Anda di halaman 1dari 4

Noise removal of MRI data with Edge Enhanceing

Xu Yan, Min-Xiong Zhou, Ling Xu, Wei Liu, Guang Yang


Shanghai Key Laboratory of Magnetic Resonance, Institute of Cognitive Neuroscience, East China Normal University, Shanghai 200062, China gyang@phy.ecnu.edu.cn
AbstractMagnetic Resonance Imaging (MRI) data always suffer from noise, especially for fast scanning sequence, such as EPI sequence which is basis for Functional MRI. Thus smoothing or denoising methods are always applied to suppress noise, but unavoidably blurs fine anatomical structure. While detail enhancement of denoised images is desirable, it often comes with kinds of artifacts. In this paper, a new strategy was proposed to combine image denoising and Multi-scale contrast enhancing techniques. The procedure extracted a continuous and noiseresistant edge from original image first, and then used it to combining the denoised image and High Frequency Components from Laplacian Pyramid technique. The algorithm was tested with both synthetic and real data, the experimental results showed that it smoothed image in homogeneous regions while preserved or enhanced details and edges of anatomical structure without introducing evident artifacts.) Keywords- Magnetic Resonance Imaging; Noise Removal; Image denoising; Non-Local Means; Image enhancement

denoising result in smoothing area. However, after using advanced denoising method, such as NLM, losing signal is highly mixed with noise. Using the same strategy will add noise back as well as edges. Besides, the edge detectors are usually influenced by noise, resulting in artifacts during the combining process. In this paper, we extended the combination strategy from compensation to detail enhancing, which merged NLM denoising and Multi-scale contrast enhancement method together. Thus, the sharpness of image edges are preserved or enhanced. Moreover, a noise-resistant edge detector based on Susan operator was proposed to guarantee this merging result. II. METHODS

I.

INTRODUCTION

Magnetic Resonance Imaging (MRI) has been successfully applied in medicine and psychology research during the last decade. Functional magnetic resonance imaging (FMRI), for example, uses EPI based sequence which scans a whole brain in less than 2 second, and makes study of cognitive process become possible. However, the cost is image quality. FMRI applications frequently suffer from noise and artifacts. . In this case, image postprocessing provides a convenient way to improve the image quality without introducing extra cost on hardware or imaging time. Image denoising has been a hot topic in image processing for decades. A great many types of algorithms in either image space or transform space have been proposed, such as Gaussian smoothing, Wavelet and Partial Differential Equation (PDE) based method. Recently, neighborhood based methods, such as bilateral filter [1], trilateral filter [2], and Non-Local Means filter (NLM) [3], become increasingly popular. NLM method, in particular, has demonstrated outstanding denoising power over a wide range of images and has been applied to both natural images[3] and MRI images [4-6]. However, it bases on weighted spatial averaging, which may meanwhile lose the sharpness of anatomical structure. In paper [7], G. Placidi al. proposed an image restoration strategy that makes use of edge information to compensate the blurred edge from denoising method with original image. Compared with the simple denoising algorithm, it effectively preserves sharp edges of the original image while maintains the

A. NLM Denoising Non-Local Means filter gets its name from the way it calculates the filtered pixel[3, 4]. Each filtered pixel is essentially a weighted average of all pixels in the image. The weights used in the calculation reflect the similarity between the patterns of neighboring pixels of the current pixel and those of the contributing pixels. It shows great noise removal power, see Fig. 1. Suppose for image I, u(x) is the gray value of pixel x, then the similarity between the neighborhoods of pixel x and y can be expressed by the weighted Gaussian Euclidian distance of all the pixels in each neighborhood:
d ( x , y ) = G || u ( R x ) u ( R y ) ||2

(1)

where Rx and Ry refer to all pixels in the neighborhood of the size (1+2f)2, centered at x and y respectively, and G is a normalized Gaussian weighting function to penalize pixels far from the center of the neighborhood. As mentioned earlier, NLM filtered pixel is a weighted average of all pixels in the image:
NLM(u ( x )) = 1 d ( x, y ) exp( )u ( y ) Z ( x ) yI h2

(2)

where h is a parameter for exponential decay control and Z(x)

Fig. 1. Non-Local Means (NLM) filter. From left to right: original Phantom, noisy one (5% rician noise), NLM denoised one and residual map.

This work was supported by Science and Technology Commission of Shanghai Municipality (08DZ1900700) and National Natural Science Foundation of China (10575038).

978-1-4244-5089-3/11/$26.00 2011 IEEE

is used for normalization:


Z ( x) =

y I

exp(

d ( x, y ) ) h2

(3)

SUSAN operator [11]. Since MRI images often come with a relatively low SNR, we choose SUSAN operator for its good noise resistance, see Fig 3. To further removal noise artifacts of edge extracted from Susan operator, we again use the excellent noise resistant property of NLM. It can be achieved by averaging the edge map with the NLM weights calculated from original MRI images (Fig. 3).For computational efficiency, this step can be combined with NLM denoising, since the averaging weights are exactly the same for smoothing MRI images shown in (2).

Considering computational burden, the calculation of weighted average is often limited in a local region Sx of size (1+2t)2, which is called searching region. NLM compares the similarities between neighborhoods and uses pattern redundancy in the image to reduce the noise. It can be also extends the use of pattern redundancy to the entire image without space distance penalty (thus non-local).

B. Multi-scale Enhancement In general, image enhancement technique can be divided into two categories, global contrast enhancement and local (or detail) enhancement. Global contrast enhancements, such as linear contrast stretching, histogram equalization, and other histogram equalization based algorithms, operate on the whole image, whereas local contrast enhancement is intended to selectively enhance edges and details in the image. Multiresolution methods, such as algorithms based on fast wavelet transformed (FWT) [8] or Laplacian Pyramid (LP) [9] have been used successfully in local contrast enhancement. It is reported that visible artifacts are often present in images enhanced by FWT based algorithm [8], thus an LP based approach is used for detail enhancement in this work. Typical LP decomposition method [9, 10] down samples each layer by a factor of two, which are not suitable for directly enhancing. We up sampled each layers to the same size as original image, see Fig 2. And then the High Frequency Components (HFC), namely each layer of the pyramid, can be calculated as follow,
HFC ij =

Fig. 3. Edge detector. From left to right: Noisy image, Binary Sobel edge, Susan edge and Susan edge with smoothing.

D. Combining Denoising and Enhancement As pointed out in [7], wavelet based methods may over smooth image details and edges, so Placidi et al suggested to replace the denoised image with original one in regions rich of details or edges. And Binary Sobel operator was used for edge extraction. However, such compensation strategy only works when residual image (the difference between original image and denoised image) still contains considerable signal. It has been reported in [3] that with NLM denoising, almost no noticeable anatomical structures can be found in the image residuals or method noise, see Fig 1. Thus, such strategy is not applicable to NLM. In contrast, HFC provided by multi-scale enhancing method contains more significant structure information, and has potential to recover or enhancing the lost sharpness of image edge, see Fig. 2. Moreover, a robust edge detector plays an important role in this combination. The performance of Binary Sobel edge operator used in [7] was not robust against noise and its binary property may introduce discontinuity near the edge when the denoising procedure changes greatly the appearance of the original image. In their following works [12], Canny operator was used instead to detecting more continuous edge. But the resulting edge map were obviously too thin (width = 1 pixel) to cover most regions with valuable anatomical structures. Thus, we modified Susan operator to acquire more noiseresistant and continuous edge map. Its value lies in the range of [0, 1], where 1 indicates a 100% edge pixel. Furthermore, a parameter can also be introduced to control the crispness of edges to be enhanced. Finally, our proposed algorithm can be formulized as:
OUT = C i Edge i HFC + NLM

LPnij
n =1

(4)

where LPn denotes the nth layer of LP, N is the total number of layers used in enhancing. As layer increases, more features of lower frequency will be involved, which might result in undesirable change of image impression. Thus we conservatively choose N = 3 in this paper.

Fig. 2. Laplacian pyramid. From left to right: first, second and third layers of Laplacian Pyramid and its High Frequency Components.

C. Edge Detection Many existing edge detectors are available, including first order gradient methods (Sobel, Robert and Prewitt operator), second order differentiate methods (Canny operator), and

(5)

where OUT, NLM and HFC denote the final image, denoised image by NLM and high frequency components derived using Multi-scale method respectively. A large allows only small

brain data was corrupted with 1~7% ( = 41, 82, 123, 164, 205, 246, 287) Rician noise, which is originated from independent Gaussian noise in real and imaginary channels. Normally it can be simulated by
IN = ( I 0 + n1 ( )) 2 + n 2 ( ) 2

(6)

where I0, IN are true signal and noisy one, n is Gaussian noise with zero means and standard deviation . And the Peak signal-to-noise ratio (PSNR) was used to measure the differences between the resulted image and ground truth, which are defined for 8-bit (0~255) images as follow:
PSNR = 20 log 10 255 RMSE

(7)

where RMSE is the root mean square error between ground truth and result image. Since enhancement will decrease PSNR which is unfair for our method, we also tested the mere noise removal ability of our strategy. To do so, the enhancing component (HFC) was replaced by residual map of NLM denoising. Fig. 4 shows the results of synthetic T1 images with 5% Rician noise (SNR = 13, 14) processed by original Placidis and our proposed method. The result of Placidis method suffered from dot like noises in both edge area and smooth area. This may be due to insufficient noise suppressing ability and discontinuity of binary Sobel operator. While in our proposed method, no obvious artifacts were found. A PSNR comparison between NLM, our proposed method with or without enhancement and the original Placidis method (corr_d = 4, haar wavelet) was given in Fig. 5. Our pure noise removal method gained highest PSNR at low noise level, then as noise increases, it converged with NLM and outperform the others. It implied that our modified edge detector successfully found rich detail region and had less impacts on vast homogeneous region. Due to strong edge enhancement, our enhancing method showed low PSNR at low noise level. And

Fig 4. Method comparison. From left to right, the upper row shows synthetic T1 weighted image and noisy one (5% noise). Lower row are results of Placidis method our proposed one ( = 0.75, C = 0.6).

amount of sharp and significant edges to be enhanced and thus can be used in images with relatively lower SNR to avoid the impact of noise. On the other hand, a small can be used in high quality images. Parameter C controls the extent to which the edges are enhanced. Because both edge and HFC can be calculated in advance, this algorithm can also be implemented in a way that either C or can be adjusted interactively by the end user to achieve personalized enhancement. III. EXPERIMENT AND RESULTS

MRI images used in this section are synthetic T1 weighted MRI images (217*181*181, slice thickness = 1mm) from the BrainWeb Simulated Brain Database [13] and real images acquired on an OPM35I 0.35T MRI system from Shanghai Colorful Magnetic Resonance Technology Co. Ltd (range = [0, 4095], resolution = 256x256, scan matrix 256x192). In this section, proposed algorithm is compared with that of Placidi's. For the parameters for NLM denoising, exhausting tests have been done in previous works [3, 4, 14], giving recommended settings of t = 5 (11x11 search window), f = 2 (5x5 neighborhood), h = 1.2, where is the noise standard deviation evaluated from the background of the squared magnitude image. Here C was conservatively set to 0.6 to avoid over enhance artifacts, and was set to 0.75. There is no absolute criterion of image enhancement, so their evaluations are mostly done by their visual appearances. But detail restoration ability between Placidis and our strategy can still be compared quantitatively. The original synthetic

Fig. 5 Performance of methods for different level of noise.

Fig. 6 Performance of methods for real MRI data. From left to right, original T2 weighted image, NLM denoised one and our method with enhancement.

then as noise increases, less fine structure could be found for enhancing, resulting in similar PSNR as NLM method. For real MRI data, our proposed method kept the outstanding denoising result of NLM while give a subtle enhancement of the details, see Fig. 6. IV.
CONCLUSION

[5]

[6]

In this paper, a new strategy was proposed to combine MRI image denoising and multi-scale detail enhancement technique. And to achieve robust merging result, a new modified noiseresistant Susan edge detector was also proposed. Our method can effectively reduce the noise in image while keeping or enhancing the image details without introducing undesirable artifacts. Moreover the algorithm can also be used to enable the end user to enhance the image interactively to achieve more personalized enhancement while the image is denoised. REFERENCES
[1] C. Tomasi and R. Manduchi, Bilateral filtering for gray and color images, in Book Bilateral filtering for gray and color images, Series Bilateral filtering for gray and color images, Editor ed.^eds., City, 1998, pp. 839-846. [2] W. Wong, A. Chung, and S. Yu, Trilateral filtering for biomedical images, in Book Trilateral filtering for biomedical images, Series Trilateral filtering for biomedical images, Editor ed.^eds., City, 2004, pp. 820-823. [3] A. Buades, B. Coll, and J.M. Morel, A Review of Image Denoising Algorithms, with a New One, Multiscale Model. Simul., vol. 4, (no. 2), pp. 490-530, 2005. [4] Jos V. Manjn, Jos Carbonell-Caballero, Juan J. Lull, Gracin GarcaMart, Lus Mart-Bonmat, and M. Robles, MRI denoising using Non-

[7]

[8]

[9] [10] [11] [12] [13] [14]

Local Means, Medical Image Analysis, vol. 12, (no. 4), pp. 514-523, 2008. M. Descoteaux, N. Wiest-Daessle, S. Prima, C. Barillot, and R. Deriche, Impact of Rician adapted Non-Local Means filtering on HARDI, Med Image Comput Comput Assist Interv Int Conf Med Image Comput Comput Assist Interv, vol. 11, (no. Pt 2), pp. 122-30, 2008. N. Wiest-Daessle, S. Prima, P. Coupe, S.P. Morrissey, and C. Barillot, Rician noise removal by non-Local Means filtering for low signal-tonoise ratio MRI: applications to DT-MRI, Med Image Comput Comput Assist Interv Int Conf Med Image Comput Comput Assist Interv, vol. 11, (no. Pt 2), pp. 171-9, 2008. G. Placidi, M. Alecci, and A. Sotgiu, Post-processing noise removal algorithm for magnetic resonance imaging based on edge detection and wavelet analysis, Phys Med Biol, vol. 48, (no. 13), pp. 1987-95, Jul 7 2003. S. Dippel, M. Stahl, R. Wiemker, and T. Blaffert, Multiscale contrast enhancement for radiographies: Laplacian Pyramid versus fast wavelet transform, IEEE Trans Med Imaging, vol. 21, (no. 4), pp. 343-53, Apr 2002. Peter J. Burt and E.H. Adelson, The Laplacian Pyramid as a Compact Image Code, IEEE Transactions on Communications, vol. COMM31, (no. NO. 4), pp. 532-540, 1983. M. Do and M. Vetterli, Framing pyramids, IEEE Transactions on Signal Processing, vol. 51, (no. 9), pp. 2329-2342, 2003. S. Smith and J. Brady, SUSANA new approach to low level image processing, International Journal of Computer Vision, vol. 23, (no. 1), pp. 45-78, 1997. I. Delakis, O. Hammad, and R.I. Kitney, Wavelet-based de-noising algorithm for images acquired with parallel magnetic resonance imaging (MRI), Phys Med Biol, vol. 52, (no. 13), pp. 3741-51, Jul 7 2007. Available: http://www.bic.mni.mcgill.ca/brainweb. P. Coupe, P. Yger, S. Prima, P. Hellier, C. Kervrann, and C. Barillot, An optimized blockwise nonlocal means denoising filter for 3-D magnetic resonance images, IEEE Trans Med Imaging, vol. 27, (no. 4), pp. 425-41, Apr 2008.

Anda mungkin juga menyukai