References 1. Nekula, Heřman, Vomáčka, Kocher. Radiologie. 1

advertisement
Segmentation of CT arteriography based on combination of segmentation methods1A. I.
Pirner2, M. Jiřík3, M. Železný4
2
The University of West Bohemia, ivaneck@students.zcu.cz
3 The University of West Bohemia, mjirik@kky.zcu.cz
3 The University of West Bohemia, zelezny@kky.zcu.cz
This paper describes a solution of CT arteriography vessel segmentation using a
combination of methods for segmentation of image data. Computer tomography (CT) is
one of the most useful approaches for investigating the arterial system and its
pathologies. There are three ways for segmenting the vessels from the CT image –
manually, fully automatically and with a user interaction. We considered only the two
automatical approaches, because the manual way is too time-consuming and humandependent. We used three segmentation techniques: thresholding, edge detection and
region growing. Results of these methods were consulted with a medicine expert to
determine the best procedure. None of these methods can be used separately, but a
combination of them brought very good results, which can be used in medicine.
Summary
Introduction
Computed tomography is the best noninvasive method to investigate the shape of
any organs in human body [1]. It has a better
resolution than MRI (magnetic resonance
imaging) and supplies us with a 3D density
image of the diagnosed area. A disadvantage
of this procedure is the radiation dose the
patient receives when being scanned. There
are two ways, how to get an arteriogram a. k.
a. image of arteries. The first option is two
scan the patient twice – once without and once
with added contrast substance into the blood –
so-called bolus. These two images are aligned
so that each pixel of one image corresponds
with a pixel of the other image and then
digitally subtracted. All that remains is the
arterial system. However it’s generally
difficult to proper align those images, which
can be disturbed, deformed etc.
The other option is to capture just one image
with the contrast substance and to use
segmentation
techniques
to
get
the
arteriogram. In this case the using time of the
CT machine, as well as the medical staff and
radiation dose is approximately a half of the
preceding procedure, which is a big advantage.
The CT scanner consists of a stable circular
frame, where the X-ray emitters on one side
and X-ray sensors on the other side rotate at
constant speed around the axis of the frame.
The patient is placed on a bed which moves
constantly into the circle so that the final path
of each emitter according to the patient’s body
forms a helix. This capture is then recomputed
into slices, a series of single images. The
projection plane of each of the slices is
orthogonal to the axis of examined part of the
body. The value of each pixel is the so-called
X-ray density, which defines the ability of
stuff to halt the X-rays. Density values ranges
are generally known for each type of tissue,
but the ranges of arteries with bolus and bones
overlap, so we can determine a threshold to
separate easily these two stuffs, but we can’t
discriminate between them. That means we
can’t use only thresholding, but a combination
of techniques.
1
Support of the grant (if is available)
Problem statement
Our task is to develop a procedure for
segmenting the arterial system from a series of
CT images with or without a user interaction
and to state, whether a user interaction is
needed or not. Results of the developed
procedure are to be evaluated by a medicine
expert, who tells the final decision, if the
procedure is suitable for using or not. The final
algorithm can be implemented in a medical
image processing computer in a hospital then
to serve the doctors for CT investigation of the
arterial system. The algorithm will be designed
and evaluated in MATLAB.
Approach and techniques
One of the simplest segmentation methods is
thresholding. This method converts a greylevel image into a binary image according to
the rule
and maximal value of pixel intensity of the
arteries. The initial region is defined as a seed
set. The seed set can be a single point, multiple
points or regions. The region is computed
iteratively then according to the rule: If a
neighboring pixel has its value between the
maximal and the minimal threshold, let it be
part of the region. This procedure is repeated
until no more pixels can be added to current
region. The result of this segmentation method
strongly depends on the choice of the seed set.
Our goal is to segment the arterial system,
which usually consists of a single, maximally
two 3D-connected regions. Therefore the seed
set can be composed of minimally 1-2 points,
which are surely part of the searched region.
Results
1 𝑓𝑜𝑟 𝐺(𝑖, 𝑗) ≥ 𝑇
𝐵(𝑖, 𝑗) = {
0
𝑒𝑙𝑠𝑒
where G(i,j) is a pixel in the original greylevel image at position (i,j), T is the threshold
value and B(i,j) is a pixel of the final binary
image at position (i,j). Logical 1 in the binary
image stands for foreground (object), 0 for
background. It can be used in case that the
objects to be segmented in the image are
characterized by a different brightness level
range than the background. This implies the
necessity of proper estimation of the threshold
value. When using the thresholding, many
small regions grow up. Filtering according to
the region’s size (area, volume) can be made
using the labeling algorithm, which goes
through all pixels (voxels) of the image and
assorts each connected region a label – a
unique number. When counting the occurrence
of each label, we can assess the size of each
region and delete all regions smaller than some
specific value.
Another approach to segmentation is the edge
detection, which seeks for discontinuities in
the brightness of an image. There are many
variations of operators, which approximate the
first derivative (difference in this discrete
case), we used Sobel’s.
The third main approach to image
segmentation is region growing. We used
concretely a modified confidence-connected
algorithm. We set two thresholds for minimal
The first procedure we developed uses
thresholding. Considering the constant
“lighting” of the CT images, the minimal
density value was estimated manually from a
single density histogram of the 3D-image. This
value can be used for any CT images, because
the X-ray density depends only on the scanned
tissue. Thresholding provides us with a binary
image, where bones and arteries together form
multiple connected regions. The resulting
image is 3D – the thresholding proceeds for all
slices at once. 3D-connected regions are sorted
by their size (volume) then and “small”
regions are cropped, so that only bigger
regions remain. For removing the bones we
need a mask, which is created using edge
detection. The edge image is morphologically
filled – connected regions are built from an
image of region borders. This mask is
subtracted from the thresholded image and we
get the arteriogram without bones.
The second examined procedure uses region
growing, where the seed set is defined as the
result of the preceding thresholding, which
produced a raw segmentation consisting of
multiple connected regions. The goal of this is
to get a single connected region saving some
computer time using a preceding “cheaper”
method. Thresholding is i.e. faster than region
growing and supplied in our case a partial
segmentation. The missing parts can be filled
in using region growing.
The last procedure uses region growing with a
seed set, where the seed set is defined as
separated points, in this case two, one for each
arterial tree. The visualization of results of
particular segmentation procedures is made by
means of volume rendering, which produces
2D-projections of 3D-images to be understood
easily by physicians.
growing. But the problem of choice of the size
filtering threshold value persists. This
procedure gives good results, but needs a too
complicated user interaction.
Fig. 3. Region growing.
Fig. 1. Thresholding.
As shown in Fig. 1, the thresholding
segmentation looks smooth after removing
smaller objects. However it’s not clear how to
choose the threshold of the object size, it was
chosen manually. The resulting image misses
some parts of the arterial system, that’s why
the result is not suitable for practical use.
The third technique grant as good results as the
second one, but there is no size threshold value
to choose. The user only labels two points, one
for each main artery and the process goes
automatically on. These key points could be
chosen automatically in the future, we didn’t
investigate this option yet.
Example
Let’s show the influence and output of
particular used methods upon the image. Fig. 4
shows the source CT image, 1 slice of the
series. It’s a 16 bit grey-scale image, depicted
in 8 bits.
Fig. 2. Thresholding and region growing.
The result of region growing based on the raw
thresholding looks much better; it contains
more details and doesn’t lack any arteries. It’s
pretty fast, because the basic shapes are
defined by quick thresholding and only the
missing parts are filled in using the region
Fig. 4. Original CT slice.
Fig. 5 shows the effect of thresholding.
Separate arteries are segmented well, but some
parts of bones are marked as foreground, too.
Fig. 5. Thresholded image.
In Fig. 6 we can see the result of edge
detection, which finds securely borders of
bone areas. These borders form a closed area
(scull bones), which is filled up and creates a
mask for removing bones.
medicine expert, this form was evaluated as
the most useful.
Fig. 7. Resulting segmentation with transparent
bones.
Conclusion
We developed a method consisting of a
combination
of
several
segmentation
techniques for reliable arterial system
segmentation from the images of CT contrast
angiography.
Three
approaches
were
examined and their results compared. These
procedures were consulted with a medicine
expert to determine the best of them. The
resulting procedure requests a user-interaction;
the user has to select two points in the images,
where the major arteries are located.
References
Fig. 6. Edge detection result.
Interpretation
To make it easier to understand such an image
and to locate concrete arteries, we decided to
depict both of bones and arteries in one image
with given transparency for bones. The
resulting image is shown in Fig. 7. For our
1. Nekula, Heřman, Vomáčka, Kocher.
Radiologie. 1. vydání. Olomouc : Univerzita
Palackého v Olomouci, 2001.
2. Fitzpatrick, M. Handbook of Medical
Imaging: Medical Image Processing and
Analysis. Bellingham, Washington : SPIE
Society of Photo-Optical Instrumentation
Engi, 2000. 0819436224.
3. Šonka, Hlaváč, Boyle. Image Processing,
Analysis and Machine Vision. 3rd edition.
Toronto : Thomson Learning, 2007. ISBN
049508252X.
Download