2009 IEEE International Conference on Signal and Image Processing Applications

Enhanced Anisotropic Diffusion with Noise

Amplification Suppression BaIza Achmad #1, Mohd. Marzuki Mustafa #2, Aini Hussain #3

# Department of Electrical, Electronic and Systems Engineering, Universiti Kebangsaan Malaysia 43600 UKM Bangi, Selangor Darul Ehsan, Malaysia

1 balzach@ugm. ac. id 2 marzuki@vlsi.eng. ukm.my

3 aini@vlsi.eng. ukm.my

Abstract- In order for medical doctors to be able to effectively utilize ultrasound images to support their diagnosis, the images require to be enhanced. The noise contained in the image needs to be reduced and the edges need to be sharpened.

In this paper, an enhanced technique based on anisotropic diffusion that is capable of carrying out simultaneously image smoothing and enhancement is presented. The technique (EAD) is equipped with noise amplification suppression to prevent

unwanted enhancement of noise. The technique performs well for image containing noise up to 30%. The tuning parameter is simpler compared with other anisotropic diffusion enhancements.

I. INTRODUCTION

Up to recently, ultrasonography is one of the most popular and widely used imaging technique utilized to support diagnostic of a patient by a medical doctor. This technique utilizes ultrasound waves that go through subcutaneous tissue of the patient. The information gathered from the reflected ultrasound waves is reconstructed to visualize body structures including tendons, muscles, joints, vessels and other internal organs.

Normally, ultrasonography machine gives grayscale images which reflect the internal condition of the patient. Unfortunately, these images suffered from speckle noise unavoidably created by the interaction between the emanated ultrasound waves. This drawback has to be overcome in order to obtain good quality images proper for diagnostic. Many researchers have developed various image processing methods to eliminate speckle noise in ultrasound images by utilizing linear as well as non-linear filters, including median-based, wavelet-based [1], artificial intelligence-based [2], and PDE­based filters [3, 4, 5, 6].

Elimination of speckle noise, however, is not the only important task required to improve the quality of ultrasound image. The image needs also to be enhanced to help the doctors differentiate between body structures and determine the organs shape. Therefore, proper image processing method has to be able to simultaneously carry out noise reduction and edge enhancement [4]. Some of the abovementioned methods have already performed these two basic image processing tasks [2, 4, 5, 6]. However this paper focuses on a filter family

which has been drawing interests, for the last decade i.e. anisotropic diffusion (AD).

Gilboa et al developed a forward and backward diffusion technique for adaptive image enhancement and denoising, which needs to optimize four coefficients to balance the forward and backward forces [5]. Li and Meng introduced a simpler conductance function that performs forward-and­backward diffusion [6]. However, the main purpose of the technique was to enhance the contrast of the image and it was merely tested on low contrast images. If the technique is applied on an image that contain significant amount of speckle noise, the noise will also be amplified and hence degrades the quality of the image. This technique has two tuning parameters; thus there is still a chance to simplify the conductance function by reducing the number of tuning parameters.

In this paper, we develop a technique based on AD with noise amplification suppression, which we call Enhanced Anisotropic Diffusion (EAD).

II. PROPOSED TECHNIQUE

AD borrows a mechanism of heat conduction from mechanical engineering discipline in which heat diffuses through a medium from a location with higher temperature to another location with lower temperature [5]. If the conduction coefficient is uniform all over the medium, this heat will be conducted isotropically. Analogously in image processing realm; noise, considered as hotspots, in an image will be dispersed to the neighboring pixels. Isotropic diffusion is equivalent to a Gaussian convolution, which naturally has blurring effect on original the image. In order to prevent the edges from being blurred by the diffusion process, the

conduction coefficients were varied as a function of intensity gradient of each pixel.

Perona and Malik [3] proposed two different conduction coefficient function as shown in (1) and (2).

(1)

978-1-4244-5561-4/09/$26.00 ©2009 237

(2)

where dI is the intensity gradient of a pixel and K is a tuning parameter.

These functions always give positive conduction coefficients, as shown in Fig. 1, which in turn result in smoothing effect to the image. Although it has been minimized, the smoothing of the edges that have higher intensity gradient still exists for some extent. In order to enhance an edge instead of smoothing it, a function is introduced as follow.

As shown in Fig. 1, this function produces positive values of conduction coefficient at lower intensity gradient but negative values at higher gradient. The negative conduction coefficient leads to enhancement process since it reverses the diffusion flow. Therefore, using the function given in (3), the smoothing of homogenous area and enhancement of edges can be performed simultaneously. Here, K sets the gradient threshold between the smoothing and the enhancement process.

0,8

0,6

0,4

""0 0,2 U

0

-0,2

-0,4

-0,6 EAD

200 250

Fig. 1. Conduction coefficient as a function of intensity gradient for Perona­Malik and EAD

Some researchers proposed conduction coefficient functions similar to the proposed function in this paper. Gilboa et al developed forward-and-backward anisotropic diffusion (FAB) [5]. However, their technique needs to optimize four tuning parameters, i.e. edge threshold for forward force as well as backward force, and the ratio

between the two forces. Li and Meng also used similar conduction function, which had the parameters that control the forward and backward diffusion. As mentioned before, these two parameters have to be optimized to obtain best result. In contrast, the proposed method contains only single parameter (K), hence make it simpler to optimize.

The negative conduction coefficient has a risk that can cause amplification of unwanted noise [5]. Therefore, the proposed technique is designed such that it will detect the presence of noise explosion and differentiate it from the edges. Then, if noise explosion is found, the conduction coefficient for this pixel is set to 1. This will force the noise to be diffused to the neighbouring pixels and hence prevent it from being amplified. The conduction coefficient of the EAD with noise amplification suppression (EADNAS) can thus be formulated as follow.

(dJ) -

{c EAD (dJ), if edge (4) C EADNAS

-

1, if noise explosion

Noise explosion is determined by examining the intensity gradients of the neighbouring pixels. If the gradients of the observed pixel have the same sign for all directions with amplitude larger than gradient threshold (K), then the pixel is considered as noise.

III. RESULTS AND DISCUSSION

The technique was tested using noise free image in order to be able to control the amount of noise that will be added to the image. The test image is given in Fig 2. From this image, spackle noise were added in various amount, i.e. 5%, 10%, 15%, 20%, 25%, and 30%.

• • Fig. 2. Test image

Firstly, the EAD technique was applied to an image containing low noise percentage (5%) and then compared to the original AD using PMl and PM2 conduction coefficient functions. Fig. 3 shows the resulted images of both techniques (EAD and AD). The figure also present the grayscale profiles at row 11 0 to show the performance of the technique in smoothing the homogenous area while enhancing the edges. The result shows that both techniques managed to smooth the homogenous area while preserving the edges. However, the blurring effects appeared on the images which resulted from the AD technique. On the other hand, EAD was able to enhance the edges.

238

� r--------

r l � r---'

(c) result of AD with PM2 (d) result ofEAD

Fig. 3. Images and grayscale profiles at row 110 resulted from 5% noisy image

Next, EAD was applied and tested on images containing higher quantity of noise, without and with noise amplification suppression feature. EAD without noise amplification suppression failed to eliminate the noise completely, as can be seen in Fig. 4. In fact, some amount of noise was amplified hence worsening the quality of the original image. The severity level of noise amplification was higher as the amount of noise increases. It can be seen that for images containing relatively large amount of noise, EAD does not work properly without the noise amplification suppression feature.

(a) image with 10% noise (b) result of EAD

(f) result of EAD

Fig. 4. Images and grayscale profiles at row 110 resulted from 10%, 15% and 20% noisy images using EAD without noise amplification suppression

When the noise amplification suppression is utilized, EAD was able to perform proper noise suppression while enhancing the edges. Fig. 5 presents the EAD result when noise amplification suppression is in used onto an image containing speckle noise up to 30%. It can be seen that noise amplification could be avoided. However, for images with higher noise percentage, the smoothing was not as good as the one for with lower noise percentage.

(a) from 5% noisy image (b) from 10% noisy image

Fig. 5. Images and grayscale profiles at row 110 resulted from 10% up to 30% noisy images using EAD with noise amplification suppression (continues)

239

(c) from 15% noisy image (d) from 20% noisy image

(e) from 25% noisy image (t) from 30% noisy image

Fig. 5. Images and grayscale profiles at row llO resulted from lO% up to 30% noisy images using EAD with noise amplification suppression (continued)

Lastly, EAD was applied to an ultrasound image and compared to the original AD. The ultrasound image was an image of a kidney generated using Field II ultrasound imaging simulation developed by the Department of Information Technology, Technical University of Denmark [7]. Again, both AD and EAD techniques were able to suppress the speckle noise, as shown by Fig. 6. However, the edges of the image resulted from EAD were sharper then those of AD.

(a) kidney ultrasound image (b) result of AD with PM 1

(c) result of AD with PM2 (d) result of EAD

Fig. 6. Results for kidney ultrasound image

IV. CONCLUSION

An enhanced anisotropic diffusion technique featured with noise amplification suppression has been developed and tested. The technique is able to simultaneously suppress noise and enhance edges. The optimization of the process is simpler compare to other AD techniques since there is only one parameter needs to be tuned.

ACKNOWLEDGMENT

The authors would like to express gratitude to the University Kebangsaan Malaysia for the support to this research through research grant contract number UKM-GUP­TKP-08-24-080.

REFERENCES

[I] V.P.S. Rallabandi, "Enhancement of Ultrasound Images using Stochastic Resonance-Based Wavelet Transform", Computerized Medical Imaging and Graphics, vol. 32, p.p. 316-320, 2008.

[2] E.d.S. Filho, M. Yoshizawa, T. Iwamoto, A. Tanaka, Y. Saijo, "Morphological Fuzzy Filter for Enhancement of Intravascular Ultrasound Images", in Proc. SICE Annual Conference in Sapporo, Hokkaido Institute of Technology, August 4-6, 2004.

[3] P. Perona and 1. Malik, "Scale-Space and Edge Detection Using Anisotropic Diffusion", IEEE Transl. Pattern Analysis and Machine Intelligence, vol. 12, no. 7, pp. 629-639, July 1990.

[4] 1. Monteil and A. Beghdadi, "A new interpretation and improvement of the nonlinear anisotropic diffusion for image enhancement", IEEE Transl. Pattern Analysis and Machine Intelligence, vol. 21, no. 9, pp. 940-646, Sept. 1999

[5] G. Gilboa, N. Sochen, and Y. Y. Zeevi, "Forward-and-backward diffusion processes for adaptive image enhancement and denoising", IEEE Trans. on Image Processing, vol. 11, no. 7, pp. 689-703, July 2002.

[6] B. Li, M. Q. H. Meng, "Wireless capsule endoscopy images enhancement using contrast driven forward and backward anisotropic diffusion", in Proc. International Conference on Image Processing (ICIP), 2007, pp. II 437-440.

[7] 1.A. Jensen and P. Munk, "Computer Phantoms for Simulating Ultrasound B-Mode and CFM Images", Acoustical Imaging, vol. 23, pp. 75-80, Eds.: S Lees and L. A. Ferrari, Plenum Press, 1997.

240