Você está na página 1de 7

ALGORITHMS FOR

VISUALIZATION

A Bayes-Based
Region-Growing Algorithm
for Medical Image Segmentation
A new Bayesian-analysis-based region-growing algorithm for medical image segmentation
can robustly and effectively segment medical images. Specifically, the approach studies
homogeneity criterion parameters in a local neighbor region. Using the multislices
Gaussian and anisotropic filters as a preprocess helps reduce an image’s noise.

M
edical images from computerized shape properties. We get the parameters by ana-
tomography (CT), magnetic reso- lyzing statistical information via a specific cluster
nance imaging (MRI), and other algorithm. To reduce image noise, we use the
imaging modalities are playing in- anisotropic and multislices Gaussian filters. Finally,
creasingly important roles in clinics. Accordingly, we apply the morphological closing operation as a
analyzing these images for computer-aided diag- postprocess to remove holes in the result images.
nosis and therapy planning is becoming a more By applying a Bayes-based statistical methodol-
challenging research task. ogy to adjust parameters during the region-searching
Region growing is the most conventional and ef- process and by using multi-image-processing meth-
ficient approach in medical image segmentation, ods, we can improve the final image quality. We
which is the preliminary stage of image analysis. tested our algorithm framework on CT and MRI
(See the “Related Work in Medical Image image segmentation, and our experimental results
Segmentation” sidebar on p. 34 for details on other show that the approach is reliable and efficient.
approaches.) The key step in region-growing
methods is to define a classifying criterion that re- Bayes-Based
lies on image properties and user interaction. Region-Growing Algorithm
Bayesian analysis is a powerful tool for pattern clas- Medical volume data sets are made up of a series of
sification, so we propose a Bayes-based region- parallel slices with uniform spacing; the space be-
growing algorithm that estimates parameters by tween the nearby slices is small, so the image data
studying characteristics in local regions and con- in adjacent slices are highly correlated. As Regina
structing the Bayes factor as a classifying criterion. Pohle pointed out, values in CT images represent
Our method uses a Gaussian mixture model to de- average x-ray absorption distorted by noise and ar-
scribe the local regions’ homogeneity and image tifacts.1 The absorption itself is a constant value for
a given anatomical structure, and the noise can be
assumed to be zero-mean Gaussian noise with an
1521-9615/07/$25.00 © 2007 IEEE unknown standard deviation based on the image-
Copublished by the IEEE CS and the AIP
formation process.
ZHIGENG PAN AND JIANFENG LU In the first step of our algorithm, we use mul-
Hangzhou Dianzi University tislices and anisotropic filtering to remove the
noise in medical images. Figure 1 illustrates the

32 THIS ARTICLE HAS BEEN PEER-REVIEWED. COMPUTING IN SCIENCE & ENGINEERING


Gauss convolution filtering between multislices,
and Figure 2 gives experimental results, with
Figure 2c showing that we effectively eliminated
the noise from Figure 2b by using Gauss convo-
lution filtering.

Anisotropic Filtering
Pietro Perona and Jitendra Malik2 formulated
anisotropic filtering as a diffusion process. The Figure 1. Gauss convolution filtering between multislices. We
equation is combine every nearby slice into the last result according to the left
curve. The noise can be depressed by the nearby reference points.
It(x, y, t) =   C(|I(x, y, t)|) I(x, y, t), (1)

where I(x, y, t) is the input image’s intensity value,


and t refers to the iteration steps. C(|I(x, y, t)|) is
the diffusion function and is a monotonically de-
creasing function of the image gradient magnitude.
This is how the diffusion process will primarily oc-
cur in the regions’ interior (because of the low gra-
dient magnitude); it won’t affect region boundaries
where the gradient magnitude is large. Here, we
choose a Gauss function as function C: (a) (b) (c)

|∇I|2 Figure 2. Experimental results of using Gauss convolution filtering.


C (| ∇I |) =
2
e 2K . (2) Note the progression from (a) the original image to (b) the image
with Gauss noise to (c) the image after Gauss convolution filtering.
Parameter K controls the diffusion process and
adaptive changes with the gradient magnitude

⎛ 1 n n ⎞
K = C ⋅ sqrt ⎜ 2
⎜⎝ n
∑ ∑ |∇I ( x, y, t )|2⎟⎟ , (3)
y =1 x =1 ⎠

where n is the neighborhood’s size. Normally, we


adopt n = 3.
Figure 3 illustrates our experimental results with
anisotropic filtering.

Local Gaussian Mixture Model


A connected region forms during the region-grow- (a) (b)
ing process as we calculate each neighboring pixel
with the homogeneity criterion. Generally, this cri- Figure 3. Experimental results of using anisotropic filtering. We took
terion remains unchanged throughout the process. (a) the original image distorted with mostly random noises, and
Our algorithm calculates the homogeneous parame- achieved (b) this result after the anisotropic filtering operation.
ters by analyzing the histogram’s shape in neighbor-
hood areas. Generally speaking, we can define
homogeneity as a likelihood of belonging to a Gauss- ian approach to model selection is based on poste-
ian distribution of gray values with a given mean and rior model probabilities. Considering a set of
standard deviation. The mean is the absorption value classes, {1, 2, …, N}, and data x, the posterior
of the tissue in a CT image and the magnetization in probability of class i is
an MRI, and the standard deviation is correlated to
the noise. Thus, the mean and standard deviation p( x |ω i ) P (ω i )
help us distinguish different tissues. P (ω i | x ) = N
, (4)
We use a Bayes factor to label pixels to different ∑ p( x |ω j )P (ω j )
classes in the region-growing process. The Bayes- j =1

JULY/AUGUST 2007 33
RELATED WORK tial-information-based segmentation methods. Other meth-
ods are called clustering-based techniques. k-means is a clas-
IN MEDICAL IMAGE SEGMENTATION
sical and powerful clustering algorithm that researchers still
Rolf Adams broadly classifies segmentation techniques into apply to many applications.10–11
four classes:1 interactive thresholding,2,3 boundary detec- All these segmentation techniques have their advantages
tion,4,5 region splitting and merging,6,7 and hybrid and drawbacks, which encourages researchers to combine
methods.8 Thresholding techniques compare the value of all one or more of them into hybrid methods. Malik Jitendra,
pixels with a threshold value and identify the values of pixels for example, used a spectral graph theoretic framework of
that lie within a particular region’s range. Selecting the normalized cuts and combined intensity, texture, and con-
threshold value is the key to this method. Boundary-based tour measurements in her algorithm.12 Susanta Mukhopad-
methods use edge-detect techniques such as a gradient filter hyay’s algorithm applies multiscale morphological
to locate different regions’ boundaries, but this method is segmentation techniques.13,14
sensitive to image noise. We can also use anisotropic filtering techniques to effec-
Region-based segmentation techniques postulate that tively reduce noise in the image process while still preserving
pixels in the same region have similar features. The general region boundaries and fine details. In 1990, Pietro Perona
growing procedure starts from some initial points as seed proposed the Perona-Malik anisotropic diffusion method to
points, compares every pixel with its surrounding neighbors, overcome the drawbacks of conventional linear filters such
and if a certain merging criterion is satisfied, the pixel is clas- as Gauss filters, which usually blur an image’s sharp bound-
sified into the same class. Naturally, the choice of a merging aries and small structures.15 Perona also analyzed how the
criterion is critical to the segmentation’s success in this different parameters in filter function affect image bound-
method. Many researchers focus on this important issue to aries. Based on this technique, Guido Gerig developed multi-
improve the algorithm’s effectiveness and accuracy, so most channel anisotropic diffusion for smoothing in MRI brain
of their algorithms are sensitive to the start position and se- images.16 The major advantages of this method are efficient
quence of the initial seed points. As an alternative, Lin Zheng noise reduction in homogeneous regions and edge en-
has developed an unseeded region-growing algorithm for hancement and preservation of object contours/boundaries
image segmentation.9 Region-growing algorithms are spa- between different regions. Used as a preprocess step,

70 70 where p(x| i ) is the function of likelihood of


60 60
class  I ; and P(i) is the prior probability of class
50 50
 i and is usually obtained from training data.
40 40
30 30 The quantity
20 20
p( x |ω i ) ⋅ P (ω i )
10 10 Bij = (5)
p( x |ω i ) ⋅ P (ω i )
0 50 100 150 200 0 50 100 150 200
(a) (b) (c)
is defined as the Bayes factor for class i against
70 35
class i.
60 30
As an example, consider the gray-level distribu-
50 25
40 20
tions of the abdomen image in Figure 4. In Figure
30 15 4a, the red region is inside the tissue, so the shape
20 10 of the gray histogram in Figure 4b should approx-
10 5 imately obey the Gaussian distribution. In Figure
4e, we can clearly see that the gray histogram is
0 50 100 150 200 0 50 100 150 200 made up of two Gaussian distributions because the
(d) (e) (f) red region is just on the boundaries of the two tis-
sues. To classify the pixels according to the Gauss-
Figure 4. Gaussian distribution of an abdomen image’s ian mixture model correctly, we first estimate the
homogeneity model. Note the changes from (a) the image with a Gaussian model’s parameters and then use the
selected area inside the tissue and (b) a histogram of that area; to Bayes factor as the selection rule. To do so, we use
(c) a Gaussian curve fitting for (b); to (d) the image with a selected the expectation-maximization (EM) algorithm to
area on the edge of the tissue and (e) a histogram of that area; to fit the parameters of the Gaussian mixture densi-
(f) a Gaussian curve fitting for (e). ties. Unfortunately, the iteration algorithm uses a

34 COMPUTING IN SCIENCE & ENGINEERING


anisotropic filters can effectively eliminate noise in medical 1684–1699.
image segmentation.17 9. L. Zheng, J. Jesse, and T. Hugues, “Unseeded Region Growing for 3D Im-
age Segmentation,” Proc. ACM Pan-Sydney Workshop Visualization, ACM
References Press, vol. 2, 2000, pp. 31–37.
1. R. Adams and L. Bischof, “Seeded Region Growing,” IEEE Trans. Pattern 10. T. Kanungo et al., “An Efficient k-Means Clustering Algorithm: Analysis
Analysis and Machine Intelligence, vol. 16, no. 6, 1994, pp. 641–647. and Implementation,” IEEE Trans. Pattern Analysis and Machine Intelli-
2. P.K. Sahoo, S. Soltani, and A.K.C. Wong, “A Survey of Thresholding gence, vol. 24, no. 7, 2002, pp. 881–892.
Techniques,” Computer Vision, Graphics, and Image Processing, vol. 41, 11. L. Yingli, J. Tianzi, and Z. Yufeng, “A Split-Merge Based Region Growing
no. 2, 1998, pp. 233–260. Method for the Analysis of fMRI Data,” Human Brain Mapping, vol. 22,
3. T. Wen-Bing, T. Jin-Wen, and L. Jian, “Image Segmentation by Three- no. 4, 2004, pp. 271–279.
Level Thresholding Based on Maximum Fuzzy Entropy and Genetic Algo- 12. M. Jitendra et al., “Contour and Texture Analysis for Image Segmenta-
rithm,” Pattern Recognition Letters, vol. 24, no. 16, 2003, pp. 3069–3078. tion,” Int’l J. Computer Vision, vol. 43, no. 1, 2001, pp. 7–27.
4. R. Goldenberg et al., “Fast Geodesic Active Contours,” IEEE Trans. Image 13. S. Mukhopadhyay and B. Chanda, “Multiscale Morphological Segmenta-
Processing, vol.10, no. 10, 2001, pp. 1467–1475. tion of Gray-Scale Images,” IEEE Trans. Image Processing, vol.12, no. 5,
5. E. Sharon, A. Brandt, and R. Basri, “Segmentation and Boundary Detec- 2003, pp. 533–549.
tion Using Multiscale Intensity Measurements,” Proc. IEEE Computer Soc. 14. P. Maragos, “A Representation Theory for Morphological Image and Sig-
Conf. Computer Vision and Pattern Recognition, vol. 1, IEEE CS Press, 2001, nal Processing,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol.
pp. 469–476. 11, no. 6, 1989, pp. 586–599.
6. X. Rihua and W. Runsheng, “Range Image Segmentation Based on Split- 15. P. Perona and J. Malik, “Scale-Space and Edge Detection Using
Merge Clustering,” Proc. 17th Int’l Conf. Pattern Recognition (ICPR 2004), Anisotropic Diffusion,” IEEE Trans. Pattern Analysis Machine Intelligence,
vol. 3, IEEE CS Press, 2004, pp. 614–617. vol. 12, no. 7, 1990, pp. 629–639.
7. G.A. Borges and M.J. Aldon, “A Split-and-Merge Segmentation Algo- 16. G. Gerig et al., “Nonlinear Anisotropic Filtering of MRI Data,” IEEE Trans.
rithm for Line Extraction in 2-D Range Images,” Proc. 15th Int’l Conf. Pat- Medical Imaging, vol. 11, no. 2, 1992, pp. 221–232.
tern Recognition (ICPR 00), vol. 1, IEEE CS Press, 2000, pp. 441–444. 17. J. Weickert, “Applications of Nonlinear Diffusion in Image Processing and
8. K. Haris et al., “Hybrid Image Segmentation Using Watersheds and Fast Computer Vision,” Acta Mathematica Universitatis Comenianae, vol. 70,
Region Merging,” IEEE Trans. Image Processing, vol. 7, no. 12, 1998, pp. no. 1, 2001, pp. 33–50.

lot of computing time, so to counteract this affec- model’s parameters.


tion, we use a fuzzy c-means algorithm to estimate In our algorithm, we want to find the suitable
the parameters space. Then, we apply these para- functions u and the centroids c to minimize the
meters into the Bayes-factor-based region-grow- objective function (Equation 6). We can simplify
ing process as parameters. Equation 6 on the assumption that the two
Fuzzy c-means (FCM)—first proposed by J.C. classes cluster. This is acceptable in our segmen-
Dunn3 and frequently used in pattern recogni- tation process, and it dramatically reduces com-
tion—is a clustering method that lets one piece of puting time.
data belong to two or more clusters. The method The following pseudocode charts our algorithm:
is based on minimizing the following objective
function: 1. First, assume iNumofClasses=1 represents
pixels inside the border.
N C 2 2. Use FCM to calculate the u, c, and the error
J m = ∑ ∑ uijm x i − c j , 1  m  , (6) E(u, c).
i =1 j =1 3. If E > Eset, then set iNumofClasses=2 and go
to step 2.
where m is any real number greater than 1, uij is the 4. Get the final parameters (u, c) as the result.
degree of membership of xi in cluster j, xi is the ith
of d-dimensional measured data, cj is the cluster’s d- The parameter iNumofClass is the number of
dimension center, and ||*|| is any norm expressing classes during the clustering process, and function
the similarity between any measured data and the E(u, c) represents the degree of errors.
center. Medical image segmentation approaches of-
ten employ this clustering algorithm to distinguish Algorithm Framework Description
different tissues in the whole image data. Here, we After obtaining the parameters, we can construct
use it to estimate the local Gaussian distribution the region-growing criterion:

JULY/AUGUST 2007 35
1. Do preprocess filtering.
2. Begin the loop of the region-growing process.
3. Estimate the parameter space using FCM al-
gorithm.
4. Construct the criterion for region growing.
5. Investigate nearby pixels to see if they meet
the criterion.
(a) (b) 6. End loop.

Figure 5. Morphological operation experimental results. Some In Step 3, we calculate the parameters of the
obvious holes exist in (a) the original image, which we were able to Gaussian mixture model. Then, we apply these pa-
remove by using (b) the closing operator. rameters in Equation 7 to construct the region-
growing process’s homogeneity criterion.
Steps 4 and 5 give the Floodfill algorithm’s
loop cycle, which travels all the current region’s
neighbors and repeats until no more seeds meet
the criteria.

Morphological Operation
After the region-growing process, we have some
holes in the resulting images because of existing
noise. To fix this, we adopt a morphological oper-
ation to remove the holes by first using an erosion
operation and then applying a dilation operation.
We define the erosion operation in morphological
operators as

E(x) = XB = {x, y|Bxy  X}; (8)

the dilation operation is

(a) (b) (c) (d) D = X  B = {x, y|Bxy X = }. (9)

Figure 6. Experimental results comparisons. We achieved the Figure 5 shows some experimental results.
results with parameter k threshold values of (a) 5, (b) 10, and (c)
30 in a conventional region-growing algorithm. However, (d) our Experimental Results
Bayes-based region-growing algorithm improves results by We implemented our algorithm on the Win2k and
automatically estimating the parameters. VC++ 6.0 platforms. Figure 6 shows some experi-
mental results, using skull and liver medical data
sets. The images in the left three columns are the
⎧ ⎧⎪ Bij > 1, Pixel belongs to class i results from a conventional region-growing algo-
⎪On the border ⎨ rithm, and those on the right are the result of our
⎪ ⎪⎩ Bij < 1, Pixel belongs to class j
C grow =⎨ (7) Bayes-based region-growing algorithm. As Figure
⎪ ⎧true, if μ − cσ < x < μ + cσ
⎪ Inside the border ⎨ false, otherwise, 6 illustrates, the conventional algorithm relies on
⎩ ⎩ user interaction, forcing users to continuously ad-
just parameter k for better results. The algorithm
where the FCM computes  and  in the current we proposed overcomes this drawback by auto-
neighborhood, and C is the parameter factor. As we matically estimating the parameters.
described in the last section, we calculate the num- Table 1 compares the computational time taken
ber of classes according to the local histogram’s by the conventional algorithm and ours. Our al-
shape, and the parameter iNumofClasses repre- gorithm takes more time because it needs to com-
sents whether the pixel is inside or on the border of pute model parameters in the region-growing
the region. process, but solutions such as reducing the size of
A C-like pseudocode of our algorithm is as the local neighbor area can improve this bottle-
follows: neck. In future research, we will focus on ap-

36 COMPUTING IN SCIENCE & ENGINEERING


Table 1. Time comparison between conventional region-growing algorithm and our Bayes-based algorithm.

Data set Dimension Conventional region growing (seconds)* Our algorithm (seconds)†
k=5 k = 10 k = 30 s=5 s = 10
Skull A 256
256 0.2 0.3 0.5 3.1 4.0
Skull B 256
256 0.1 0.3 0.4 2.8 3.2
Liver 384
384 0.3 0.4 0.6 5.7 9.1

*Parameter k is the threshold value in the conventional region-growing algorithm



s is the size of the local neighbor area in the Bayes-based algorithm

proaches using graphic processing units (GPUs)


to speed up the algorithm.
Figure 7 compares the results of different seg-
mentation algorithms. The general region-grow-
ing algorithm’s results aren’t good because fixed
parameters limit the algorithm’s ability to process
growing regions. The split-and-merge algorithm
is a classical clustering method, so it only considers
the pixels’ global statistical information. Our algo-
rithm considers all the statistical and spatial infor-
mation, making the edge of segmented image exact
and smooth. (a) (b) (c) (d)
Figure 8 compares our algorithm’s computing
time for different image sizes with other segmenta- Figure 7. Experimental result comparisons. In comparing (a) the
tion algorithms. The four algorithms we compared original images (the red arrow points to the objects [teethridge
are region growing with k = 5, k = 30, a split-and- and auris] to be segmented), (b) segmentation results from
merge algorithm, and our method. Although the conventional region growing with parameters k = 30, (c) results
data size changes from 64
64 to 1,024
1,024, the from the powerful split-and-merge algorithm, and (d) results from
computing time for all the algorithms increases with our algorithm, we see that the edge of the images with our
different degrees. Conventional region growing algorithm are more exact and smooth.
gets the best performance because of its simplicity,
and the split-and-merge algorithm slows down be-
cause it must scan all the image pixels during the number NSFC 60533080, and the 973 Program of
process. Our algorithm’s computing time also in- China, under grant number 2003CB716104. We thank
creases, but in an acceptable range. Jim X. Chen for his kind suggestion to the improvement
Figure 9 shows some reconstruction experiments of this research work.
of 3D models using our algorithm; here, we used
Bayes-based region growing in each slice and then References
reconstructed the 3D models. 1. R. Pohle and K.D. Toennies, “Segmentation of Medical Images
Using Adaptive Region Growing,” Proc. SPIE Medical Imaging, M.
Sonka and K.M. Hanson, eds., vol. 4322, 2001, pp. 1337–1346.
2. P. Perona and J. Malik, “Scale-Space and Edge Detection Using

I
n the future, we plan to extend our algorithm Anisotropic Diffusion,” IEEE Trans. Pattern Analysis Machine Intel-
into 3D volume data sets in the multichan- ligence, vol. 12, no. 7, 1990, pp. 629–639.
nel color space. With the rapid development 3. J.C. Dunn, “A Fuzzy Relative of the ISODATA Process and Its Use
of PC hardware, programmable graphic in Detecting Compact Well-Separated Clusters,” J. Cybernetics,
vol. 3, no. 9, 1974, pp. 95–104.
cards can help speed up graphic applications. One
of our future research directions is to use this tech-
nology to optimize our algorithm to enhance per- Zhigeng Pan is a professor in, PhD supervisor for, and di-
formance and reduce user waiting time. rector of the Division of VR & Multimedia at the State Key
Lab of CAD&CG, Zhejiang University. He is also the vice
Acknowledgments director of the Hangzhou Center, CAD Training Network
This research work is supported by the Key National in China. His research interests are in computer graphics,
Natural Science Foundation in China, under grant virtual reality, e-commerce, multimedia computing, and

JULY/AUGUST 2007 37
50 40
k = 5 region growing 35
40 Split-and-merge 30
k = 30 region growing
30 25
Our algorithm
20
20 15
10
10
5
0 0
64 × 64 128 × 128 256 × 256 512 × 512 1024 × 1024 64 × 64 128 × 128 256 × 256 512 × 512 1024 × 1024

(a) (b)

Figure 8. Computing time comparisons. With different data sizes for (a) skull and (b) liver data sets, we see the different,
increasing time rates for all the segmentation methods.

(a) (b) (c) (d)

Figure 9. Experiment results with 3D rendering for the (a) skull model, (b) rachis model, (c) tumor model in a blood
vessel, and (d) inner ear model. We use the reconstruction method to create and render the 3D objects as the succeeding
steps for segmentation operation to make the results more observable.

Web graphics. Pan has a PhD in computer science from Jianfeng Lu is a faculty member in the Department of
Zhejiang University. He is an associate editor of the In- Computer Science in Hangzhou Dianzi University. His
ternational Journal of CAD/CAM and is a member of research interests are in scientific visualization and
IS&T, the Multimedia Committee of Image & Graphics in computer animation. Lu has a PhD in computer science
China, and the Committee of China Image and Graphics from Zhejiang University. Contact him at jflu@hziee.
Association. Contact him at zgpan@cad.zju.edu.cn. edu.cn.

How to Reach CiSE


Writers Subscription Change of Address (IEEE/CS)
Visit www.computer.org/cise/author.htm. Send an email to address.change@ieee.org. Specify CiSE.

Letters to the Editors Missing or Damaged Copies


Send letters to Jenny Stout, Lead Editor, jstout@computer.org. Contact help@computer.org. For AIP subscribers, contact
Provide an email address or daytime phone number. claims@aip.org.

Subscribe Reprints of Articles


Visit https://www.aip.org/forms/journal_catalog/order_form For price information or to order reprints, email cise@
_fs.html or www.computer.org/subscribe/. computer.org or fax +1 714 821 4010.

Subscription Change of Address (AIP) Reprint Permission


Send general subscription and refund inquiries to Contact William Hagen, IEEE Copyrights and Trademarks
subs@aip.org. Manager, at copyrights@ieee.org.

www.computer.org/cise/ or http://cise.aip.org

38 COMPUTING IN SCIENCE & ENGINEERING

Você também pode gostar