Sie sind auf Seite 1von 11

Artif Intell Rev (2013) 39:225235

DOI 10.1007/s10462-011-9268-0

New spatial based MRI image de-noising algorithm


M. A. Balafar

Published online: 14 June 2011


Springer Science+Business Media B.V. 2011

Abstract Nowadays, fast scan techniques are used to reduce scanning times. These techniques raise scanning noise level in MRI systems. Instead of progress made in image
de-noising, still, it is challenging. A novel edge-preserving neighbourhood filter for image
enhancement is proposed. The main focus of this paper is to propose an adaptive filtering
function to account for the image content while try to preserve edge of image. Proposed
algorithm uses the edges of image to do edge-preserving neighbourhood filtering. Contribution of a sample, in neighbourhood of a pixel, in filtering, depends on the space between
the pixel and the sample. In fact, the sample which there is edge between it and the pixel
dont contribute in the grey level estimation. Promising experimental results on simulated
and real brain images and comparison with state-of-art de-noising algorithm demonstrate the
potential of proposed algorithm.
Keywords

De-noising algorithms MRI Brain

1 Introduction
The Medical images almost are stored and represented digitally; and this facilitates diagnosis
(Ping-Lin and Teng 2007). Medical imaging types mostly are as follow: ultrasound images,
X-ray computed tomography, digital mammography, and magnetic resonance image (MRI)
(Tian et al. 2007). Researches commonly use MRI images in medical image processing.
MRI images mostly contain fine and complicated structure. Therefore, de-noising of MRI
images remains a challenging issue and one should choose a trade-off between degree of
de-noising in one side and resolution, degrading of fine structures and acquisition speed in
other side. However, noise and inhomogeneity (Balafar et al. 2010a) are two main factors that
degrade performance and usefulness of other Medical imaging algorithms such as segmentation algorithms (Balafar et al. 2008a,b, 2010b,c,d). Moreover, generally noise degrades

M. A. Balafar (B)
Department of Computer, Faculty of Engineering, Tabriz University, Tabriz, Azerbaijane shargi, Iran
e-mail: balafarila@yahoo.com

123

226

M. A. Balafar

the performance of clustering algorithms (Lorena et al. 2008; Fernndez and Isasi 2008;
Delibaic 2009; Smail 2009).
State of art de-noising methods refers to global and local methods (Le Pennec and Mallat 2005). They vary from standard to more advanced ones and from general to specific
ones. These methods have advantages and disadvantages and none of them is better than
others regard to computation cost, degree of de-noising, quality of de-noising and boundary
preserving.
Linear filters are conceptually simple. They update the value of a pixel by (weighted)
average of its neighbourhood. These filters reduce noise but degrade image details and the
edges of the image. In contrast to linear filters, nonlinear filters have better performance in
edge preserving but degrade fine structures. Anisotropic nonlinear diffusion (Ardizzone et al.
2008) is a modification of the linear diffusion methods. This method is more powerful in
reducing noise in flat regions and reducing the diffusivity at the edges of image. Usually this
method tends to remove some high frequency components.
Markov Random Field method (MRF) (Heng et al. 2007) incorporates spatial correlation
information to preserve fine details. Usually, this method is computationally expensive. Wavelet-based methods (Anand and Sahambi 2008) perform in frequency domain and attempt to
distinguish signal from noise while try to preserve the signal in the de-noising process. These
methods do not preserve fine details, especially in low SNR images (Tisdall and Atkins 2005).
Analytical correction method attempts to estimate noise and subsequently noise-free signal
from observed image. This method uses Maximum Likelihood Estimation (MLE) (Sijbers
et al. 1998) to estimate noise-free signal and have the problem same as wavelet-based methods. Non-Local (NL) method tries to take advantage of the redundancy in image to preserves
the edges of image (Avanaki et al. 2008; Buades et al. 2005). In the image with textured
or periodic case, due to large redundancy, this method successfully reduces noise and in
mean time, preserves the geometrical edges of the image (Gudbjartsson and Patz 1995).
In MRI images, there may be non-repeated details. The following is the reasons for these
details: noise, complicated structures, blurring in acquisition, and the partial volume effect
originating from the low sensor resolution. NL can eliminate such details.
State-of-art de-noising methods have a trade-off between de-noising in one side and preserving detail of the image in other side. In this paper, to preserve the edges of de-noised
image, a novel definition for the neighbourhood of the pixels is proposed that restrict two
pixels in different sides of edges to be the neighbour for each other. Moreover, Mean filter
based on proposed definition for neighbourhood is presented. Anisotropic filter and NonLocal (NL) methods preserve the edges of the image in some extent. However, they can
consider two points in different sides of edges as neighbour that cause edges degradation.
In the rest of this paper, we presented our proposed definition for neighbourhood, proposed
methods, experiments, conclusion, and references.
2 Methods
In this section, new Edge-preserving neighborhoods and an Edge-preserving mean which
updates the value of pixel to the average of its postulated neighbourhood are presented.
2.1 New edge-preserving neighbourhood
The focus of this section is presenting a new edge-preserving neighbourhood. The new neighbourhood uses the edges of the image to specify the neighbours of a pixel. A rectangular

123

New spatial based MRI image de-noising algorithm

227

Fig. 1 a Brain image. b Brain image along with the edges of the image specified with black points

window with size n n is centred on this pixel. A sample (one of the pixels in the window
except centred one) in the window is a postulated neighbour of the pixel if there is not any
edge between that sample and the pixel.
The following is steps of an algorithm to obtain postulated neighbourhood of a pixel:
1.
2.

3.

4.

5.

The edges of the input image are identified and marked. Figure 1a represents a T1 brain
image and Fig. 1b shows the same image with the edges marked using black points.
In order to obtain the postulated neighbours of pixel, a rectangular window with size
n n is centred on this pixel. Figure 2 represents pixel B0 and a 5 5 window around
that. B1 . . . B24 are the neighbours of pixel B0 and the neighbours with underlined letters
represent edge points.
The samples of the window are scanned in consecutive skew lines. Each scan line starts
from pixel B0 and ends in a boundary point of the window. Two adjacent scan lines start in
centre of window and end in two adjacent points in window boundary. The following set
is the set of skew lines in Fig. 2: {{B0 B9 B5 }, {B0 B9 B4 }, {B0 B8 B3 }, {B0 B7 B2 }, . . .}.
In order to obtain points in a scan line {B0 B9 B5 }, equation of line connecting centre
of window (B0 ) and point in window boundary (B5 ) is obtained. Then the points are
obtained.
Whenever scan of each line reaches an edge point, the remaining samples of line are
labelled and they are not considered as postulated neighbours. For example in line
{B0 B18 B24 }, B18 is the edge point, therefore, B18 and the points after that {B24 } are
labelled and not considered as postulated neighbours.
After finishing the scan of each line, the scanning of the consecutive line starts. In the
Fig. 2, the points in window with italic letters represent postulated neighbours.

Figure 3a shows an object, a pixel, the window around pixel and the edges of the object
fitted in the window. Figure 3b represents the window after labelling. Figure 3b shows that
in a few lines scan do not label the samples outside the edges. This happens because the line
passes the edges but do not touch any edge point. Figure 4 represents different situations

123

228

M. A. Balafar

Fig. 2 A pixel (B0 ) in image


and the window around it.
The underlined letters represent
the edges of image fitted in the
window and the italic letters
represent the postulated
neighbours

B1
B6
B11
B15
B20

B2
B7
B12
B16
B21

B3
B8
B0
B17
B22

B4
B9
B13
B18
B23

B5
B10
B14
B19
B24

Fig. 3 a A pixel, the window around the pixel (the box) and the edges of image in the window. b the postulated
neighbourhood (white points). c the postulated neighbourhood with new condition for touching points (white
points)

(a)
*

(b)
*

+
+

*
*+
+
+

+
*
+

+
*

+
*
*
*

Fig. 4 The situations when scan line pass the edges of the image in the window around the pixel. a The scan
line touches the edges of the image. b The scan line does not touch the edges of the image. Where + denotes
the edges in the window and * denotes the samples of the scan line

where line passes the edges. Where + denotes the edge points and * denotes the samples
of line.
In the Fig. 4a, the scan line passes the edges and touch them but in Fig. 4b, the scan line
passes the edges but do not touch them (other example is point B5 in line {B0 B9 B5 }). To
overcome this, sample (x, y), in the scan line, is considered touching edges if it passes one
of the following criteria:

It is an edge point.
The scan line at that sample cross two connected edge points (Fig. 4b).

Figure 3c represents the window after labelling with the new definition. The black points
demonstrate labelled points. Figure 3c shows that scan labelled all the points outside of the
edges.

123

New spatial based MRI image de-noising algorithm

229

2.2 Edge-preserving mean


Edge-preserving mean updates the value of pixel to the average of its postulated neighbourhood. The Edge-preserving mean updates Ii , the value of considered pixel i to the average
value of samples in the postulated neighbourhood of i as follow:
si =

ei j I j /

jI

ei j

(10)

jI

where si is the value of pixel i in average image.



1 if I j is in postulated neighbourhood of pixel i
ei j =
0 otherwise

(11)

Image averaging reduce noise but have degrading effects such as blurring. In low level of
noise, image degrading effect is more than noise reduction effect and clustering of original
image produce better results than clustering of average image. In order to overcome this
shortcoming, a threshold based on variance of noise is used to trade off between original
image and averaged image. Threshold value of 3 is used which is obtained by error and trial.
In this algorithm, the xi average of neighbour pixels around pixel xi is calculated. Then a
linearly-weighted sum of xi and xi is used for image de-noising. That is:
oi = (1 )xi + xi

(12)

where oi represents the value of pixel i in generated image. The parameter determines the
weight of neighbourhood information. We set its value to ( 2). Where is the variance
of noise. The value 2 is obtained using error and trial.
In MRI, noise is Rician distributed (Gudbjartsson and Patz 1995). Noise distribution
approaches Gaussian with increasing Signal to Noise Ratio (SNR) and approaches Rayleigh
with decreasing SNR (Gudbjartsson and Patz 1995). Rician distribution in background is
Rayleigh because there is not any signal. The Rayleigh PDF of the statistically independent
observations is
p({Oi }) =

n

Oi (O 2 )/(2 2 )
e i
2

(13)

i=1

where O is observations and 2 is the variance of noise. The variance is obtained by maximizing the log-likelihood of PDF function with respect to variance:
N2 oise =

n
1  2
Oi
2n

(14)

i=1

3 Results and discussion


Manjon et al. (2008) reported best parameters values for NL-mean and showed that
NL-mean with these parameters performs better than other de-noising algorithms. Therefore, the proposed de-noising (Edge-preserving mean) and NL-mean de-noising algorithms
with mentioned parameters are simulated in MATLAB and tested on the simulated images
from BrainWeb and real MRI images from the IBSR. The results of two algorithms are
compared quantitatively to investigate their performance. The neighbourhood size N for

123

230

M. A. Balafar

Edge-preserving mean is set to 3 3. The similarity index (Zijdenbos and Dawant 1994) is
used to evaluate the algorithms quantitatively. The similarity index is the degree of matching
between ground truth and segmentation result for a class pixels. It is defined as
=

2|X i + Yi |
.
|X i | + |Yi |

(15)

where X i represents class i in ground truth and Yi represents the same class in the segmentation result.
3.1 Simulated image
The simulated MRI images are obtained from the BrainWeb. Extra-cranial tissues are removed
prior to segmentation. A simulated data volume with T1-weighted sequence from Brainweb,
slice thickness of 1 mm, volume size of 217 181 181 are used. The number of tissue
classes in the segmentation is set to three, which corresponds to grey matter (GM), white
matter (WM) and CSF. Background pixels are ignored in experiment.
Two de-noising algorithms are applied on a slice of simulated image. Figure 5 represents
the segmentation results of applying both algorithms on 90th slice of the simulated 3D volume
with 7% rician noise. Part (a) is noisy image. Part (b) is ground truth. Parts (c),(d) are segmentation results of Edge-preserving mean and state-of-art de-noising algorithm (NL-mean),
respectively. Figure 5 shows that NL-mean produced some artefacts. Edge-mean produced
clear result. The places with artefact in NL-mean result are specified.
The similarity index of results is used to evaluate two algorithms quantitatively. The similarity indices () of segmentation results in Fig. 5 are presented in Fig. 6. Figure 6 shows that
similarity index of Edge-preserving mean for the images with 7% is higher than NL-mean.
Then, two de-noising algorithms are applied simulated MRI brain volume (3D) and
average similarity value is used to evaluate them. The average similarity indices for
Edge-preserving mean and NL-mean are shown in Fig. 7.
Edge-preserving mean in image with 9% noise has similarity indices higher than NL-mean
and when the noise level is increased from 3% to 9%, its similarity indices decreases more
slowly than NL-mean.
The speed of two algorithms is also investigated. Figure 8 represents time required to
cluster a slice using different algorithms. Edge-preserving mean takes 2 s to de-noise a slice
and NL-mean takes 28 s to de-noise a slice. These results show that Edge-preserving is much
faster than NL-mean. NL-mean uses a similarity window to calculate similarity of each pixel
to its neighbours which is time consuming.
Since NL does not consider the edges of the image. Therefore two pixels in different
side of the edges can be considered as neighbour of each other. In addition, NL eliminates
non-repeated details which may be exist in MRI images. That is why Edge-preserving mean
produces results better than NL-mean.
As mentioned the best parameters (such as neighbourhood size) for NL-means already are
investigated and we used those parameters. Therefore, it is not necessary to investigate the
effect of neighbourhood size on performance of NL. The effect of different neighbourhood
sizes on performance of Edge-preserving mean is investigated.
Figure 9 shows the average similarity index of Edge-preserving mean when with different neighbourhood sizes is applied on simulated image with 9% noise. Figure 9 shows
that when the neighbourhood size is increased, the similarity index of Edge-preserving mean
decreases slightly.

123

New spatial based MRI image de-noising algorithm

231

Fig. 5 The segmentation results of applying two methods on 90th slice of the simulated image with 7% Rician
noise. b Ground truth. c Segmentation results of Edge-preserving mean. d Segmentation results of NL-mean

Similarity index

0.934
The proposed

0.933 noise reduction


0.932

algorithm, 0.9337

0.931
0.93

NL-mean, 0.93

0.929
0.928

Algorithms

Fig. 6 The similarity index of results in Fig. 5

123

M. A. Balafar

Average similarity index

232
0.98
0.96
0.94
0.92
The proposed noise
reduction algorithm
NL-means

0.9
0.88
0.86
3%

5%

7%

9%

Noise levels

Average time (second)

Fig. 7 The average similarity indices for Edge-preserving mean and NL-mean in different noise level

30

NL-mean, 28

25
20
15
10
5

The proposed
noise reduction
algorithm, 2

Algorithms

Fig. 9 The average similarity


index of Edge-preserving mean
when with different
neighbourhood sizes is applied on
simulated image with 9% noise

Average similarity index

Fig. 8 Average times required filtering a slice using different algorithms

0.92
0.87
0.82

The proposed noise


reduction algorithm

0.77
0.72
3

Neighbourhood sizes

3.2 Real images


The superiority of proposed algorithm is also demonstrated on real MRI image. The real
MRI images are obtained from the IBSR by the Centre for Morphometric Analysis at Massachusetts General Hospital. 20 normal data volumes with T1-weighted sequence are used.
In IBSR, manual segmentation results are provided along with brain MRI data to encourage introducing new segmentation algorithms and evaluate their performance.
Trained investigators used a semi automated histograms on the spatially normalized images
to obtain manually segmentation.
Edge-preserving mean (proposed algorithm) and NL-mean are applied to a real MRI
volume with size 256 256 65 and similarity index is used to compare the results, quantitatively. Figure 10 shows the similarity indices of proposed algorithm for WM and GM in

123

New spatial based MRI image de-noising algorithm

233

Similarity index

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0

Slice numbers
GreyMatter

WhiteMatter

Fig. 10 The similarity index of proposed algorithm when are applied for a real image

The proposed noise


reduction
algorithm, 0.7767
0.78

Average similarity index

0.77

0.76

NL-means, 0.7428
0.75

0.74

0.73

0.72

Algorithms
Fig. 11 The average similarity indices, for both algorithms when are applied on a real image

every MRI image slice. The average similarity index values of both algorithms for GM and
WM were presented in Fig. 11. Figure 11 shows that Edge-preserving mean produces the
higher similarity indices .
Edge-preserving mean (proposed algorithm) and NL-mean are also applied to all 20 normal real MRI volumes and similarity index is used to compare the segmentation results,
quantitatively. The average similarity index values of both algorithms for all 20 normal
images were presented in Fig. 12. Figure 12 shows the average similarity index values of
both algorithms for all 20 normal images.

123

M. A. Balafar

Average similarity index

234

0.765
0.76

The proposed noise


reduction algorithm,
0.7610 0.0449

0.755
0.75
0.745
0.74
0.735

NL-means,
0.7341 0.0417

0.73
0.725
0.72

Algorithms

Fig. 12 The average similarity index of two de-noising algorithms for all 20 real images

4 Conclusion
Data acquisition, processing and visualization techniques facilitate diagnosis. Medical image
de-noising has very important rule in many computer aided diagnostic tools. One of main part
of these tools is to design an efficient de-noising algorithm. Medical images mostly contain
fine and complicated structure. Therefore, de-noising of medical images is a challenging and
complex task.
In this paper, an adaptive filtering function is proposed that account for the image content
and try to preserve the edges of image. Proposed algorithm uses the edges of image to do
edge-preserving neighbourhood filtering.
In order to demonstrate the effectiveness of our method, it is applied to several medical
images. Experiments demonstrate the effectiveness of our algorithm in compare to state-of-art
algorithm.

References
Anand CS, Sahambi JS (2008) MRI denoising using bilateral filter in redundant wavelet domain. TENCON
2008. In: Proceedings of IEEE region 10 conference, 1921, pp 16
Ardizzone E, Pirrone R, Gallea R, Gambino O (2008) Noise filtering using edge-driven adaptive anisotropic
diffusion. computer-based medical systems, 2008. CBMS08. In: Proceedings of 21st IEEE international
symposium on, 1719, pp 29 34
Avanaki AN, Diyanat A, Sodagari S (2008) Optimum parameter estimation for Non-Local means image
de-noising using corner information. signal processing, 2008. ICSP2008. In: Proceedings of 9th international conference on, 2629, pp 861863
Balafar MA, Ramli AR, Mashohor S, (2010) A new method for MR grayscale inhomogeneity correction. Artif
intell rev. doi:10.1007/s10462-010-9169-7
Balafar MA, Ramli AR, saripan MI, Mahmud R, Mashohor S (2008) Medical image segmentation using
anisotropic filter, user interaction and Fuzzy C-Mean (FCM), communications in computer and information science. Springer, Berlin, pp 169176
Balafar MA, Ramli AR, saripan MI, Mahmud R, Mashohor S (2008) Medical image segmentation using
Fuzzy C-Mean (FCM), Learning Vector Quantization (LVQ) and user interaction, communications in
computer and information science. Springer, Berlin, pp 177184
Balafar MA, Ramli AR, Saripan MI, Mashohor S (2010) Review of brain MRI image segmentation methods.
Artif Intell Rev 33(3):261274
Balafar MA, Ramli AR, saripan MI, Mashohor S, Mahmud R (2010) Improved fast Fuzzy C-Mean and its
application in medical image segmentation. J Circuits Syst Comput 19(1):1324
Balafar MA, Ramli AR, Saripan MI, Mashohor S, Mahmud R (2010) Medical image segmentation using
Fuzzy C-Mean (FCM) and user soecified data. J Circuits Syst Comput 19(1):114
BrainWeb [Online]. Available: www.bic.mni.mcgill.ca/brainweb/
Buades A, Coll B, Morel J-M (2005) A Non-Local algorithm for image denoising. CVPR, pp 6065

123

New spatial based MRI image de-noising algorithm

235

Delibaic B, (2009) Reusable components for partitioning clustering algorithms. Artif Intell Rev 32(14):
5975
Fernndez F, Isasi P (2008) Nearest prototype classification of noisy data. Artif Intell Rev 30(14):5366
Gudbjartsson H, Patz S (1995) The rician distribution of noisy MRI data. Magn Reson Med 34:910914
Heng S, Shi-Xi W, Ke-Feng J, Wen-Xian Y (2007) Bayesian despeckling to SAR images based on the membrane MRF model, synthetic aperture radar, 2007. APSAR 2007. In: Proceedings of 1st Asian and Pacific
conference on, 59, pp 347350
IBSR. Available: http://www.cma.mgh.harvard.edu/ibsr/
Le Pennec E, Mallat S (2005) Sparse geometric image representations with bandelets. IEEE Trans on Image
Process 14(4):423438
Lorena AC, de Carvalh AC, Gama JMP (2008) A review of the combination of binary classifiers in multiclass
problems. Artif Intell Rev 30(14):1937
Manjon V et al (2008) MRI denoising using Non-Local means. Med Image Anal 12:514523
Ping-Lin C, Teng W-G (2007) Exploiting the self-organizing map for medical image segmentation. CBMS,
pp 281288
Sijbers J, den Dekker AJ, Scheunders P, Dyek DV (1998) Maximum likelihood estimation of rician distribution
parameters. IEEE Trans Image Process 17:357361
Smail L (2009) De-separation and computation of probability distribution in Bayesian networks. Artif Intell
Rev 31(14):8799
Tian D, Fan L (2007) A brain MR images segmentation method based on SOM neural network. ICBBE,
pp 686689
Tisdall D, Atkins MS (2005) Mri denoising via phase error estimation. In: Proceedings of SPIE medical
imaging, pp 646654
Zijdenbos A, Dawant B (1994) Brain segmentation and white matter lesion detection in MR images. Crit Rev
Biomed Eng 22(56):401465

123

Das könnte Ihnen auch gefallen