Chapter 4
Skull Stripping-2
4.1
Single Intensity Mapping and Largest
Connectivity Algorithm(SIMALC) for Skull
Stripping Brain MRI
The two methods discussed in chapter 3 have a severe limitation that they can
extract brain tissues only from slices that have a clear separation dark ring
around the brain tissues. The Single Intensity rvIapping and Largest Connectivity (SIMALC) algorithm proposed here, proceeds without any initial parameters.
In this chapter a method is proposed to regenerate the cerebrospinalfluid (CSF)
from the skull stripped image which initially has only white matter and gray
matter. This method can be successfully applied to all the images in the MRI
set. The proposed algorithm is tested on multiple slices of images from different
depths both on axial and coronal view of Tl weighted MR images.
4.2
Brain Extraction Tool A Review
A survey of existing skull stripping techniques is given in the previous chapter.
One of the popular method, the Brain Extraction Tool (BET) developed by Smith
38
4.3 SIMALC Algorithm for Skull Stripping
(M.S.Smith, 2002) is reviewed in this section. The lower and upper threshold
points are determined from the intensity histogram. This gives a rough brain
and nonbrain threshold. The centre of gravity of the head image is found, along
with the rough size of the head in the image. A triangular tessellation of a spheres
surface is initialized inside the brain and allowed to slowly deform, one vertex at
a time, following forces that keep the surface well spaced and smooth, while
attempting to move toward the brains edge. If a suitably clean solution is not
arrived at, then this process is re-run with a higher smoothness constraint. The
major fallout of this algorithm is the manual selection of the threshold points and
also the initial seed point selection which might not always yield reliable results.
The stripping results highlighting the limitation of the BET method is shown in
fig. 4.8 on page 51.
4.3
SIMALC Algorithm for Skull Stripping
In the anterior and posterior images the dark ring of separation between the brain
and nonbrain tissues become less pronounced, because of the high intensity pixels representing the eyeballs and other nonbrain tissues. The stripping technique
adopted for mid brain images will yield lot of extra tissues along with the brain
tissues which need to be cleaned up. The proposed method overcomes this difficulty by mapping pixels using intensity and connectivity. The algorithm consists
of five steps.
4.3.1
Computing Histogram
Histogram of the original image Foriginal is determined and the peaks of the histogram are identified. The peak corresponding to the lowest intensity region,
represents the CSF and bone tissues. All the pixels in this region is masked off
to zero. The rest of the pixels are mapped to a single intensity.
39
4.3 SIMALC Algorithm for Skull Stripping
4.3.2
Multiple Region Mapping
The previous step divides the entire image into two classes. The pixels with zero
intensity and the other pixels mapped to some arbitrary value. F
= {I1,I2},
Class .f1 of the image F has all low intensity values including CSF masked to
zero and .f2 is another class of image which has pixels representing all bright
pixels indusive of brain and nonbrain tissues. The image F has single valued
pixels of .f2 disjoint with 0 values of II. A new image mask is built checking for
connectivity of all pixels of class f2.
If a pixel F(:r, y) or if anyone of its neighbours F(x, y - 1), F(x - 1, y -
1), F(:z: - 1, y) or F(:z: - 1, y
+ 1)
is a nonzero pixel belonging to .f2 then the
corresponding pixel in the new mask belong to region R1. The next nonzero
pixel is similarly checked, if this pixel satisfies the above said condition then it is
added to the region R1, otherwise a new region R2 begins. The new region starts
growing with an incremented value. The next nonzero pixel is assigned to the
new region R2 and this grows until it is intercepted by another pixel belonging
to class f1.
This process is continued till all the pixels are checked.
{R1, R2, R3, R4
Fm
=
}, F m represents the new image mask grown comprising all
disjoint regions R1, R2, R3 etc. where each consecutive region is separated by
regions of zeros. This can be visualised by observing the fig. 4.1.
The following algorithm explains the different steps involved in region growing
4.3.2.1
Algorithm for Region Growing
• Initialise n
= 1, which marks the first region.
• If F(x, y) is nonzero, assign it a value say n in the mask Fm .
• If F(x, y - 1) is nonzero then F(x, y) gets its value of n. i.e it gets
added to that region.
• If the above condition is not satisfied then check for the other neigh-
bours in the previous row.
If anyone of them is nonzero, then F(x, y) gets added to that region
or simply retains the previous value for n.
40
4.3 SIMALC Algorithm for Skull Stripping
Figure 4.1: Figure shows multiple connected region with brain tissues as the
single largest connected region
• If F(x, y) is not connected to any of its neighburs then increment n.
Incrementing n means that a new region starts growing.
• Repeat the above steps for all the pixels.
A new image mask F m is thus grown, where pixels have intensity values
representing the region.
4.3.3
Region Merging
The previous algorithm results in multiple regions representing the brain. The
next step is to connect all these multiple regions in the brain into a single region
while keeping the skull as a separate region. If a pixel in the mask, Fm(x, y) E R2
and is connected to R1 thenFm(x, y) is merged with Rl. This process merges the
smaller disconnected regions into larger region. The algorithm for the same is
given below.
41
4.3 SIMALC Algorithm for Skull Stripping
4.3.3.1
Algorithm for Region Merging
• Start from left bottom and move vertically upwards column-wise.
• If F m (x, y) and Fm (x - 1, y) are nonzero pixels and belong to different
regions merge them to a single region
• Repeat this for all the pixels
4.3.4
Largest Connected Region
The histogram of the image Fm is determined. The highest peak identifies the
region with maximum number of pixels and this represents the largest connected
region. And through iterations more and more nonzero and N 4 (four connected
) pixels are added to the largest connected region. Pixels F(:r:, y - 1), F(x, y
+
1), F(:r: - 1, y) and F(:r: + 1, y) are said to be four connected neighbours of the
pixel F (:r, y) . If a pixel belong to the laregst region then its N 4 nonzero neighbours are also added to this region. Pixels are added simultaneously from bottom
upwards vertically column-wise and bottom upwards horizontally row-wise. This
reduces the number of iterations in growing the connected region. Brain, the
largest connected component is shown in fig. 4.1.
R max = 1\la:r{R1,R2,R3,R4.... }
The pixels in the largest connected region is mapped to the original values in
the MR image according to the equation given below. This results in the skull
stripped image Fss .
Fss(x, y) = Foriginal(X, V);
Fss(x, y) = 0;
if Fm(x, y) E Rmax
Otherwise
The skull stripped image F ss has both WM and GM pixels while CSF pixels
being removed along with the skull tissues. The volume atrophy of WM and GM
42
4.4 Segmentation of CSF pixels
tissues can significantly alert about the onset of AD. We have also proposed a
method to recover the CSF pixels.
The proposed SIMALC algorithm needs human intervention in some cases: l.
In the first few slices the largest connected component will be nonbrain tissues,
in such cases the brain is extracted by subtracting the stripped image from the
original image. 2. Another exception is in the images where the CSF forms a
virtual boundary at the center, in such slices the brain will be extracted in two
halves in two stages.
4.4
Segmentation of CSF pixels
The CSF pixels being low intensity gets eroded along with the pixels representing
the skull in the above proposed algorithm. Any attempt to retain this results in
unwanted extra cranial tissues along with the brain tissues. The lost CSF was
regenerated using two different approaches.
4.4.1
Boundary Outline
An outline of the boundary of the extracted brain is marked by identifying the
end pixels in each of the row. A binary image, Fbin is developed where all pixels
bounded by the end pixels are marked one. The stripped image is obtained in the
original intensity value by using eqn.4.l. The CSF pixels along the outer edge
can not be smoothly recovered through this step. The edge information is lost
in most of the slices and might result in under segmentation of the CSF pixels.
Mapping the trapped pixels enclosed by the skull stripped image is easily done
using this method. The edges could be grown through seed region growing (SRG)
and the gradient along the outer periphery of the skull as a stopping criterion,
the target pixels can be added. As SRG method is computationally expensive
another alternative method is proposed in the next section.
(4.1)
43
4.5 Result and Discussion
4.4.2
Connectivity Algorithm
The main attempt in the proposed work is to develop a segmentation algorithm
that can be generalised and applied on all the slices without any delimitation. Developing successful skull stripping technique, completes half of the segmentation
process. The multiple mapped region of the MRI has brain as the single largest
connected region. The skull is left in multiple intensity sections. The skull region
is mapped into a single largest region after removing the extracted brain tissues.
The skull has layers of scalp, muscles and fat and on the lower end extension of
the skull region. Each one of these tissues are represented using different intensity
pixels. All these are mapped using LOIlIH'ctivity criterion into a single intensity
region as shown in fig. 4.2(b). The skull is roughly sized to the dimension of the
skull stripped image. This enables in discarding the extended skull portion which
usually represents multiple cross sections and even zero intensity region similar
to the CSF. The first part of the LOnnectivity algorithm for skull stripping would
have left zero intensity pixels within the skull tissues, these zero intensity regions
are filleel with the same intensity as representeel by the skull. The sized down
skull and the skull with uniform intensity are shown in fig. 4.2(c). Removing
the skull from the original image results in the skull stripped image having all
three brain components WM, Gl'vI and CSF shown in fig. 4.3(a). Subtracting this
image from the initial skull stripped image yields CSF. The CSF thus extracted
can still have extra tissues, which could be easily discarded as they are either
isolated regions or regions of high intensity. Connectivity algorithm is applied at
the final stage to accurately segment the CSF as shown in fig. 4.3(b).
4.5
Result and Discussion
The algorithm automatically extracts the brain from all images. The results on
the images of Tl weighted axial and coronal views are shown in fig. 4.5 and
fig. 4.6. The algorithm was tested on images from different subjects at different
points of time. The slices are carefully chosen to demonstrate the robustness of
the algorithm across the data set with various topological complexities. There can
be no rule in nature which says the anatomic structures are same and stereotype
44
4.5 Result and Discussion
(a) The original image and the skull stripped image
(b) The skull extracted after mapping to single intensity and re-sized
(c) The skull extracted showing zero intensity pixels within and the second image shows the zero intensity pixels mapped.
Figure 4.2: The extration of skull through different stages
45
4.5 Result and Discussion
(a) 'fhe skull stripped image showing Wl\I, GM and CSF
(b) The Segmented CSF
Figure 4.3: The skull striPped images and the segmented CSF
46
4.5 Result and Discussion
as we move from one subject to another. This truth makes it impossible for
any single algorithm to run with perfection from one image to another. Human
interventions and modifications become a part of any automated algorithm. Thin
differences show up in few cases calling for manual intervention, in such cases
we used the previous similar slice as template image. Using similar slices as
template image reduces the computational task and it also directly extracts the
CSF. If the closely resembling slices of rvIR images are grouped together, then
one template image in each group will simplify the skull stripping process. Fig.
4.7 demonstrates one of the cases of using template image. The computational
time for skull stripping varries between 4 to 5 seconds per image, on PC with
1.73 GHz and 512 MB RAM.
Some slices in the tvIRI set show certain exceptions, where some portion of
the brain may he represented as disconnected from the largest component. Radiologist can easily add such isolated portions by introspecting the fig. 4.1. In
some of the slices the largest connected component shows a virtual split due to
the CSF running in between. In such slices growing the first largest and second
largest component yield the complete brain and it involves two stages. Fig. 4.4
shows two typical slices with the CSF running in between causing a virtual split,
in such cases the stripping results in only one half of the brain and the algorithm
is executed in two steps. The second step extracts the remaining half and thus
the complete brain is extracted.
In threshold based methods the threshold value has to be chosen after trying
out the algorithm on all the slices and the human intervention is needed. The
BET developed by Smith M.S.Smith (2002) begins with an approximate threshold
chosen between brain and skull tissues. This results in unwanted extra tissues in
many a slices and also reworking on the initial seed points. The level set method
Haihong Zhuang (2004) section works with the initial chosen center and also the
radius for growing the function is manually set by the user. Most of the existing
techniques discussed in 3.2 on page 28 needs manual intervention in every slice.
Fig. 4.8, shows the results of BET and Brain Surface Extraction (BSE), we can
notice the unwanted tissues being retained. The proposed method is thus superior
as this disposes the need to manually set a threshold between the brain and the
skull tissues and neither does it require any initial seed points to proceed.
47
4.5 Result and Discussion
•••
••
••
••
(a) Typical slice no. 140
(b) slice 141
Figure 4.4: Examples of slices showing brain as two connected component
48
4.5 Result and Discussion
Figure 4.5: Tl weighted axial view images: first column original images from
different slices, second column skull stripped images with WM and GM, third
column skull stripped images with CSF regenerated
49
4.5 Result and Discussion
Figure 4.6: Tl weighted Coronal images: first column original images from
different slices, second column skull stripped images with WM and GM, third
column skull stripped images with CSF regenerated
•
Figure 4.7: Top row: The template image, and the bottom row: Stripped image
using template image showing CSF
50
4.5 Result and Discussion
Figure 4.8: Results of BET and BSE (courtsey:Chen Yunjie[2009])
51