization, regional alignments have a high accuracy. Compared to Zhao et al., 2010, and
Mangin, F.R. (1996), ‘Shape Bottlenecks and Convservative Flow’, IEEE Work. MMBIA, San Francisco, CA, pp. 319–328. Liang et al., 2007, it allows the alignment of major structures, as well as the correction for blood
Rehm, K. (2004), ‘Putting our heads together: a consensus approach to brain/non-brain segmentation in T1- vessels and meninges and lling of subcortical structures and ventricles.
weighted MR volumes’, Neuroimage, vol. 22 (3) pp. 1262-70
Sadananthan, S.A. 2010, ‘Skull stripping using graph cuts’, Neuroimage, vol. 49 (1) pp. 225-39
Shattuck, D.W. (2009), ‘Online resource for validation of brain segmentation methods’, Neuroimage, vol. 45 (2) pp. 431-9. ACKNOWLEDGEMENTS
Zhao, L. (2010), ‘Automatic cerebral and cerebellar hemisphere segmentation in 3D MRI: Adaptive
disconnection algorithm’, Medical Image Analysis, vol. 14 (3) pp. 360-72. R.D, R.Y., & C.G. are supported by the German BMBF grants 01EV0709 & 01GW0740.
E-mail: email@example.com PDF of the poster is available at: http://dbm.neuro.uni-jena.de/HBM2010/Dahnke.pdf