PERFORMANCE EVALUATION OF ADAPTIVE SPECKLE FILTERS FOR ULTRASOUND IMAGES Abstract: L.M.Merlin Livingston #, L.G.X.Agnel Livingston *, Dr. L.M.Jenila Livingston ** #Associate Professor, ECE Dept., Jeppiaar Institute Of Technology,Chennai, India * Asst. Professor, CSE Dept., St.Xavier s Catholic College of Engineering, Nagercoil, India **Associate Professor, SCSE & Learning Research Cell VIT Chennai. The quality of radar and ultrasound images is degraded by speckle noise. To eliminate the speckle noise, several methods are available based upon different mathematical models. In this study, three state-of-art techniques Lee filter; Frost filter and Speckle Reducing Anisotropic Diffusion (SRAD) are used to despeckling the medical ultrasound images. The Lee and Frost filters exploit the coefficient of variance in adaptive filtering by use of local statistics. In this context, SRAD utilizes the instantaneous coefficient of variation by use of local gradient magnitude and Laplacian operators. The performance of three adaptive speckle filters are evaluated by using ultrasound images affected with various noise level. Keywords: speckle noise, de-noising, ultrasound images, adaptive filters 1. Introduction Image de-noising refers to the recovery of a digital image that has been affected by noises. The digital image can be affected by different types of noises such as salt and pepper noise, Poisson noise, Gaussian noise and speckle noise. A novel multi parameter system to enhance the noisy image is presented in [1] which includes directional smoothening and sharpening. To estimate the noise standard deviation piecewise linear filtering technique is used. To reduce speckle noise in the ultrasound images the non local means filter is proposed in [2]. Pearson distance is introduced for patch comparison. This method is carried out on synthetic images by varying the values of speckle. Principal component analysis (PCA) with local pixel grouping (LPG) technique is proposed in [3] for de-noising the images. PCA is a statistical technique which is used to simplify the dataset by reducing it to lower dimensions and commonly used for data reduction in statistical pattern recognition and signal processing. The proposed algorithm has two stages. Most of the noises is removed in the first stage which gives an initial estimation of the image whereas the second stage will further refines the output of the first stage with a same procedure except noise level. A multilevel thresholding technique for noise removal in curvelet transform domain which uses cycle-spinning is proposed in [4]. By thresholding curvelet coefficient the both uncorrelated and correlated noise gets removed at the 62
lowest level but while comparing both the noises the correlated noise gets removed only by a fraction. In order to calculate the threshold value it depends on three major parameters mainly the contrast ratio, median and the level dependent parameter. A novel technique despeckling the ultrasound images using lossless compression is proposed in [5]. Generalized Laplacian distribution is modeled mainly for log transformed ultrasound images. In order to achieve de speckling and quantization a simple adaptation and reconstruction levels is proposed. The performance of the proposed method is evaluated using two stages. They are filtering and compressing the speckle images. This method gives as a better results based on visual quality and PSNR value. A novel nonlinear multi scale wavelet diffusion method for ultrasound speckle suppression and edge enhancement is presented in [6]. This method is designed to utilize the techniques, sparsity, multi resolution properties of the wavelet, and the iterative edge enhancement feature of nonlinear diffusion. Neural network technique proposed in [7] is used to remove the speckle noise. Speckle noise is modeled as Rayleigh distribution and the noise and the sigma is estimated using the neural network. Comparative study of various transforms like wavelet, curvelet, ridgelet and for various frequency domain filters for suppressing speckle noise is also explained in [8]. As a result Ridgelet transform gives a better performance of PSNR while comparing with other transform. A novel mathematical morphology algorithm is proposed in [9] to remove the speckle noise in the ultrasound images and also find the edges efficiently than the existing one. The results can be evaluated for various parameters like Signal to noise ratio (SNR), correlation, Structural similarity index (MSSIM), Root mean square error (RMSE) and Edge preserving index (EPI). An efficient and fast wavelet based technique for speckle removal from SAR images is proposed in [10]. The main novelty is the use of realistic distributions of the wavelet coefficients which represent mainly speckle noise on the one hand and those that represent the useful signal on the other. In this study, the performance of three adaptive filters Lee s filter, Frost filter and speckle reducing anisotropic diffusion are evaluated. 2. Adaptive Speckle Filters 2.1 Lee s Filter Lee filters [11] is used to smoothen the noisy data that have an intensity related to the image scene and that also have an additive and/or multiplicative component. Lee filtering is a standard deviation or sigma filter that filters data based on statistics calculated within individual filter windows. Unlike a typical low-pass smoothing filter, the Lee filter and other similar sigma filters preserve image sharpness and detail while suppressing noise. The pixel being filtered is replaced by a value calculated using the surrounding pixels. In order to eliminate speckle noise while preserving edges and point features in radar imagery the Lee filter is used. The filter achieves a balance between straightforward averaging (in homogeneous regions) and the identity filter (where edges and point features exist). This balance depends on the coefficient of variation inside the moving window. 2.2 Frost Filter 63
Frost filters [12] are used to reduce speckle while preserving edges in radar images. The Frost filter is an exponentially damped circularly symmetric filter that uses local statistics. The pixel being filtered is replaced with a value calculated based on the distance from the filter center, the damping factor, and the local variance. Frost filter also strikes a balance between averaging and the allpass filter. The Frost filter uses an exponentially damped convolution kernel that adapts to regions containing edges by exploiting local statistics. f ( x, y ) and f ( x, y ) considering one of image as a noisy approximation of the other follows: Mean Square Error (MSE) (1) The PSNR is defined as (2) 2.3 Speckle Reducing Anisotropic Diffusion Speckle reducing anisotropic diffusion (SRAD) is a method which is mainly applied to ultrasonic and radar images [13]. It is one of the edge sensitive methods for speckled images. SRAD not only preserves edges but also enhances edges by inhibiting diffusion across edges and allowing diffusion on either side of the edge. SRAD is adaptive and does not utilize hard thresholds to alter performance in homogeneous regions or in regions near edges and small features. 3. Performance Metrics To evaluate the performance of the three adaptive filters, the quantitative metrics Mean Square Error (MSE) and Peak Signal to Noise Ratio (PSNR) is used. The MSE is the ratio between the maximum possible power of a signal and the power of corrupting noise that affects the fidelity of its representation. The MSE is the most commonly used measure of quality of the reconstruction. It is defined as the mean squared error (MSE) for two images Here, peak to peak value of the referenced image is the maximum pixel value of the image. When the pixels are represented using 8 bits per sample, this is 255. The peak signal to noise ratio is calculated from the error using the above formula. The higher the value of the PSNR, the better is the performance of that particular local operator for the noise added. 4. Experimental Results In order to assess the performance of the three adaptive filters, experiments with 4 ultrasound images are used. Speckle noise with different noise level form 0.1 to 0.5is added to the original input image. Then the three filters Lee, Frost and SRAD are applied to remove the speckle noise. The MSE values for different methods are calculated and the comparative chart is shown Table 1. From the table, MSE value of SRAD filter is lower than Lee and Frost filters. The PSNR values are calculated by using the speckle filters are calculated and the comparative chart is shown Table 2. From the table, PSNR value of SRAD filter is higher than other filters. Figure 3 shows the corrupted image and 64
filtered output from the Lee, Frost and SRAD filters respectively. The noise variance of the first row corrupted image is 0.1, and the second row is 0.2 and so on. From the results, the Lee and Frost filtered images having more smooth area than the SRAD filtered output. Table 1 Comparison of MSE values of Lee, Frost and SRAD speckle filters for various noise levels Noise Level 0.1 0.2 0.3 0.4 0.5 Filter MSE Values Average Image 1 Image 2 Image 3 Image 4 Value Lee 164.69 183.85 265.23 181.42 198.80 Frost 138.02 154.35 200.68 152.67 161.43 SRAD 111.02 150.50 175.33 147.68 146.13 Lee 195.29 207.33 307.54 206.38 229.13 Frost 158.54 167.57 231.41 165.18 180.67 SRAD 123.80 144.14 195.29 142.19 151.35 Lee 209.25 231.51 341.42 226.24 252.10 Frost 169.37 188.73 264.29 183.52 201.47 SRAD 147.23 162.54 225.63 162.08 174.37 Lee 229.93 248.97 362.93 253.39 273.80 Frost 191.13 205.62 279.61 210.24 221.65 SRAD 178.10 192.13 251.24 188.47 202.48 Lee 253.35 275.37 408.14 268.34 301.30 Frost 216.30 237.04 331.67 229.49 253.62 SRAD 206.63 224.92 290.74 214.95 234.31 65
350 300 PSNR values 250 200 150 100 50 0 0.1 0.2 0.3 0.4 0.5 Frost SRAD Lee SRAD Frost Lee Noise Level Figure 1 Graphical representation of MSE values of speckle filters with varies noise levels Table 2 Comparison of PSNR values of Lee, Frost and SRAD speckle filters for various noise levels Noise Level 0.1 0.2 0.3 0.4 0.5 Filter PSNR Values Average Image 1 Image 2 Image 3 Image 4 Value Lee 25.96 25.49 23.89 25.54 25.22 Frost 26.73 26.25 25.11 26.29 26.09 SRAD 27.68 26.36 25.69 26.44 26.54 Lee 25.22 24.96 23.25 24.98 24.60 Frost 26.13 25.91 24.49 25.95 25.62 SRAD 27.20 26.54 25.22 26.60 26.39 Lee 24.92 24.49 22.80 24.59 24.20 Frost 25.84 25.37 23.91 25.49 25.15 SRAD 26.45 26.02 24.60 26.03 25.77 Lee 24.51 24.17 22.53 24.09 23.82 Frost 25.32 25.00 23.67 24.90 24.72 SRAD 25.62 25.29 24.13 25.38 25.10 Lee 24.09 23.73 22.02 23.84 23.42 Frost 24.78 24.38 22.92 24.52 24.15 SRAD 24.98 24.61 23.50 24.81 24.47 66
27 26 PSNR Values 25 24 23 Lee Frost 22 21 0.1 0.2 0.3 0.4 0.5 Lee SRAD Frost SRAD Noise Level Figure 2 Graphical representation of PSNR values of speckle filters with varies noise levels 67
(a) (b) (c) (d) Figure 3 (a) Corrupted image (b) Lee filtered image. (c) Frost filtered image. (d) SRAD filtered image 5. Conclusion In this study, the despeckling of medical ultrasound images by using three adaptive speckle filters is evaluated. The speckle filters used in this study are Lee s filter, Frost Filter and SRAD. To evaluate the performance, ultrasound images corrupted with varies noise density from 0.1 to 0.5 is used. The performance of the three filters is evaluated by MSE and PSNR metrics. Experimental results show that the SRAD produces better despeckling effect than the other filters. The PSNR value of SRAD filter is approximately +1.5 db higher than the Lee filter and also +0.5 db than Frost filter. References: [1] Fabrizio Russo, An Image-Enhancement System Based on Noise Estimation, IEEE Transactions on Instrumentation and Measurement, vol. 56, no. 4, August 2007. [2] Pierrick Coupé, Pierre Hellier, Charles Kervrann, and Christian Barillot, Nonlocal Means-Based Speckle Filtering for Ultrasound Images, IEEE transactions on Image Processing, vol. 18, no. 10, October 2009. [3] Y. Murali Mohan Babu, Dr. M.V. Subramanyam and Dr. M.N. Giri Prasad, PCA based image denoising, An International Journal on Signal & Image Processing, vol.3, no.2, April 2012. [4] Binh NT and Khare A, Multilevel threshold based image denoising in curvelet domain Journal of Computer Science and Technology, vol.25, no.3, May 2010. [5] Nikhil Gupta, M. N. S. Swamy and Eugene Plotkin, Despeckling of Medical Ultrasound Images Using Data and Rate Adaptive Lossy Compression, IEEE Transactions on Medical Imaging, vol. 24, no. 6, June 2005. [6] Yong Yue, Mihai M. Croitoru, Akhil Bidani, Joseph B. Zwischenberger and John W. Clark, Nonlinear Multiscale Wavelet Diffusion for Speckle Suppression and Edge Enhancement in Ultrasound Images, IEEE Transactions on Medical Imaging, vol. 23, no. 3, March 2006. [7] Mohammad R. N. Avanaki, P. Philippe Laissue, Adrian G. Podoleanu and Ali Hojjat, Denoising based on noise parameter estimation in speckled OCT images using neural network, Proceedings of SPIE, vol. 7139, 2008. 68
[8] Tajinder kaur, Manjit Sandhu and Preeti Goel, Performance Comparison of Transform Domain for Speckle Reduction in Ultrasound Image, International Journal of Engineering Research and Applications (IJERA), vol. 2, Issue 1, pp.184-188. [9] Arpit Singhal and Mandeep Singh, Speckle Noise Removal and Edge DetectionUsing Mathematical Morphology, International Journal of Soft Computing and Engineering (IJSCE), vol-1, Issue-5, November 2011. [10] Alexendra Pizurica, Wilfried Philips, Ignance Lemahieu and Marc Acheroy, Despeckling SAR Images Using Wavelets and a New Class of Adaptive Shrinkage Estimators, Proceedings of International Conference on Image Processing, vol.2,2001. [11] John Sen Lee, Digital Image enhancement and noise filtering by use of local statistics,ieee Transaction on Pattern analysis and Machine Intelligence, no.2,march.1980. [12] V. S. Frost, J. A. Stiles, K. S. Shanmugan, and J. C. Holtzman, A model for radar images and its application to adaptive digital filtering of multiplicative noise, IEEE Transactions on Pattern Analysis and Machine Intelligence., vol. 4 pp. 157 165, 1982. [13] Yongjian Yu and Scott T. Acton, Speckle Reducing Anisotropic Diffusion, IEEE Transactions on Image Processing, vol. 11, no. 11, November 2002. 69