BEMD Based Ultrasound Image Speckle Reduction Technique Using Pixel-Wise Wiener Filtering

DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 19 |NUMBER: 2 |2021 |JUNE BEMD Based Ultrasound Image Speckle Reduction Technique Using Pixel-Wise Wiener Filtering Bhawna GUPTA , Vineet KHANDELWAL Department of Electronics and Communication Engineering, Jaypee Institute of Information Technology, Sector-62, 201309 Noida, Uttar Pradesh, India bha[email protected], vineet.khandelw[email protected] DOI: 10.15598/aeee.19i2.4100 Article history: Received Jan 30, 2021; Revised Mar 17, 2021; Accepted Apr 01, 2021; Published Jun 30, 2021. This is an open access article under the BY-CC license. Abstract. In this paper, an improved Bidimensional Empirical Mode Decomposition (BEMD) based speckle reduction technique for ultrasound images has been proposed. The noisy image has been decomposed into its Intrinsic Mode Functions (IMFs) and a residue. The noise component of the low order IMFs is removed with the pixel-wise Wiener filtering. The image is reconstructed with these filtered low order IMFs, high order IMFs and the residue. The performance of the proposed method has been tested on synthetic as well as real ultrasound images having noise components of different variance. The experimental results show that the proposed algorithm performs better than other existing methods for synthetic images as well as real ultrasound images in terms of various image quality matrices. Keywords BEMD, noise reduction, speckle, ultrasound, Wiener. 1. Introduction Speckle, which is a multiplicative noise, is an unwanted phenomenon that is present in ultrasound images due to scattering at the time of acquiring the image [1]. Speckle reduction in ultrasound images is an essential step and is targeting improvement in the quality of the image in terms of PSNR (Peak Signal to Noise Ratio), CC (Correlation Coefficient), SNR (Signal to Noise Ratio), FoM (Figure of Merit) and SSIM (Structural SIMilarity) [2]. Due to the high-frequency characteristic of the noise component, the main challenge of speckle reduction technique is that while removing noise the information in the edges should not be lost as this is important for diagnosis. The metric used for measuring the same is the EKI (Edge Keeping Index). Image denoising specifically speckle reduction in ultrasound images has been studied extensively, and it has been broadly classified into different domains, such as spatial domain techniques, transform domain techniques, and hybrid techniques as shown in Fig. 1. The spatial domain techniques work directly on the image pixels while transform domain technique applies an appropriate transform to convert the image to frequency domain before processing. Furthermore, there are hybrid techniques that are combination of spatial domain and transform domain methods. The spatial domain techniques use local statistics or information redundancy between similar patches and replace the pixel value by processing the nearby pixel values. Most successful amongst this category are diffusion-based filters like Speckle Reducing Anisotropic Diffusion (SRAD), Detail Preserving Anisotropic Diffusion (DPAD), Perona-Malik’s Anisotropic Diffusion (PMAD) [3], [4], [5] and [6], Bilateral filters [7] and [8], and patch-based methods like Non Local Mean filter (NLM) [9], [10], [11] and [12] and Optimized Bayesian Nonlocal Mean filter (OBNLM) [13] and [14]. The similar patch-based methods need to select the candidate patch reasonably so that the removal of noise does not lead to lose of edge information. Therefore, recent work as proposed in [15], [16] and [17] uses modified NLM and BM3D algorithms while trying to preserve the edge information. Wiener filters [18] and [19] are the optimum linear filters, that are widely in use for image processing. Performance of the classical pointwise Wiener filter is enhanced for noise 168 ©2021 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 19 |NUMBER: 2 |2021 |JUNE Fig. 1: Classification of Speckle reduction techniques. corrupted images, with the use of non-local parameter estimation proposed by Andre et al. [20]. Transform domain methods assume that the image can be sparsely represented by its low-frequency components and the speckle is present in the highfrequency components of the image [21]. Thereby, the image high-frequency components are dealt with by applying the thresholding (hard or soft) as in wavelet thresholding method to remove the noise [22]. While these methods give satisfactory results, choosing the right threshold is tricky and should avoid the Gibbs phenomenon and over smoothening effects. Thus, further extension in the transform domain techniques includes data-driven techniques and modeling techniques. Hybrid techniques are the ones that combine spatial domain methods with transform domain methods for the improvement of denoising performance. Some of the recent hybrid techniques are as introduced in [23], [24] and [25]. Empirical Mode Decomposition (EMD) was introduced by Huang et. al [26] in 1998. This is a very useful algorithm that decomposes the signal in its Intrinsic Mode Functions (IMFs) [27]. This decomposition is done based on their local frequency or oscillation in the spatial domain. Unlike Fourier transform and Wavelet transforms, the basis functions calculated are signal-dependent and are used to estimate a series of IMFs via an iterative procedure called sifting [28]. EMD was introduced in images in 2003 [29] and the Bidimensional Empirical Mode Decomposition (BEMD) for images was introduced in 2005 in [30]. BEMD is also a signal-dependent adaptive technique decomposing the image in a series of IMFs and a residue. The low-order IMFs are the high-frequency components and the high order IMFs are low-frequency components. BEMD is an adaptive multi-resolution analysis technique that is driven by the input signal. Therefore, it is used for various applications in image processing [31]. As the noise in image is mainly concentrated in the high-frequency components, the low order IMFs are having more noise components as compared to the high order IMFs. Accordingly, some BEMD based denoising algorithms use this fact to suppress the noise existing in low order IMFs. But this may not always be true and a significant noise component may also be present in further IMFs as well. Some denoising algorithms deal with this issue and use mutual information [32] or other filtering techniques [33]. These techniques use mutual information or a similarity measure between the probability density functions of the input signal and IMFs, to determine the noise dominant low order IMFs [28], [34] and [35]. These noise-dominant IMFs are discarded and the signal dominant IMFs are retained to obtain the denoised signal. However, most of the highfrequency components of the image also contain detail information such as edges. These algorithms therefore lose important edge details although giving good denoising performance. These details are of importance for diagnostic purposes when dealing with ultrasound images in particular. This paper introduces an improved technique, in which the low-order high-frequency IMFs, which are noise dominant, are not completely discarded. Rather pixel-based selective filtering in form of Wiener filter is applied to the low order IMFs, to reduce the noise component, while preserving the edge details. The remainder of the paper is organized as follows. Section 2. , provides a brief overview of BEMD algorithm along with parameter metrics used for comparison. In Sec. 3. , the proposed denoising method is described. The performance evaluation of the proposed method is illustrated in Sec. 4. , and Sec. 5. presents the conclusions. ©2021 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 169 DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 19 |NUMBER: 2 |2021 |JUNE 2. Background 2.1. Bidimensional Empirical Mode Decomposition EMD in signals was introduced by Huang et. al [26], which is an adaptive technique, which does not have predefined basis functions and decomposes the signal into various IMFs. Similar to EMD in one dimension, BEMD is an adaptive technique being applied to the images [31]. This decomposes the signal into IMFs and a residue. For the sake of illustration, BEMD algorithm is presented here briefly. Let f(m, n)be the given image. Let rl(m, n)represent the residue of l-th IMF. To find the next IMF, the residue of the previous IMF is taken as the input. Let il,k(m, n)be the input image for the generation of an IMF, where the first index is l-th number IMF, l= 1, . . . , L, the second index is k-th iteration of the sifting process, k= 1, . . . , K and (m, n)being two spatial dimensions. Step 1: Start with the given image as the input signal. i1,0(m, n) = f(m, n). Step 2: Extract all local maxima and minima of il,k(m, n). Step 3: Interpolate all local maxima to get the upper envelope eu(m, n)and interpolate all local minima to get the lower envelope el(m, n). The applicationspecific spline can be used for interpolation. Step 4: Calculate the envelope mean el,k(m, n)of the upper and the lower envelopes obtained in step 3: el,k(m, n) = eu(m, n) + el(m, n) 2.(1) Step 5: The input signal is updated by subtracting the envelope mean el,k(m, n)for the next iteration: il,k(m, n) = il,k−1(m, n)−el,k(m, n), k →k+1.(2) Step 6: This step is to check if the result obtained in step 5 is an IMF or not. For this standard deviation, is calculated as: = M−1 X m=0 N−1 X n=0 |il,k(m, n)−il,k−1(m, n)|2 i2 l,k−1(m, n).(3) Step 7: Check the standard deviation to be less than a predefined value (generally 0.2–0.3). If the value is greater than the criterion, repeat steps 2–6. When the value of is below the predefined value, the result of step 5 is the required l-th IMF, fl(m, n): fl(m, n) = il,k(m, n).(4) Step 8: The residue of the l-th IMF is defined as: rl(m, n) = il,0(m, n)−fl(m, n).(5) Step 9: The next IMF is calculated by taking the residue calculated as the input signal and starting over from step 2: il+1,0(m, n) = rl(m, n).(6) All subsequent IMFs are calculated by repeating the steps from 2–9. The process is stopped when the residue calculated has no more extrema. So, with the total Lnumber of IMFs calculated and the last residue without extrema rL, the original signal can be represented as: f(m, n) = L X l=1 fl(m, n) + rL(m, n).(7) It is important to mention here that the low order IMFs, are corresponding to the high frequency while the high order IMFs are corresponding to the low frequency. 2.2. Image Quality Performance Metrics The performance of the proposed BEMD based noise reduction technique using a Wiener filter is analyzed and compared both quantitatively and qualitatively with the existing techniques. For synthetic US images generated using Field II software [36] developed by J. A. Jensen, a quantitative analysis is carried out using performance metrics viz., Peak Signal to Noise Ratio (PSNR) [37], Edge Keeping Index (EKI) [38], Structure SIMilarity Index Measures (SSIM) [39], Correlation Coefficient (CC) [40], Signal to Noise Ratio (SNR) [37] and Figure of Merit (FoM) [41]. However, for real ultrasound images, as no noise-free image is available, the Mean to Variance Ratio (MVR) [42] and an Equivalent Number of Looks (ENL) [43] are used for quantitative evaluation. For a noisy image f(m, n)and the reconstructed denoised image fr(m, n), the metrics used for quantitative evaluation are defined as follows: •Peak Signal to Noise Ratio (PSNR) is a frequently used measure to access the efficacy of the denoising algorithm. It is the ratio of peak signal power to the noise power given by Eq. (8), where l= 255 for an 8-bit grayscale image of size M×N. •Edge Keeping Index (EKI) is a parameter to measure the edge keeping capability of a denoising algorithm, see Eq. (9), where ∆fand ∆frare high pass filtered version of fand fr, respectively, obtained using 3×3Laplacian operator and ∆fand 170 ©2021 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 19 |NUMBER: 2 |2021 |JUNE PSNR = 10 log10 l2 1 MN M−1 P m=0 N−1 P n=0 (f(m, n)−fr(m, n))2 ,(8) EKI = M−1 P m=0 N−1 P n=0 ∆f(m, n)−∆f∆fr(m, n)−∆fr sN−1 P n=0 ∆f(m, n)−∆f2∆fr(m, n)−∆fr2 ,(9) CC = M−1 P m=0 N−1 P n=0 f(m, n)−ffr(m, n)−fr sM−1 P m=0 N−1 P n=0 f(m, n)−f2M−1 P m=0 N−1 P n=0 fr(m, n)−fr2 .(12) ∆frare the mean values of ∆fand ∆fr, respectively. •Structural SIMilarity (SSIM) is for measuring the structure saving capability of the denoising algorithm. If fand frare the mean and σiand σir are the standard deviation of original and reconstructed images respectively, then: SSIM = 2f(m, n)fr(m, n) + c1(2σf,fr+c2) f2+fr 2+c1σ2 i+σ2 f,fr+c2, (10) where c1= (0.01l)2,c2= (0.03l)2and covariance image matrix is given by: σ2 f,fr=1 N−1 N−1 X k=0 fk−ffrk −fr.(11) •Correlation Coefficient (CC) defines the interdependence of the noisy image and the reconstructed image. It is defined as shown in Eq. (12). •Signal to Noise Ratio (SNR) is defined as the ratio of signal power to noise power: SNR = 10 log10 M−1 P m=0 N−1 P n=0 (f(m, n))2 sM−1 P m=0 N−1 P n=0 (f(m, n)−fr(m, n))2 . (13) •Figure of Merit (FoM) defined by Eq. (14) is a measure of preserving the edge information while denoising the image: FoM = 1 max (ND, NI)X n=1 ND1 1 + γd2 n,(14) where NDand NIare the numbers of edge pixels that are detected and ideally present respectively; dnis the Euclidean distance between the n-th detected edge pixel and the closest ideal pixel that is present; γis a scalar usually equal to 1 9for image calculations. •Mean to Variance Ratio (MVR) is calculated for real images for a selected local region as: MVR = µl σ2 l ,(15) where µland σ2 lare the local mean and variance. •Equivalent Number of Looks (ENL) is another parameter used for real images. It is the ratio of the square of mean to the variance for a selected local region given by: ENL = µ2 l σ2 l .(16) 3. BEMD Based Pixel-Wise Wiener Filtering In this section, the proposed BEMD based speckle reduction in Ultrasound images using Wiener filtering has been introduced. Due to the high-frequency characteristics of the speckle noise, it is mainly constrained in the low order IMFs of the ultrasound image. Therefore, considering the low order IMFs for noise reduction is the choice that has been considered. At the same time, the edge information is also a piece of crucial information from the point of view of the diagnostic importance in ultrasound images. As the edges are signified by the abrupt changes in the amplitude in the spatial domain, the low order IMFs cannot be discarded all together. The proposed scheme utilizes the fact that low order IMFs has maximum noise ©2021 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 171 DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 19 |NUMBER: 2 |2021 |JUNE component and thus low pass filters these IMFs using an adaptive Wiener filter for preserving the edge information present in these low order IMFs. Fig. 2: Block Diagram of the proposed scheme. The steps in the proposed algorithm are as follows: Step 1: Calculate the BEMD of the ultrasound image corrupted with speckle noise. Step 2: Select the low order IMFs which have the high-frequency components and are therefore having both the noise and the edge information. Step 3: Calculate the pixel-wise Wiener filtering of the selected IMFs [44] assuming additive noise v(n1, n2)with zero mean and variance σ2 v. Wiener filter estimates the local mean and variance around each pixel of the chosen IMF: µe=1 NM X m,n∈w fl(m, n),(17) σ2 e=1 NM X m,n∈w f2 l(m, n)−µ2 e,(18) where wis the window corresponding to the N×M neighborhood of each pixel in the IMF. This filter then creates a pixel-wise estimate given as: fe(m, n) = µe+σ2 e−σ2 v σ2 e (fl(m, n)−µe).(19) Step 4: The low-order filtered IMFs combined with the high-order IMFs and the residue are used to reconstruct the denoised image. 4. Experimental Results The kidney and fetus Field II images have been simulated and used for the experiments performed on MATLAB. The speckle noise has been added to the image with σ2= 0.1,σ2= 0.2, and σ2= 0.3. Figure 3 shows the BEMD IMFs for synthetic kidney image and synthetic fetus image, that are obtained for noise variance σ2= 0.1. The spline interpolation that is used in our case is a cubic spline. As can be seen in Fig. 3, the low order IMFs are having only the highfrequency components corresponding to speckle and edges. The Wiener filter is best defined for additive noise. Therefore, the multiplicative speckle noise has been log-transformed and converted to additive noise before applying BEMD. (a) (b) Fig. 3: BEMD IMFs for (a) Syntetic Kidney Field II image and (b) Synthetic Fetus Field II image. Pixel-wise Wiener filtering has been applied on low order IMFs. It has been observed that the results were best for the Wiener filter applied to IMF1 and IMF2. While applying Wiener filtering, 8-neighborhood has been utilized for calculation of local region. The Wiener filtered IMFs along with higher-order IMFs and the residue is reconstructed to get the de172 ©2021 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 19 |NUMBER: 2 |2021 |JUNE Tab. 1: Correlation Coefficient (CC) for various techniques. Technique CC Noise variance σ2 Synthetic fetus image Synthetic kidney image 0.1 0.2 0.3 0.1 0.2 0.3 Proposed EMD 0.96234 0.94607 0.92693 0.94284 0.90968 0.87969 Conventional EMD 0.95102 0.92555 0.87480 0.92645 0.87332 0.81060 Bilateral 0.91368 0.84291 0.78464 0.84947 0.74776 0.67000 SRAD 0.94191 0.87916 0.81739 0.91540 0.8292 0.75551 NLM 0.93126 0.87093 0.82044 0.87618 0.78484 0.71328 OBNLM 0.95155 0.90697 0.86630 0.90923 0.83734 0.77571 PMAD 0.94638 0.88481 0.82668 0.90595 0.80801 0.72333 Tab. 2: Signal to Noise Ratio (SNR) for various techniques. Technique SNR Noise variance σ2 Synthetic fetus image Synthetic kidney image 0.1 0.2 0.3 0.1 0.2 0.3 Proposed EMD 14.0620 12.6740 11.4590 15.6020 13.6940 12.4430 Conventional EMD 13.9140 11.0110 10.2900 14.2440 12.0820 10.2030 Bilateral 10.7200 7.9790 6.4897 10.7320 8.0182 6.5425 SRAD 12.6000 9.2820 7.3429 13.7580 10.2860 8.4398 NLM 11.7240 8.8585 7.3089 11.7650 8.9261 7.3852 OBNLM 13.3050 10.3730 8.6871 13.3320 10.4920 8.8753 PMAD 12.9410 9.4890 7.5779 13.2020 9.6131 7.6633 Tab. 3: Figure of Merit (FoM) for various techniques. Technique FoM Noise variance σ2 Synthetic fetus image Synthetic kidney image 0.1 0.2 0.3 0.1 0.2 0.3 Proposed EMD 0.82919 0.86146 0.84317 0.86334 0.86188 0.82560 Conventional EMD 0.82761 0.84320 0.82133 0.86112 0.83281 0.81143 Bilateral 0.81285 0.74764 0.70383 0.77549 0.72620 0.69278 SRAD 0.86349 0.78487 0.73090 0.83815 0.75237 0.73401 NLM 0.87531 0.79219 0.75194 0.83151 0.78397 0.74317 OBNLM 0.89364 0.84390 0.80875 0.88295 0.78922 0.76806 PMAD 0.86438 0.79341 0.73724 0.85802 0.76742 0.71880 noised image. The qualitative assessment is shown in Fig. 4 for synthetic kidney and in Fig. 5 for synthetic fetus for the noise of variance σ2= 0.1. As can be seen, the denoised image is perceptually of the same quality as that of the original image. (a) (b) (c) Fig. 4: Synthetic kidney for σ2= 0.1: (a) original image, (b) noisy image and (c) denoised image. Table 1, Tab. 2, Tab. 3, Tab. 4, Tab. 5 and Tab. 6 shows the comparison of various parameters obtained for different values of σfor the two Field II synthetic images, fetus and kidney. As can be seen, the proposed (a) (b) (c) Fig. 5: Synthetic fetus for σ2= 0.1: (a) original image, (b) noisy image and (c) denoised image. algorithm performs better than the existing techniques in terms of PSNR, SNR, and CC, for noise variance σ2= 0.1. The results obtained for these parameters give similar results even at higher noise variances, σ2= 0.2and σ2= 0.3. Moreover, the result in the tables show that the proposed technique gives comparable results with the existing techniques in terms of EKI, SSIM, and FoM for ©2021 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 173 DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 19 |NUMBER: 2 |2021 |JUNE Tab. 4: Peak Signal to Noise Ratio (PSNR) for various techniques. Technique PSNR Noise variance σ2 Synthetic fetus image Synthetic kidney image 0.1 0.2 0.3 0.1 0.2 0.3 Proposed EMD 24.946 23.577 22.383 24.322 22.388 21.158 Conventional EMD 22.398 20.671 20.330 22.317 20.672 19.821 Bilateral 21.056 18.009 16.260 19.025 15.992 14.254 SRAD 23.058 19.463 17.260 22.192 18.459 16.398 NLM 22.336 19.310 17.594 20.288 17.239 15.518 OBNLM 24.002 21.030 19.306 21.971 19.054 17.350 PMAD 23.539 19.837 17.690 21.730 17.878 15.686 Tab. 5: Edge Keeping Index (EKI) for various techniques. Technique EKI Noise variance σ2 Synthetic fetus image Synthetic kidney image 0.1 0.2 0.3 0.1 0.2 0.3 Proposed EMD 0.64603 0.59358 0.53587 0.54885 0.50024 0.49630 Conventional EMD 0.61720 0.52081 0.50012 0.51320 0.48230 0.45321 Bilateral 0.59360 0.54006 0.52306 0.53703 0.49294 0.47574 SRAD 0.55134 0.49815 0.49498 0.49201 0.45619 0.43757 NLM 0.59768 0.54228 0.51775 0.53614 0.49024 0.47184 OBNLM 0.62562 0.54757 0.52198 0.55135 0.49023 0.46734 PMAD 0.55606 0.50568 0.49567 0.49833 0.46407 0.44813 Tab. 6: Structural SIMilarity (SSIM) for various techniques. Technique SSIM Noise variance σ2 Synthetic fetus image Synthetic kidney image 0.1 0.2 0.3 0.1 0.2 0.3 Proposed EMD 0.70575 0.66684 0.62823 0.61088 0.53961 0.49747 Conventional EMD 0.72345 0.65482 0.61291 0.61098 0.49308 0.45320 Bilateral 0.71808 0.61478 0.5552 0.53328 0.39544 0.32945 SRAD 0.76831 0.67000 0.59414 0.65411 0.50453 0.41941 NLM 0.74678 0.64120 0.58082 0.56396 0.42391 0.35323 OBNLM 0.78688 0.68623 0.62356 0.62079 0.47589 0.39955 PMAD 0.78014 0.68009 0.60847 0.63845 0.47608 0.38675 σ2= 0.1. At the same time, based on the analysis of the values obtained by the experiments, the scheme is giving better results in terms of these parameters for higher values of σ2= 0.2and σ2= 0.3. Thus, we can summarize that the technique performs acceptably well even at higher noise levels. For completeness of the efficacy evaluation of the proposed algorithm, an experiment has also been performed on the real ultrasound image database taken from [45]. The real images have three sets of data namely kidney, liver, and gall bladder images each having around 85 images. Three regions were selected randomly for all three sets and the MVR and ENL have been calculated. Figure 6(a) shows the real liver ultrasound image and the three regions that are taken for calculations of MVR and ENL. Figure 6(b) and Fig. 6(c) are the results that are obtained for the selected regions. Table 7 shows the MVR and ENL for the real liver ultrasound database. Similar experiments were also performed on real ultrasound image data sets of kidney and gall bladder, however, due to paucity of space we have shown MVR and ENL plots for liver database only. The results obtained for the other sets are also in conjunction with the results shown here. Tab. 7: MVR and ENL for Real Liver Ultrasound Database. Technique MVR ENL Proposed EMD 18.51 ±3.76 5.25 ±2.79 Conventional BEMD 17.91 ±5.32 5.01 ±2.54 Bilateral 15.42 ±5.16 3.96 ±2.32 SRAD 17.66 ±4.52 4.87 ±2.35 NLM 17.01 ±4.14 4.91 ±2.15 OBNLM 17.81 ±4.71 4.95 ±2.61 PMAD 16.39 ±6.21 4.33 ±2.79 174 ©2021 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 19 |NUMBER: 2 |2021 |JUNE (a) (b) (c) Fig. 6: (a) Real liver image with selected regions, (b) MVR plot, (c) ENL plot. 5. Conclusion In this paper, a BEMD based image denoising algorithm using pixel-wise Wiener filtering has been presented. The proposed method solves the problem of losing the edge information in the conventional BEMD based denoising technique. The conventional BEMD based algorithm is able to remove low order IMF, but loses important edge information which has been preserved with the help of the method proposed herein. The performance of the method has been verified quantitatively in terms of PSNR, EKI, SSIM, FoM, SNR, and CC for synthetic ultrasound images. The quantitative analysis is also done on the real ultrasound images in terms of MVR and ENL. It has been found that the proposed algorithm performs better than many other existing state-of-the-art techniques and can preserve the edge information. Author Contributions V.K. conceived the presented idea. B.G. developed the theory and performed the experiments. V.K. encouraged B.G. to investigate and supervised the findings of this work. Both authors discussed the results and contributed to the final manuscript. References [1] BURCKHARDT, C. B. Speckle in ultrasound B-mode scans. IEEE Transactions on Sonics and Ultrasonics. 1978, vol. 25, iss. 1, pp. 1–6. ISSN 2162-1403. DOI: 10.1109/T-SU.1978.30978. [2] SZABO, T. L. Diagnostic Ultrasound Imaging: Inside Out. 2nd ed. Amsterdam: Academic Press, 2014. ISBN 978-012-396487-8. [3] YU, Y. and S. T. ACTON. Speckle reducing anisotropic diffusion. IEEE Transactions on Image Processing. 2002, vol. 11, iss. 11, pp. 1260–1270. ISSN 1941-0042. DOI: 10.1109/TIP.2002.804276. [4] RAHIMI, M. and M. YAZDI. A new hybrid algorithm for speckle noise reduction of SAR images based on mean-median filter and SRAD method. In: 2015 2nd International Conference on Pattern Recognition and Image Analysis (IPRIA). Rasht: IEEE, 2015, pp. 1–6. ISBN 978-1-4799-8445-9. DOI: 10.1109/PRIA.2015.7161623. [5] LIU, X., J. LIU, X. XU, L. CHUN, J. TANG and Y. DENG. A robust detail preserving anisotropic diffusion for speckle reduction in ultrasound images. In: The 2010 International Conference on Bioinformatics and Computational Biology (BIOCOMP 2010): Genomics. Las Vegas: BMC Genomics, 2011, pp. 1–10. DOI: 10.1186/1471-216412-S5-S14. [6] PERONA, P. and J. MALIK. Scale-space and edge detection using anisotropic diffusion. IEEE Transactions on Pattern Analysis and Machine Intelligence. 1990, vol. 12, iss. 7, pp. 629–639. ISSN 1939-3539. DOI: 10.1109/34.56205. [7] TOMASI, C. and R. MANDUCHI. Bilateral filtering for gray and color images. In: Sixth International Conference on Computer Vision. Bombay: IEEE, 1998, pp. 839–846. ISBN 81-7319-2219. DOI: 10.1109/ICCV.1998.710815. [8] BALOCCO, S., C. GATTA, O. PUJOL, J. MAURI and P. RADEVA. SRBF: Speckle Reducing Bilateral Filtering. Ultrasound in Medicine &Biology. 2010, vol. 36, iss. 8, pp. 1353–1363. ISSN 1879-291X. DOI: 10.1016/j.ultrasmedbio.2010.05.007. [9] COUPE, P., P. HELLIER, C. KERVRANN and C. BARILLOT. Nonlocal means-based speckle filtering for ultrasound images. IEEE Transactions on Image Processing. 2009, ©2021 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING 175 DIGITAL IMAGE PROCESSING AND COMPUTER GRAPHICS VOLUME: 19 |NUMBER: 2 |2021 |JUNE vol. 18, iss. 10, pp. 2221–2229. ISSN 19410042. DOI: 10.1109/TIP.2009.2024064. [10] ZHAN, Y., M. DING, L. WU and X. ZHANG. Nonlocal means method using weight refining for despeckling of ultrasound images. Signal Processing. 2014, vol. 103, iss. 1, pp. 201–213. ISSN 18727557. DOI: 10.1016/j.sigpro.2013.12.019. [11] LI, X., H. HE, R. WANG and J. CHENG. Superpixel-guided nonlocal means for image denoising and super-resolution. Signal Processing. 2016, vol. 124, iss. 1, pp. 173–183. ISSN 1872-7557. DOI: 10.1016/j.sigpro.2015.09.021. [12] GUO, Y., Y. WANG and T. HOU. Speckle filtering of ultrasonic images using a modified non local-based algorithm. Biomedical Signal Processing and Control. 2011, vol. 6, iss. 2, pp. 129–138. ISSN 1746-8108. DOI: 10.1016/j.bspc.2010.10.004. [13] KERVRANN, C., J. BOULANGER and P. COUPE. Bayesian Non-local Means Filter, Image Redundancy and Adaptive Dictionaries for Noise Removal. In: SSVM: International Conference on Scale Space and Variational Methods in Computer Vision. Ischia: Springer, 2007, pp. 520–532. ISBN 978-3-540-72823-8. DOI: 978-3-540-72823-8_45. [14] COUPE, P., P. HELLIER, C. KERVRANN and C. BARILLOT. Bayesian non local meansbased speckle filtering. In: 2008 5th IEEE International Symposium on Biomedical Imaging: From Nano to Macro. Paris: IEEE, 2008, pp. 1291–1294. ISBN 978-1-4244-2002-5. DOI: 10.1109/ISBI.2008.4541240. [15] YANG, J., J. FAN, D. AI, X. WANG, Y. ZHENG, S. TANG and Y. WANG. Local statistics and non-local mean filter for speckle noise reduction in medical ultrasound image. Neurocomputing. 2016, vol. 195, iss. 1, pp. 88–95. ISSN 1872-8286. DOI: 10.1016/j.neucom.2015.05.140. [16] MEI, F., D. ZHANG and Y. YANG. Improved non-local self-similarity measures for effective speckle noise reduction in ultrasound images. Computer Methods and Programs in Biomedicine. 2020, vol. 196, iss. 1, pp. 1–14. ISSN 1872-7565. DOI: 10.1016/j.cmpb.2020.105670. [17] HUANG, S., C. TANG, M. XU, Y. QIU, and Z. LEI. BM3D-based total variation algorithm for speckle removal with structurepreserving in OCT images. Applied Optics. 2019, vol. 58, iss. 23, pp. 6233–6243. ISSN 2155-3165. DOI: 10.1364/AO.58.006233. [18] JADWA, S. Wiener Filter based Medical Image De-noising. International Journal of Science and Engineering Applications. 2018, vol. 7, iss. 9, pp. 318–323. ISSN 2319-7560. DOI: 10.7753/IJSEA0709.1014. [19] KAY, S. M. Fundamentals of Statistical Processing, Volume I. 1st ed. Englewood Cliffs: Prentice Hall PTR, 1993. ISBN 978-0-1334-5711-7. [20] BINDILATTI, A. A., M. A. C. VIEIRA and N. D. A. MASCARENHAS. Poisson Wiener filtering with non-local weighted parameter estimation using stochastic distances. Signal Processing. 2018, vol. 144, iss. 1, pp. 68–76. ISSN 1872-7557. DOI: 10.1016/j.sigpro.2017.10.001. [21] JAIN, P. and V. TYAGI. A survey of edgepreserving image denoising methods. Information Systems Frontiers. 2016, vol. 18, iss. 1, pp. 159– 170. ISSN 1572-9419. DOI: 10.1007/s10796-0149527-0. [22] DONOHO, D. L. De-noising by soft-thresholding. IEEE Transactions on Information Theory. 1995, vol. 41, iss. 3, pp. 613–627. ISSN 1557-9654. DOI: 10.1109/18.382009. [23] CHOI, H. and J. JEONG. Speckle Noise Reduction Technique for SAR Images Using Statistical Characteristics of Speckle Noise and Discrete Wavelet Transform. Remote Sensing. 2019, vol. 11, iss. 10, pp. 1184–1210. ISSN 2072-4292. DOI: 10.3390/rs11101184. [24] CHOI, H. and J. JEONG. Despeckling Algorithm for Removing Speckle Noise from Ultrasound Images. Symmetry. 2020, vol. 12, iss. 6, pp. 938–963. ISSN 2073-8994. DOI: 10.3390/sym12060938. [25] CHEN, Y., M. ZHANG, H.-M. YAN, Y.-J. LI and K.-F. Yang. A New Ultrasound Speckle Reduction Algorithm Based on Superpixel Segmentation and Detail Compensation. Applied Sciences. 2019, vol. 9, iss. 8, pp. 1693–1705. ISSN 2076-3417. DOI: 10.3390/app9081693. [26] HUANG, N. E., Z. SHEN, S. R. LONG, M. C. WU, H. H. SHIH, Q. ZHENG, N.-C. YEN, C. C. TUNG and H. H. LIU. The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences. 1998, vol. 454, iss. 1971, pp. 903–995. ISSN 1471-2946. DOI: 10.1098/rspa.1998.0193. [27] LINDERHED, A. Image Empirical Mode Decomposition: a New Tool for Image Processing. Advances in Adaptive Data Analysis. 2009, vol. 1, iss. 2, pp. 265–294. ISSN 1793-7175. DOI: 10.1142/S1793536909000138. 176 ©2021 ADVANCES IN ELECTRICAL AND ELECTRONIC ENGINEERING