A novel Skull Stripping Method for T1 Coronal and
T2 Axial Magnetic Resonance Images of Human Head
Scans Based on Resonance Principle
K.Somasundaram1, R. Siva Shankar2
Image Processing Lab, Department of Computer Science and Applications,
Gandhigram Rural Institute-Deemed University, Gandhigram-624302, Tamil Nadu, India.
1&2
Abstract - In this paper we propose a novel method for skull
stripping or brain extraction from T1 Coronal and T2 Axial
Magnetic Resonance Images (MRI) of human head scans.
Brain extraction is done by first detecting the boundary
separating brain tissue and non-tissues. The pixel values at
the bright boundary will be ≈ 255. This property is utilized to
generate resonance behaviour at the boundary. We make use
of the exponential function to generate the resonance
condition using which the boundary is detected. Using the
boundary, the skull portion is removed and the brain is
extracted. The experiments on two T1 and a T2 volumes show
satisfactory results.
Keywords: Skull stripping, segmentation, resonance method,
MRI processing
1
Introduction
Magnetic resonance image (MRI) analysis is a noninvasive, non ionising and non-destructive imaging technology
to study the structural anatomy of human organs. This can
produce high quality and highly detailed images, which can
almost give every angle of organs and tissues. MRI of the
brain gives the anatomy of brain that is helpful to diagnose the
brain related diseases. MRI guided surgery like angiogram,
breast biopsy is directed accurately after knowing the result of
MRI scan. Skull striping is an important pre-processing
technique in MRI analysis. Skull stripping methods are
classified into three types: intensity based, morphology based
and deformable model based. Region based methods view the
brain regions as a group of connected pixel data sets. These
regions will have muscles, cavities, skin, optic nerves, etc. The
extraction of the brain region from the non-brain region is
done by using methods like region growing, watershed and
morphological methods.
Several works for segmentation of brain have been
reported in [1] - [8]. Justice et al.[1] proposed a semi
automatic segmentation method using 3D seeded region
growing (SRG). In that a semi-automatic method which
effectively segments imaging data volumes by using 3D region
growing guided by initial seed points has been used. Seed
voxels may be specified interactively with a mouse or through
the selection of intensity thresholds. Segmentation proceeds
automatically following seed selection only on few slices in
the volume due to the 3D nature of the region growth. The
method proposed by Adams et al.[2] requires the input value
for the number of seeds, either individual pixels or regions,
which controls the formation of regions into which the image
will be segmented. Brummer et al.[3] proposed a fully
automatic algorithm that starts with a histogram-based
thresholding preceded by an image intensity correction
procedure. This step is followed by a morphological
operations which refines the binary mask images. Anatomical
knowledge essential for the discrimination between desired
and undesired structures is implemented in this step through a
sequence of conventional and novel morphological operations,
using 2-D and 3-D operations. The final step of the procedure
performs overlap test between current and previous slice.
Lemieux et al.[4] proposed an automated algorithm to segment
the brain portion from T1-weighted volume MRI. The
algorithm uses automatic computation of intensity threshold
and morphological operations. It is a three-dimensional
method and therefore independent of scan orientation. Hohne
et al.[5] proposed a semi-automated segmentation algorithm
based on region growing and morphological operations. This
segmentation is performed concurrently with 3D visualization
providing direct visual feedback to guide the user in the
segmentation process. Jong and Lee.[6] proposed an
algorithm, after eliminating the background voxels using
histogram analysis. Two seed regions of the brain and nonbrain regions were automatically identified using a mask
produced by morphological operations. Then these seed
regions are expanded with a 2D region growing algorithm
based on general brain anatomy information. An automatic
method for brain extraction was proposed by and Stella et
al.[7]. This method uses an integrated approach which
employs image processing techniques based on anisotropic
filters, snake contouring technique, and a priori knowledge,
which is used to remove the eyes, a tricky structure in brain
MRI. It is a multistage process, involving removal of the
background noise leaving a head mask, finding a rough outline
of the brain and refinement of the rough brain outline to a
final mask. In an earlier work[8] we used Ridler’s method,
morphological operations to extract brain from T2 weighted
MRI.
In this paper we propose a brain extraction scheme using
resonance method to detect brain-skull boundary. The
remainder of the paper is organized as follows. In section 2,
we present our method. In section 3 results and discussions are
given. In section 4 the conclusion is given.
2
Proposed Method
2.1
Skull – Brain Interface
We model the skull-brain boundary as an interface of
two regions. It is well known, at the interface of two media,
interfacial waves propagate along the boundary. The
amplitude of such waves remain constant along the interface
boundary and decay exponentially in a direction perpendicular
to the interface. In a plane geometry, at the interface of waterair, hydrodynamic surface waves propagate along the
interface with constant amplitude [9], but decay exponentially
in a direction perpendicular to the interface.(see Fig.1).
Fig.1 Decay of amplitude in an interfacial wave.
Hydromagnetic interface waves propagate in a similar way in
a plasma-plasma interface embedded in a magnetic field [10].
We make use of a similar property to detect the boundary
between skull-brain interface. At the boundary, made of white
pixels, the intensity value will be ≈ 255 in a gray scale image.
Therefore, the boundary can be detected by using the
resonance function:
(1)
R(x,y)=A
Where, A is an arbitrary constant, △f(x,y)=255-f(x,y), and
f(x,y) is the intensity value of the input image at the coordinate points (x,y). Therefore, R will be very large
(resonance condition) at the boundary, where f(x,y) ≈ 255 and
will be small at the points away from the boundary. Hence by
computing the value of R and traversing the co-ordinates (x,y)
where R(x,y) gives highest value, the boundary of the brainskull can be identified and extracted.
2.2
Brain Boundary Detection
In MRI of head scans, two boundaries prominently appear.
The inner boundary is the brain-CSF (Cerebro spinal fluid)
interface, and the outer boundary is the CSF-Skull
boundary(Fig.2). If we are able to detect the inner boundary
then the brain portion can be extracted easily.
Fig.2 The prominent boundaries in MRI.
To detect the inner boundary we start computing the
resonance function R from the mid point of each row of the
middle slice. Since in both T1 and T2 MRI volumes, the
middle slice contains the brain as a single largest region.
Hence identifying the brain area in the middle slice is easy.
The mid point of each row is computed by dividing the total
width and height of the image by 2.
midx = image_width/2
- (2)
midy = image_height/2
- (3)
Hence the process starts from the midpoint to get the seed
point which is the first occurrance of the resonance(R) on both
sides from mid point. We repeat this process for each row for
the whole image. The closely placed inner most resonance
points (CPP) are connected to form a boundary. The boundary
will be formed by analysing the value of R at 5 co-ordinate
points at each row in both left and right hand side from the
middle. The points that are not close to the innermost
boundary are discarded. The innermost contour thus formed is
the boundary of the brain. The flow chart of our scheme is
given in Fig.3.
Start
A
Input the next Slice
Start from mid point of the slice
Compute R for each Row
Find first R values from mid to right and left
Select co-ordinate point of the highest intensity
value among the R values
Connect closely placed points (CPP)
Form Contour by selecting points satisfying R & CPP
Extract brain from original MRI
Select mask from extracted edge
A
Yes
Any more
Slice ?
In the IBSR T1 1_24 dataset brain portions wont be
available after 55th slice, but our method detect a small part in
61 and 62 as brain portion. In the WBA T2 dataset brain
portion is absent after 50th slice. Our method is not able to
detect brain portion in 50th slice, because the brain portion
appears in multiple parts, where as the skull contains closely
placed pixels with good intensity values. after 50th slice brain
portions wont be available.
3.2
Performance Evaluation
We carried out experiments by applying our method on the
Three volumes of T1 and T2 weighted images. For
performance evaluation of our method we made quantitative
and qualitative analysis. For quantitative analysis we
computed Jaccard and Dice coefficient similarity indices.
The Jaccard coefficient is given by[17]:
(4)
The Dice coefficient is given by[18]:
No
Stop
Fig.3 Flow Chart of the Proposed Method.
In any MRI brain volume, the middle slice contains brain as
a single largest region and is the largest brain portion in the
entire volume. Hence identifying the brain area in the middle
slice is easy. After finding the brain in the middle slice, The
extracted brain portion of the middle slice is used as a
reference to extract brain portion from adjacent slices lying
above and below it. We then move from middle slice to top
and middle to bottom slice, one direction at a time. For each
slice, the mark of the previous slice is used as a reference to
extract brain in the current slice.
3
3.1
Results and Discussions
Materials Used
For our experiments we used three MRI volumes. The first
two are T1 weighted coronal datasets collected from
International Brain Segmentation Repository (IBSR)[15]
maintained by Center for Morphometric Analysis
Massachusetts General Hospital, USA (1_24 and 13_3). The
third is a T2 weighted axial MRI dataset collected from Whole
Brain Atlas(WBA).[16] maintained by Department of
Radiology and Neurology at Brigham and women’s hospital,
Harward Medical school, Boston, USA. T1 weighted 1_24
data set contains 65 slices and 13_3 date set contains 57 slices.
Slice thickness = contiguous 3.1mm and each of 256*256
pixel size. T2 weighted dataset contains 56 slices. slice
thickness ≈ 5mm and each of 256*256 pixel size. For each
dataset, the hand segmented brain portion or gold standard is
available in the respective websites.
(5)
where, A is a data segmented using our method and B is hand
segmented data. The value of J and D vary between 0 to 1.
The best results will be very close to 1, when both results are
similar. The computed values of J and D using the proposed
method and that of Brain Extraction Tool(BET).[19] are given
in Table 1. The best values are given in bold. We observe
from Table 1 that our method gives better results than that of
the popular method BET.
Table 1: Computed average values of Jaccard and Dice
Co-efficient.
DataSet
BET
Proposed method
Jaccard
Dice
Jaccard
Dice
T1 1_24
.9459
.9722
.9568
.9746
T1 13_3
.9453
.9615
.9595
.9786
T2
.9531
.9760
.9583
.9867
For visual inspection we also give the brain portion extracted
using our method. For visual comparision we give original
images, extracted brain portion by BET and by our method.
Fig.4 shows the original slices of T1 weighted coronal 1_24
dataset. Fig.5 brain portion extracted from the dataset 1_24.
Fig.6 shows the original slices of T2 weighted dataset. Fig.7
shows the extracted from dataset of T2 weighted axial image
from 9th slice to 48th slice. Fig.8 shows BET slices of
containing regions like neck and skull portions and extracted
brain portions by our method.
Fig.4 The original slices of T1 weighted coronal dataset 1_24.
Fig.5 Extracted brain portions from T1 weighted coronal 1_24
by our method.
Fig.7 Extracted brain portions from T2 axial MRI by our
method.
Fig.6 The original slices of T2 weighted axial dateset.
Slice 12
Slice 26
Slice 34
Slice 48
Fig.8 Brain extraction by the proposed method where BET
failed. Row 1 : the original slices 12, 26, 34, 48 of 1_24
coronal dataset. Row 2 : brain extracted by BET method. Row
3 : the extracted brain portions by our method.
4
Conclusions
We have proposed a novel method based on interfacial
resonance phenomena to extract brain portion from T1 coronal
and T2 axial MRI of head scan images. This method is able to
detect the boundary of brain skull directly and thus avoids the
processing of background and skull areas. The proposed
method gives better results in terms of the Dice and Jaccard
co-efficients for both T1 and T2 Images than that of popular
method BET.
ACKNOWLEDGEMENTS
The authors like to thank University Grants Commission
(UGC), New Delhi, Grant no:M.R.P, F.No-37-154/2009(SR),
for supporting this work.
5
References
[1]
Justice, R.K., Stokely, E.M., Strobel, J.S., Ideker, R.E.
and Smith, W. M., Medical image segmentation using
3D seeded region growing. Proc. SPIE Med. Imag.
vol.3034 ,pp.900-910, 1997.
Adams, R. and Bischof, L., Seeded region growing,
IEEE Trans. Pattern Anal. Mach. Intell. vol 16 , pp.641646., 1994.
Brummer, M.E., Mersereau, R.M., Eisner, R.L., and
Lewine, R.R.J.. Automatic detection of brain contours in
MRI data sets. IEEE Trans. Med. Imag. vol.12, pp.153166, 1993.
Lemieux, L., Hagmann, G., Krakow, K., and Woermann,
F.G. Fast, accurate, and reproducible automatic
segmentation of the brain T1-Weighted volume MRI
data. Mgn. Reson. Med. vol.42, pp. 127-135., 1999.
Hohne.K.H and Hanson.W.A., Interactive
3D
segmentation of MRI and CT volumes using
morphological operations. J. of Comput.Assist. Tomogr.
vol.16, pp. 285-294, 1992.
Jong Geun park, Chulhee Lee., “Skull Stripping based on
Region growing for Magnetic resonance brain
images”.,NeuroImage, vol.40, pp.1394-1407, 2009.
Stella Atkins, and Blair T Mackiewich., “Fully
Automatic segmentation of the Brain in MRI”., IEEE
tracnsations of Medical Imaging, vol.17, pp.98-107,
1998.
Somasundaram.K., and Kalaiselvi.T., “Fully Automatic
brain extraction algorithm for axial T2-weighted
magnetic resonance images”., computers and biology and
Medicine, vol.40, pp.811-822, 2010 .
Fluid Mechanics, L.D.Landau and E.M.Liftshiftz,
Pegraman Press, Oxford, pp.238, 1959.
Satyanarayanan.A., and Somasundaram.K., “Alfven
surface waves along coronal streamers”., Astrophysics
and Space Science, vol.109, p.p. 357-364, 1985.
Somasundaram.K. and Siva Shankar.R., “Skull Stripping
of MRI Using Clustering and Resonance method “,
International Journal on Knowledge Management & ELearning, vol.3,pp.19-23,2011.
Somasundaram.K, and Siva Shankar.R., “Skull Stripping
based on Exponential Function“, National conference on
Signal and Image Processing, pp. 154-156,Feb.2012.
James B.Scarborough., “Numerical Mathematical
Analysis”., Oxford & IBH publishing Co., Sixth edition,
1966.
Milan Sonka, Vaclav Hlavac and Roger Boyle., “Image
Processing, Analysis and Machine Vision”., Thomson
Learning Inc., Second Edition, 2007.
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15] International Brain Segmentation Repository, Center for
Morphometric Analysis Massachusetts General Hospital,
CNY-6, Building 149, 13th Street, Charlestown,
MA, 02129-USA.
http://www.cma.mgh.harvard.edu/ibsr/ibsr_data/sec3.sub
2.html
[16] The Whole Brain Atlas(WBA), Department of
Radiology and Neurology at Brigham and women’s
hospital, Harward Medical school, Boston, USA.
http://www.med.harvard.edu/aanlib/navt.html
[17] Jaccard.P., The Distribution of Flora in Alpine Zone,
New Phytol, vol.11, pp.37-50,1912.
[18] Dice.L., Measures of the Amount of Ecologic
Association between Species, Ecology, vol.26, pp.297302, 1945.
[19] Smith.S.M., Fast robust automated brain extraction.
Human Brain Mapping, vol.17, pp.143-155, 2002.
http://www.fmrib.ox.ac.uk/analysis/research/bet.