Pattern Classification in Subcortical Structures of the Human Brain

Pattern Classification in Subcortical Structures of the Human Brain
Umut Orçun Turgut1,2, Mehmet Özer Metin2
1
C/S Information Technologies R & D Center, Istanbul, Turkey
2
Informatics Institute, Department of Medical Informatics, Middle East Technical University, Ankara, Turkey
Introduction
Subcortical brain segmentation aims to separate a brain image into a number of different and disjoint sets of
voxels that correspond to anatomically meaningful regions. The segmentation of subcortical structures is a
challenging task, since the signal intensity alone is not sufficient to distinguish between different subcortical
grey matter structures. [1] Thus, segmentation algorithms usually incorporate a priori information about their
expected location and shape. Several studies, which are based not only on intensities, have been performed for
distinguishing structures from each other. [2][3][4]
The aim of this work to propose a new method for classifying the subcortical structures, namely thalamus,
caudate, putamen, amygdala and hippocampus, in both hemispheres of the human brain.
Method
The classification algorithm that we
propose is based on Support Vector
Machines (SVM) with RBF kernel.
Subjects: 34 female subjects with the
mean age of 30.3±9.5 years. None of the
subjects has brain disorder, substance
abuse, pregnancy or physical illness.
Images: MRI acquisitions are done on
1.5 T Siemens Magnetom Vision with 2
mm slice thickness, 0.859 mm in-plane
resolution, FOV=220, Flip angle=12o,
T1=300, TE=4, TR=9.7. Images are
saved in AFNI and ANALYZE formats.
AFNI is used for performing manual
Talairach
registrations,
whereas
FreeSurfer is used for automatic
extraction of the whole brain, intensity
normalization and segmentation of the
subcortical structures. Preprocessing:
Brain
extraction,
intensity
normalization, Talairach registration and
region of interest (ROI) selection were
performed on all of the 34 subject data
so as to simplify the whole process.
Figure 1: Pipeline of the method
Feature Vector: Each input vector
consisted of 27 elements consisting of 1
probability value from the probabilistic atlas (LONI), 3 spherical coordinates, 9 signal intensity values along the
largest gradient including the current voxel under examination, 12 signal intensity values along each of the three
orthogonal values (±2 voxels from the current voxel along x, y, and z), mean value of the intensities of cubic 1
and cubic 10 (mean value of the intensity of ±1/±10 voxels from the current voxel along x, y, and z, respectively)
voxels. All input vector values were normalized to get a value between 0 and 1. Training and Classification: 5
of the subject data were blindly selected for training operations and the rest were used for classification
operations. Moreover, these 29 data had also been examined with FreeSurfer scripts and results were gathered
for further comparisons with the ones that our method produced. The training unit is SVMLight and called from
out Matlab script. The output of SVMLight component is our model (SVM classifier) for that structure. The
classification step is going to deal with the parcellation of five subcortical structures. The segmentation
algorithm that was used is SVM Classifier, which is a part of STPRtool of Matlab. In this step, the models for
each subcortical structure and the feature vector for each voxel in the input brain were used in the decision of the
class that the voxel belongs to. The pipeline of our study is displayed in figure 1.
Results
Totally 29 subject’s brain data were processed under the same steps. The results that we have obtained by the
method of this study and their comparisons with the ones obtained from FreeSurfer, for one of our subjects, are
displayed in the figures 2 and 3.
Conclusion
Atlas and probability based automated segmentation methods had been developed to automatically delineate
subcortical and cerebellar regions of interest. Relative overlap of the subcortical and cerebellar structures by the
template based method ranged from 0.59 to 0.84. [5] The probabilistic atlas approach had similar results as
compared to the template method. In this study, we have calculated the relative overlaps ranging from 0.61 to
0.91 for the structures. In table 1, the dice coefficients of the 5 different subjects are displayed.
Figure 2: Segmentation of caudate, displayed in axial slice, by our model (left) and FreeSurfer
(right).
Figure 3: Segmentation of thalamus, displayed in coronal slice, by our model (left) and FreeSurfer
(right).
Table 1: Dice coefficients calculated from the overlap of subcortical structures classified by the model and the manually
segmented atlas for the 5 subjects.
References
[1] Fischl, B., Salat, D.H., Busa, E., Albert, M., Dieterich, M., Haselgrove, C., van der Kouwe, A., Killiany, R.,
Kennedy, D., Klaveness, S., Montillo, A., Makris, N., Rosen, B., Dale, A.M. (January 2002), ‘Whole brain
segmentation: automated labeling of neuroanatomical structures in the human brain’, Neuron 33, 341–355.
[2] Khan, A.R., Wang, L., Beg, M.F. (2008), ‘Freesurfer-initiated fully-automated subcortical brain
segmentation in MRI using large deformation diffeomorphic metric mapping’, NeuroImage 41, 735–746.
[3] Pohl, K.M., Bouix, S., Nakamura, M., Rohlfing, T., McCarley, R.W., Kikinis, R., Grimson, E.L., Shenton,
M.E., Wells, W.M. (2007), ‘A hierarchical algorithm for MR brain image parcellation’, IEEE Trans. Med. Imag.
26 (9).
[4] Powell, S., Magnotta, V.A., Johnson, H., Jammalamadaka, V.K., Pierson, R., Andreasen, N.C. (2008),
‘Registration and machine learning-based automated segmentation of subcortical and cerebellar brain structures’,
NeuroImage 39 (1), 238–247.
[5] Wu, M., Carmichael, O., Lopez-Garcia, P., Carter, C.S., Aizenstein, H.J. (2006), “Quantitative comparison of
AIR, SPM, and the fully deformable model for atlas-based segmentation of functional and structural MR
images”, Hum. Brain Mapp. 27, 747–754.