Study of Denoising Algorithm for MR Image Based on Contourlet Transform

Lei Guo1, Guizhi Xu1, Lei Zhao2,3, Ming Hu1, and Ying Li1
1

Province-Ministry Joint Key Laboratory of Electromagnetic Field and Electrical Apparatus Reliability, Hebei University of Technology, Tianjin 300130, China 2 Department of Radiology, Harvard Medical School and Brigham & Women's Hospital, Boston, MA 02115, USA 3 XinAoMDT Technology Co., Ltd., Langfang, Hebei 065001, China filter banks that can deal effectively with images having smooth contours. The CT not only possess the main features of The WT (multi-resolution and time-frequency localization), but also show a high degree of directionality and anisotropy. Its approximation behavior for 2-D function is smooth everywhere except discontinuities along a smooth curve. The main difference between the CT and other multi-scale directional systems is that the CT allows for a different and flexible number of directions at each scale. In addition, the CT employs iterated filter banks which makes it computationally efficient [4]. As a new effective signal representation tool, the CT was applied in many image processing tasks including denoising, compression and feature extraction [5]. Denoising image by means of the CT, however, introduces the Gibbs-like artifacts in the reconstruction image due to its lack of the translation invariance [6]. A Cycle Spinning (CS) technique is employed to develop translation invariant contourlet denoising scheme to improve the quality of the image. One of the most important characteristics of the brain MR image is the complicated changes of gray level, i.e., abundant in high frequency information. In this study, Combined the CT and the CS, the CTCS was investigated to remove the three kinds of noise in the brain MR image. II. CONTOURLET TRANSFORM WITH CYCLE SPINNING

AbstractIt is more difficult to remove the noise and keep the edge information in the same time in the brain MR image than the ordinary image since the boundaries of encephalic tissue are highly complicated. As one kind of the multi-scale and multidirectional geometrical analysis method, the Contourlet Transform (CT) is an optimal representation of the contour and texture information in an image and overcomes the drawbacks of the Wavelet Transform (WT). Due to the lack of translation invariance of the CT, a Cycle Spinning (CS) technique was employed to smooth Gibbs-like artifacts. In this study, Combined the CT and the CS, the CTCS with different threshold for each direction of each scale was investigated to remove the noise in the brain MR image. Theoretical analysis and experimental results showed that the CTCS is outperforms the CT and the WT both visually and in terms of the SNR. Keywords-medical image processing; denoising algorithm; contourlet transform; cycle spinning

I.

INTRODUCTION

With the development of medical imaging technology, clinical diagnosis and image processing have become more and more dependent on the high quality of medical image. However, noise is introduced into MR image due to the inhomogeneity of magnetic field, the excursion of temperature and the motion of tissue etc. [1]. So, denoising process is obligatory to obtain high quality of MR image. The edge information and the noise are the high frequency part in an image. The aim of denoising image is to remove the noise and keep the edge information. Traditional Fourier Transform method makes the image blurred since the edge information and the noise are removed in the same time. Due to the idea of the multi-scale, the Wavelet Transform (WT) can keep the edge information to some extent while the noise is removed since it exhibits strong dependencies both across scales and between neighbor coefficients within a subband, especially around image edges [2]. 2-D WT is the tensor product of the 1D WT. As a result of a separable extension from 1-D bases, the major drawbacks of the WT in 2-D are the following: it is only good at isolating the discontinuities at edge points, but can not see the smoothness along the contours in an image structure; it has limited ability in capturing the directional information [3]. The Contourlet Transform (CT), proposed by M. N. Do and M. Vetterli in 2002, can overcome the drawbacks of the WT. The CT representation is based on an efficient 2-D non-separable
This work was supported in part by the National Natural Science Foundation of China (No. 50937005 and No. 51077040) and the Natural Science Foundation of Hebei Province, China (No. H2012202035).

A. Contourlet Transform The CT, a true two-dimensional transform that can capture the intrinsic geometrical structure, is an optimal or sparse representation, especially for the contour and texture information in an image. It is a double iterated filter bank structure, the Pyramidal Directional Filter Bank (PDFB) which can simultaneously hold multi-scale, localization, nearly critical sampling, flexible directionality and anisotropy. The PDFB is a cascade of a Laplacian Pyramid (CT) and a Directional Filter Bank (DFB). In essence, it first uses a wavelet-like transform (LP) to capture point discontinuities for image edge using multi-scale analysis, then followed by the DFB to link point discontinuities into linear structures using directional analysis [7]. The overall result is an image expansion using basic elements like contour segments. So this method is named contourlet transform. The structure of the contourlet filter bank is shown in Figure 1. The frequency partition is shown in Figure 2.

978-1-4577-1759-8/12/$26.00 2012 IEEE

lowpass lowpass LP

{
bandpass directional subbands

(l j ) j 0 , n, j , k , n ( t ) j j 0,0 k 2 l j 1, nZ 2

(5)

(2,2)

DFB bandpass

For the denoising algorithm based on the threshold, optimal threshold is a key problem. A soft threshold method is proposed inspired by literature [8]. This method is fit for image denoising based on the CT since the threshold is different for each direction of each scale. It can be described as:
j ,k 2 log( j + 1) j Sth j ,k =
(l ) j,k = Median W j ,kj c

DFB image bandpass

bandpass directional subbands

(6) (7)

( , )

Where Sth j ,k is the threshold of k th direction of j th scale; j , k is the noise variance; c = 0.6745 is an empirical value.

( , )
Figure 2 The frequency partition of contourlet transformation

The contourlet filter bank has an associated continuousdomain expansion in square intergrable space L2 R 2 . With orthogonal filters, through its iterated filter bank, the space is decomposed into multi-scale and multi-directional sub-spaces:

( )

B. Cycle Spinning Like the WT, both the LP and the DFB lacks of the feature of translation invariance due to sub-sampling in its filter bank structure. In the neighborhood of the singularity points, the CT can exhibit the Gibbs-like artifacts in the reconstruction image, alternating undershoot and overshoot of a specific level. The CS is introduced into the CT with translation invariant scheme. The effect of the CS is to weaken the Gibbs-like phenomenon through changing the location of singularity points. For an image g (x, y ) = f ( x, y ) + ( x, y ) with the size of N M and noise , the operator of the CS is defined as:

Ci , j (g ) = g [mod(x + i, N ), mod( y + j, M )]

(8)

2 (l ) L2 R 2 = V j 0 W j ,kj j j 0 k = 0

( )

lj

(1)

Where j , k and n specify the scale, direction and location respectively. The detail subspace Wl at scale 2 j is decomposed into 2
lj

Its inverse transform is Ci , j ( g ) 1 = C i , j (g ) . The process of the CT with the CS including decomposition stage and reconstruction stage is described as:
= CT 1 Ci , j Sth CT Ci , j (g )

( ( ( (

)))1 )

(9)

directional subspaces W j , kj . The approximation

(l )

subspace V j 0 is spanned by an orthogonal basis of a scaling function and its translation:

j 0, n ( t ) l

= j 0 (t 2 j n ) nZ 2

(2)

where CT denotes the CT decomposition; CT 1 denotes the CT reconstruction; Sth denotes the soft threshold operator. n1 and n2 denote circulant shift numbers of horizontal and vertical direction respectively. The output image is the averaging shift as following:
O= 1 1 2 i , j n1n2 i =1, j =1
n ,n

Each subspace W j ,jk is spanned by a tight frame of a translation family of contourlet functions:

(10)

(l j ) j , k , n (t )

j = j,k t 2 j 1 Sk j n

(l )

(l )

)}

nZ 2

(3)

III.

EXPERIMENTAL RESULTS AND ANALYSIS

Sk

(l j ) diag ( 2

l j 1 l 1 , 2) for 0 k 2 j 1 l j 1 l 1 l ) for 2 j k 2 j 1 diag (2, 2

(4)

Combined all scales, directions and locations, the following family with a sequence of finite positive integers l j j j 0 is a tight frame for L2 R 2 :

( )

{}

Two MRI datasets were acquired using a fast-spin echo pulse sequence in a Centauri 0.3T MRI system (XinAoMDT Technology Co., Ltd., China) under approved IRB: one is T1 weighted MR images with 256 256 matrix size and 5mm slice thickness; the other is T2 weighted MR images with 256 256 matrix size and 1mm slice thickness. Figure 3 shows the original T1 and T2 weighted MR images. To verify the denoising performance of the CTCS, three kinds of noise were added into MR images: the first is

Gaussian noise; the second is salt and is pepper noise; the third is speckle noise. For the LP stage, the 9-7 biorthogonal filter was used decompose the image into 4 scales; for the DFB stage, direction is partitioned into 3, 4, 8 and 16 directional subbands from coarse to fine scales respectively. Threshold selection is based on Eq. (6) and Eq. (7). The number of circulant shifts is 64. Compared with the CTCS, the CT and the WT were also implemented for denoising the same MR images. The Signal to Noise Ratio (SNR) was used to assess the performance of denoising algorithms.

weighted MR images are shown in Figure 7 and Figure 8 respectively for three algorithms. The performances of three algorithms for salt and pepper noise are shown in Table II. Figure 9 shows the 4 curves of the SNRs with different salt and pepper noise variances ranged [0 0.1] for one noisy image and three denoisy images respectively.
1 6 1 4 1 2 1 0 8 S N R 6 4 2 2 0 -2 -4 0 0 .0 1 0 .0 2 0 .0 3 0 .0 4 0 .0 5 0 .0 6 Va ria n ce 0 .0 7 0 .0 8 0 .09 0 .1 0 -2 0 S N R 8 6 4 G a ussia n -S N R W T -S N R C T -SN R C T C S-S N R 1 4 1 2 1 0 G a ussia n -S N R W T -S N R C T -SN R C T C S-S N R

0 .0 1

0 .0 2

0 .0 3

0 .0 4

0 .0 5 0 .0 6 Va ria n ce

0 .0 7

0 .0 8

0 .09

0 .1

(a) the curves for T1 image Figure 6 (a) T1 MR image Figure 3 (b) T2 MR image Original MR images

(b) the curves for T2 image

The curves of SNRs with variances for Gaussian noise

The Gaussian noise with variance=0.03 was added to the MR images. Three denoising algorithms (the CTCS, the CT and the WT) were implemented to remove the noise. The visual denoising results of T1 and T2 weighted MR images are shown in Figure 4 and Figure 5 respectively. The performances of three algorithms for Gaussian noise are shown in Table I. Figure 6 shows the 4 curves of the SNRs with different Gaussian noise variances ranged [0 0.1] for one noisy image and three denoisy images respectively.

(a) noisy image Figure 7

(b) the WT

(c) the CT

(d) the CTCS

The visual denoising results for T1 image with salt and pepper noise

(a) noisy image Figure 8 (a) noisy image (b) the WT (c) the CT (d) the CTCS Figure 4 The visual denoising results for T1 image with Gaussian noise

(b) the WT

(c) the CT

(d) the CTCS

The visual denoising results for T2 image with salt and pepper noise

TABLE II THE PERFORMANCES OF THREE DENOISING ALGORITHMS FOR SALT AND PEPPER NOISE Noise WT CT CTCS SNR(dB) SNR(dB) SNR(dB) SNR(dB) T1 MR image 4.16 6.16 9.90 10.90 T2 MR image 5.11 6.94 10.13 11.09

2 0

1 5 S a lt & p e p pe r-S N R W T -SN R C T -S N R C T C S -S N R 1 0 S a lt & p e p pe r-S N R W T -SN R C T -S N R C T C S -S N R

(a) noisy image Figure 5 TABLE I

(b) the WT

(c) the CT

(d) the CTCS

1 5

The visual denoising results for T2 image with Gaussian noise

1 0 S N R 5 5 0 -5 0 0 .0 1 0 .0 2 0 .0 3 0 .0 4 0 .0 5 0 .0 6 Va ria n ce 0 .0 7 0 .0 8 0 .09 0 .1 0 0

THE PERFORMANCES OF THREE DENOISING ALGORITHMS FOR GAUSSIAN NOISE Noise WT CT CTCS SNR(dB) SNR(dB) SNR(dB) SNR(dB) T1 MR image 2.37 5.30 6.83 7.31 T2 MR image 2.85 5.90 7.42 8.02

0 .0 1

0 .0 2

0 .0 3

0 .0 4

0 .0 5 0 .0 6 Va ria n ce

0 .0 7

0 .0 8

0 .09

0 .1

(a) the curves for T1 image Figure 9

(b) the curves for T2 image

The curves of SNRs with variances for salt and pepper noise

The salt and pepper noise with variance=0.03 was added to the MR images. The visual denoising results of T1 and T2

The speckle noise with variance=0.03 was added to the MR images. The visual denoising results of T1 and T2 weighted

MR images are shown in Figure 10 and Figure 11 respectively for three algorithms. The performances of three algorithms for speckle noise are shown in Table III. Figure 12 shows the 4 curves of the SNRs with different speckle noise variances ranged [0 0.1] for one noisy image and three denoisy images respectively.

(a) noisy image

(b) the WT

(c) the CT

(d) the CTCS

Figure 10 The visual denoising results for T1 image with speckle noise

of denoising speckle noise using CTCS is not stable for different kinds of images. As shown in Figure 6, 9, 12, the experimental curves also show that the performance of the CTCS is the best and the CT is better than the WT with the different variances. And performances of denoising algorithms are stable for the Gaussian noise and the salt and pepper noise with the different variances, but they are not stable for speckle noise, especially at the range [0 0.04]. We concludes that the CTCS denoising algorithm is fit for the Gaussian noise and the salt and pepper noise, but not for speckle noise since the Gaussian noise and the salt and pepper noise are considered as additive noise and the speckle noise is considered as multiplicative noise. The average running times were 0.88 seconds, 8.73 seconds and 572.31 seconds for the WT, the CT and the CTCS respectively (implemented in MATLAB on a PC with a 2.16GHz CPU and 2GB of RAM). Although the time complexity of the CTCS is much higher than that of the CT and WT, it is still practical and tolerable. Also, the improvement of medical image quality is very important for clinical practice, such as surgical planning and image-guided therapy applications etc. IV. CONCLUSION

(a) noisy image

(b) the WT

(c) the CT

(d) the CTCS

Figure 11 The visual denoising results for T2 image with speckle noise TABLE III THE PERFORMANCES OF THREE DENOISING ALGORITHMS FOR SPECKLE NOISE Noise WT CT CTCS SNR(dB) SNR(dB) SNR(dB) SNR(dB) T1 MR image 13.70 17.61 18.76 19.92 T2 MR image 12.84 14.42 14.50 14.75
2 2 S p eckle -SN R W T -SN R C T -S N R C T C S -S N R 2 0 S pe ckle -S N R W T -SN R C T -S N R C T C S -S N R

2 0

1 8

1 8

1 6

1 6 S N R S N R

1 4

1 4

1 2

High quality of medical image has the significance for clinical practice and image analysis. Denoising image is a important means. Overcoming the drawbacks of the WT, the CT has the advantages of multi-scale, directionality, localization, critical sampling and anisotropy. In order to eliminate the Gibbs-like artifacts resulting from the CT denoising process, the CT technique is employed in developing a translation invariant contourlet-based denoising scheme. In this study, the CTCS is investigated to remove three kinds of noise in the brain MR images. Theoretical analysis and experimental results showed that the CTCS is outperforms the CT and the WT both visually and in terms of the SNR. In addition, the CTCS is fitter for denoising the Gaussian noise and the salt and peper noise than denoising the speckle noise. REFERENCES

1 2

1 0

1 0

8 0

0.0 1

0 .0 2

0 .0 3

0.0 4

0 .05 0 .0 6 V a ria n ce

0.0 7

0 .08

0 .0 9

0 .1

6 0

0 .0 1

0 .0 2

0 .0 3

0 .0 4

0 .0 5 0 .0 6 Va ria n ce

0 .0 7

0 .0 8

0 .09

0 .1

[1] [2]

(a) the curves for T1 image

(b) the curves for T2 image

Figure 12 The curves of SNRs with variances for speckle noise

The CT is the optimal representation for two dimensional image information and capture smooth contours in images using a variety of elongated shapes with different aspect ratios. As shown in Figure 4, 5, 7, 8, 10 and 11, the visual denoising results shows that the CTCS is the best and the CT is better than the WT since the CT can get more edge and direction information than the WT and the Gibbs-like artifacts are eliminated by the technique of the CS. As shown in Table I, II and III, the experimental data also shows that the performances of the CTCS is the best and the CT is better than the WT with variance=0.03 in term of the SNR. Furthermore, compared the CTCS with the WT, the SNRs is improved 2.01 (T1 image) and 2.12 (T2 image) for the Gaussian noise; the SNRs is improved 4.74 (T1 image) and 4.15 (T2 image) for the salt and pepper noise; the SNRs is improved 2.31 (T1 image) and 0.30 (T2 image) for the speckle noise. It hints that the performance

[3]

[4]

[5]

[6]

[7]

[8]

X. P. Zhao, Principle, equipment and applications of magnetic resonance imaging, Science Press, China, 2000. Kazubek, M, Wavelet domain image denoising by thresholding and Wiener filtering, IEEE Signal Processing Letters, vol. 10, no. 11, pp. 324326, 2003. M. N. Do, and M. Vetterli, The contourlet transform: an efficient directional multiresolution image representation, IEEE Transactions on Image Processing, vol. 14, no. 12, pp. 20912106, 2005. H. Fang, C. Y. Zheng, Z. K. Yin, et al., Method for image denoising based on improved contourlet transform, Computer Engineering and Applications, vol. 46, no. 17, pp. 197199, 2010. A. L. Cunha, J. Zhou, and M. N. Do, he nonsubsampled contourlet transform: Theory, design, and applications, IEEE Transactions on Image Processing, vol. 15, no. 10, pp. 30893101, 2006. R. Eslami and H. Radha, Translation-invariant contourlet transform and its application to image denoising, IEEE Transaction on Image Processing, vol. 15, no. 11, pp. 33623377, 2006. D. D. Po, and M. N. Do, Directional multiscale modeling of images using the contourlet transform, IEEE Transactions on Image Processing, vol. 15, no. 6, pp. 16101620, 2006. W. Dai, S. L. Yu, and S. Sun, Image de-noising algorithm using adaptive threshold based on contourlet transform, Acta Electronica Sinica, vol. 35, no. 10, pp. 19391943, 2007.