INFORMATICAInformatica1822-88440868-49520868-4952Vilnius UniversityINFO121010.15388/Informatica.2019.194Research ArticleNon-Subsampled Shearlet Transform and Log-Transform Methods for Despeckling of Medical Ultrasound ImagesAbazariRezaabazari-r@uma.ac.ir1∗
R. Abazari received the MS degree in applied mathematics from University of Mohaghegh Ardabili, Ardabil, Iran, in 2008 and PhD degree in applied mathematics in 2017 from University of Tabriz, Tabriz, Iran. He is now an adjunct lecturer in applied mathematics at University of Mohaghegh Ardabili since September 2009. He is also an research associate at the Young Researchers and Elite Club, Ardabil, Iran. His research interests include mathematical method in physics, medical sciences and image analysis.
LakestaniMehrdad2
M. Lakestani received the BS degree in applied mathematics from University of Tabriz, Iran, in 1998 and the MSc and PhD degree both in applied mathematics from Amirkabir University of Technology – Tehran Polytechnic, Tehran, Iran, in 2000 and 2005, respectively. Since 2005, he is a professor at University of Tabriz. His research interests include numerical analysis, with special emphasis on wavelet theory, time-frequency analysis and image processing.
Medical Ultrasound is a diagnostic imaging technique based on the application of ultrasound in various branches of medical sciences. It can facilitate the observation of structures of internal body, such as tendons, muscles, vessels and internal organs such as male and female reproductive system. However, these images usually degrade by a special kind of multiplicative noise called speckle. The main effects of speckle noise in the ultrasound images appear in the edges and fine details which lead to reduce their resolution and consequently make difficulties in medical diagnosing. Therefore, reducing of speckle noise seriously plays an important role in image diagnosing. Among the various methods that have been proposed to reduce the speckle noise, there exists a class of approaches that firstly convert multiplicative speckle noise into additive noise via log-transform and secondly perform the despeckling process via a directional filter. Usually, the additive noises are mutually uncorrelated and obey a Gaussian distribution. On the other hand, non-subsampled shearlet transform (NSST), as a multi scale method, is one of the effective methods in image processing, specially, denoising. Since NSST is shift invariant, it diminishes the effect of pseudo-Gibbs phenomena in the denoising. In this paper, we describe a simple image despeckling algorithm which combines the log-transform as a pre-processing step with the non-subsampled shearlet transform for strong numerical and visual performance on a broad class of images. To illustrate the efficiency of the proposed approach, it is applied on a sample image and two real ultrasound images. Numerical results illustrate that the proposed approach can obtain better performance in term of peak signal to noise ratio (PSNR) and structural similarity (SSIM) index rather than existing state-of-the-art methods.
The invention of the computational tools, especially computers, in the recent decades has led to produce many new medical imaging techniques, such as computed tomography scan (CT scan), ultrasound imaging, magnetic resonance imaging (MRI) and etc., for medical diagnosing. Among these imaging techniques, the ultrasound imaging technique is popularly used in medical diagnosing, mainly in obstetrics, gastrointestinal and cardiovascular fields. The main reasons for the popularity of ultrasound imaging technique are related to other more sophisticated imaging techniques, such as CT, MRI or Positron Emission Tomography (PET), and are its low cost, non-ionizing properties and its portability to provide real time interactive visualization of anatomical structures (Abbott and Thurstone, 1979). Ultrasound images are created by ultrasonic waves, which are produced by dispatching special sound waves through body tissues and receiving their reflex by a transducer and they are processed and transformed into a digital image (Ragesh et al., 2011). Usually, the inappropriate contact or being air gap between the transducer and body lead to create an interference pattern called speckle (Abbott and Thurstone, 1979). Speckle is a particular type of noise that is characterized by a granular pattern of bright and dark spots which tend to degrade the fine details and edges of ultrasound images and consequently lead to complication in clinical diagnosing.
Generally, two families of approaches have been proposed for reducing of speckle noise of ultrasound images:
A) The first family are those filters that perform on ultrasound images directly. In Abazari and Lakestani (2018a) the authors applied the Fourier based discrete shearlet transform to despeckle medical ultrasound images. Ritenour et al. (1984) applied the median filter to suppress speckle noise from the digital radiographic images. The adaptive weighted form of median filter is also suggested by Loupas et al. (1989) to denoise ultrasonic images. The wavelet transform and its complex form are also applied directly for ultrasound images in Deka and Bora (2013), Khare et al. (2010). Elyasi and Pourmina (2016) have employed the TV regularization with modified bayes shrink for reducing of speckle noise from ultrasound images. In Zong et al. (1998) it is shown that the linear filtering cannot be an optimal method for reducing the speckle noises. In Elmoniem et al. (2002) a denoising method based on nonlinear coherent diffusion (NCD) is utilized. Yu and Acton (2002) proposed a despeckle method based on anisotropic diffusion method (SRAD) to cast the spatial adaptive filers into diffusion model (Yu and Acton, 2002). Oriented version of SRAD (OSRAD) filter (Krissian et al., 2007), which is one of the extension of SRAD, was proposed with appraising the properties of the numerical scheme associated with SRAD filter. In Vese and Osher (2003) modified the TV minimization algorithm to reduce speckle noise from ultrasound images. Zhang et al. (2001) proposed an algorithm based on wavelet frame for denoising of Doppler ultrasound signals. Wang et al. (2014) also introduced a new denoising approach based on framelet regularization. An automated approach for segmentation of intravascular ultrasound images is also studied in Vard et al. (2012).
B) The second family are those that firstly convert the speckle noise to an additive noise via log-transform method and then a special filtering is applied to denoise the additive noise. The fundamental properties of speckle noises (Goodman, 1976), the log-transform speckle noise is studied in Hiremath et al. (2013) and their properties in the Contourlet Transform Domain are clearly explained in Kabir and Bhuiyan (2015). The discrete wavelet denoising approaches via log-transform method (Rajeshwar, 2018) are utilized for reducing the speckle noises of medical ultrasound images. In Hazarika et al. (2015), the enhanced Lee filter in lapped orthogonal transform (LOT) domain is applied to despeckle the log-transformed SAR images. Some other approaches to reduce speckle noise in medical ultrasound images are also proposed in Huang et al. (2016), Gupta et al. (2004), Bhuiyan et al. (2009) and the references given there.
As mentioned above, several methods have been proposed for reducing speckle noises, however, each method has its assumptions, advantages, and disadvantages. Among them, the methods based on wavelet transform have good efficiency in noise reduction. However, wavelets fail to capture the geometric regularity along the singularities of curves, because of their isotropic support. To exploit the anisotropic regularity of a curve along edges, the basis must include elongated functions that are nearly parallel to the edges. Several image representations have been proposed to capture the geometric regularity of a given image. Some of these representations are ridgelet (Candes, 1998), brushlet (Meyer and Ronald, 1997), curvelet (Candes and Donoho, 2000), beamlet (Donoho and Huo, 2001), contourlet (Do and Vetterli, 2005) and recently proposed shearlet (Labate et al., 2005).
The sheartlet transform as an alternative anisotropic multi-resolution system has been introduced by Labate et al. (2005) which yields nearly optimal approximation properties (Guo and Labate, 2007). Furthermore, the definition of shearlet transform is such that includes various scales, location and orientation in order to optimally represent an image. This new representation is based on a simple and rigorous mathematical framework which not only provides a more flexible theoretical tool for the geometric representation of multidimensional data, but is more natural for implementation. As a result, the shearlet approach can be associated to a multiresolution analysis (MRA) and this leads to a unified treatment of both the continuous and discrete world (Labate et al., 2005). Also, unlike the wavelets, they are optimal in sparse representation of multi-dimensional data (Guo and Labate, 2012) and, unlike curvelets, their directionality is achieved by shear matrices instead of rotation matrices (Guo and Labate, 2013). The shearlet transform has been applied in diverse areas of engineering and medical sciences, including inverse problems (Colonna et al., 2010), image separation (Kutyniok and Lim, 2011), image restoration (Patel et al., 2009), image denoising (Lakestani et al., 2016) and medical image analysis (Abazari and Lakestani, 2018a). However, due to shift variant nature of the shearlet, this method produces artifacts in the most of their process more in image denoising and image fusion (Vishwakarma et al., 2018). Some improved methods proposed to rectify the mentioned artifacts. Recently, the author proposed a hybrid denoising method based on shearlet transform and yaroslavsky’s method (Abazari and Lakestani, 2018b) to suppress the effect of the pseudo-Gibbs phenomena and shearlet-like artifacts in denoising. The non-subsampled shearlet transform (NSST) is also proposed to capture edges and line discontinuities for image fusion (Vishwakarma et al., 2018). Since the NSST is shift invariant, it diminishes the effect of pseudo-Gibbs phenomena and shearlet-like artifacts in the related processing.
In this paper, by focusing on the non-subsampled form of “discrete shearlet transform” (Hou et al., 2012), we have proposed a despeckling approach via log-transform for speckle noisy ultrasound images. To illustrate the efficiency of the proposed approach, it is applied on a sample image and two real ultrasound images. Numerical results illustrate that the proposed approach can obtain better performance in terms of peak signal to noise ratio (PSNR) and structural similarity (SSIM) index rather than existing state-of-the-art methods.
Speckle Noise
Speckle noise is characterized by a peculiar granular pattern of bright and dark spots which lead to degrade the resolution of ultrasound images. This typical pattern is also observed in other kind of images involving coherent radiation, such as Laser and Synthetic Aperture Radar (SAR).
Generally, the speckle noise is described as a multiplicative phenomenon. Suppose that the observed image fn be degraded of noise free image f by the speckle noise ns and an additive noise (such as thermal noise) na as follows from Kabir and Bhuiyan (2015):
fn=fns+na.
Since the effect of additive noise in comparison to speckle noise is very little, so equation (1) can be written as
fn=fns.
By applying the log-transformation on (2), we obtain
Fn=F+Ns,
where Fn=log(fn),F=log(f) and Ns=log(ns). The probability density function (pdf) of ns is given by a Rayleigh pdf as follow:
pdf(ns)=nsα2e−ns22α2,
where α is the shape parameter and the expected value of ns will be E(ns)=απ2. From (3) and (4), it follows that
pdf(Ns)=eNsα2e−e2Ns2α2,
where pdf(Ns) is the pdf of Ns. By applying multi-resolution transform, such as shearlet transform, on (3), we obtain
y=x+ε,
where y, x and ε represents the coefficients corresponding to Fn,F and Ns, respectively. In Goodman (1976), Goodman studied some fundamental properties of speckle noise. Also, he shows that the statistics of log-transformed speckle noise is given by a double-exponential probability density function which is known as Fisher–Tippett probability density function. For more information about properties of speckle noise, please see Goodman (1976), Kabir and Bhuiyan (2015) and the references mentioned there.
Shearlet Transform
In this section, the shearlet and its transform both in continuous and discrete form will be briefly explained. The shearlet representation is a directional representation system that provides more geometrical information and shearlets are frame elements used in this representation scheme.
For any ψ∈L2(R2), the continuous shearlet system is defined as follows
SH(ψ)={ψa,s,t(x)=a−34ψ(Aa−1Ss−1(x−t)),a>0,s∈R,t∈R2},0,\hspace{2.5pt}s\in R,\hspace{2.5pt}t\in {R^{2}}\big\},\]]]>
where Aa=(a00a) is anisotropic dilation matrix as a mean to change the resolution and Ss=(1s01) is shear transformation matrix as a means to change the orientation.
The dilation matrix Aa resembles the parabolic scaling, which has an elongated history in the literature of harmonic analysis and can be outlined back to the second dyadic decomposition from the theory of oscillatory integrals. Briefly, from the dilation matrix Aa it can be concluded that the scaling in the x-direction is square of the scaling in the y-direction. The general form of dilation matrix Aa is Aa=diag(a,aα) with the parameter α∈(0,1) that controls the degree of anisotropy, however, the value α=12 plays a special role in the discrete setting. In fact, parabolic scaling is required in order to obtain optimally sparse approximations of cartoon-like images (see Definition 2), since it is best adapted to the C2-regularity of the curves of discontinuity in the cartoon-like images class (Guo and Labate, 2007). The shearing matrix Ss also parameterizes the orientations using the variables associated with the slopes rather than the angles, and has the advantage of leaving the integer lattice invariant, provided s is an integer. The geometric effects of parabolic scaling and shearing with fixed parameter a and several parameter s are illustrated in Fig. 1. The associated continuous shearlet transform of any f∈L2(R2) is given by
SHψf(a,s,t)=⟨f,ψa,s,t⟩.
In other words, SHψ maps the function f to the coefficients SHψf(a,s,t) associated with the scale variable a>00$]]>, the orientation variable s∈R, and the location variable t∈R2.
The geometric effects of parabolic scaling and shearing with fixed parameter a and several parameter s. (a) s=0, (b) s=14 and (c) s=12.
Now, our main aim is to achieve a continuous shearlet transform, which becomes an isometry, since this is automatically associated with a reconstruction formula. To do it, the generating function ψ must be a well localized function and be compatible with admissibility condition as follows from Labate et al. (2005),
∫R2|ψˆ(ξ1,ξ2)|2ξ12dξ2dξ1⩽∞.
So that, each f∈L2(R2) has the representation
f=∫Rn∫−∞∞∫0∞⟨f,ψa,s,t⟩ψa,s,tdaa3dsdt.
Consequently, it can be easily construct examples of shearlets, including examples of admissible shearlets which are well localized. Essentially any function ψ such that ψˆ is compactly supported away from the origin is an admissible shearlet. A particular example of these representation is classical shearlet, wherein for any ξ=(ξ1,ξ2)∈Rˆ2,ξ1≠0, the generating function ψ is setting such that
ψˆ(ξ1,ξ2)=ψˆ1(ξ1)ψˆ2(ξ2ξ1),
where ψ1∈L2(R) is a wavelet which satisfies in the Calderon condition (Guo and Labate, 2007), given by
∑j∈Z|ψˆ1(2−jξ)|2=1,∀a.e.ξ∈R
with suppψˆ1⊂[−2,−12]∪[12,2] and ψ2∈L2(R) is a bump function with suppψˆ2⊂[−1,1] and
∑k=−11|ψˆ2(ξ+k)|2=1,∀a. e.ξ∈[−1,1].
Thus, a classical shearlet ψ is a function which is wavelet-like along one axis and bump-like along another one. Each element ψa,s,t of classic shearlet has frequency support on a pair of trapezoids, at various scales a, symmetric with respect to the origin and oriented along a line of slope s (see Fig. 2). Let ψ∈L2(R2) be an admissible shearlet. Define
Cψ+=∫0∞∫R|ψˆ(ξ1,ξ2)|2ξ12dξ2dξ1,Cψ−=∫−∞0∫R|ψˆ(ξ1,ξ2)|2ξ12dξ2dξ1.
If Cψ+=Cψ−=1, then SHψ is an isometry.
Support of the classical shearlets ψˆa,s,t (in the frequency domain) for different values of a and s.
To obtain the discrete form of the continuous shearlet system and related transform, it is easy to discrete by properly sampling the scale, shear and translation parameters. A (regular) discrete shearlet system, associated with ψ∈L2(R2) and denoted by SH(ψ), is defined by
SH(ψ)={ψj,k,m=234jψ(SkA2j.−m),j,k∈Z,m∈Z2},
which can be easily obtained by setting (a,s,t)=(2−j,−k2−j/2,S−k2−j/2A2−jm). Similarly to the continuous case, the discrete shearlet transform of f∈L2(R2) is defined as the following map
f→SHψf(j,k,m)=⟨f,ψj,k,m⟩,
where j,k∈Z, m∈Z2. The main goal of utilizing shearlet systems is analysis and synthesis of 2-D data, therefore, we need to provide a discrete shearlet system SH(ψ) which forms a basis or, more generally, a frame. Similar on continuous form, for each f∈L2(R2), we have the reproducing formula:
f=∑j,k∈Z,m∈Z2⟨f,ψj,k,m⟩ψj,k,m,
with convergence in the L2 sense. Also, in Labate et al. (2005), Guo and Labate (2007) it is illustrated that the classical shearlet is a well-localized function, i.e. it has rapid decay both in the spatial and in frequency domain. The well localization property of classical shearlet implies that discrete shearlet system SH(ψ) forms a frame. This property is needed for obtaining optimally sparse approximations. Tiling of the frequency plane induced by discrete shearlets ψˆj,k,m is shown in Fig. 3.
Tiling of the frequency plane induced by discrete shearlets ψˆj,k,m. The horizontal region Dh is illustrated in solid line and the vertical region Dv is in dashed line.
The class E2(R2) of cartoon-like images is the set of functions f:R2→C of the form
f=f0+f11B,
where B⊂[0,1]2 is a set with ∂B being a closed C2-curve with bounded curvature, 1B(x)={1,x∈B0,x∉B is indicator function and fi∈C2(R2) are functions with supp fi∈[0,1]2 and ‖fi‖C2=1 for each i=0,1.
The property of optimally sparse approximations of multivariate functions is one of the main motivations to propose the shearlet framework. Before stating the main results, we briefly describe how shearlet expansions are able to achieve optimally sparse approximations. Consider a cartoon-like function f and let SHψ be a discrete shearlet system of (10). For j∈Z, the elements of SHψ are approximately inside a box of size 2−j/2×2−j, it follows that at scale 2−j there exists about O(2j/2) such waveforms whose support is tangent to the curve of discontinuity. Consequently, for j sufficiently large, each shearlet coefficient ⟨f,ψj,k,m⟩ can be controlled by Guo and Labate (2007, 2012)
|⟨f,ψj,k,m⟩|⩽‖f‖∞‖ψj,k,m‖L1⩽C2−3j/4,
where C is a constant. From inequality (12) and the observation that there exists at most O(2j/2) significant coefficients, it can be concluded that the Mth largest shearlet coefficient is bounded by O(M3/2). This implies that the following result holds.
Letf∈E2(R2)be a cartoon-like image defined on a bounded domainΩ⊂R2, and letfMbe the approximation of f obtained by taking the M largest coefficient|ψj,k,m|in the shearlets expansion of f given by (11). Then the asymptotic approximation error is given by‖f−fM‖2⩽CM−2(logM)3,M→∞.
To prove see Guo and Labate (2007, 2012) and the references therein. □
Let ψj,k,m be a classical shearlet, according to Fig. 2, to attain the reproducing formula (11), it is enough to compute the inner product of ⟨f,ψj,k,m⟩ in both horizontal region Dh and vertical region Dv. In Guo and Labate (2007, 2012), for both region index d={h,v}, it can be shown that
⟨f,ψj,k,md⟩=23j2∫R2fˆ(ξ)V(2−2jξ)Wj,kd(ξ)‾e2πiξAd−jSd−kmdξ,
where
V(2−2jξ)=ψˆ1(2−2jξ1)χDh(2−2jξ)+ψˆ1(2−2jξ2)χDv(2−2jξ),Wj,kh(ξ)=ψˆ2(2jξ2ξ1−k)χDh(ξ)+ψˆ2(2jξ1ξ2−k+1)χDv(ξ),ifk=−2jψˆ2(2jξ2ξ1−k)χDh(ξ)+ψˆ2(2jξ1ξ2−k−1)χDv(ξ),ifk=2j−1ψˆ2(2jξ2ξ1−k),otherwiseWj,kv(ξ)=ψˆ2(2jξ2ξ1−k+1)χDh(ξ)+ψˆ2(2jξ1ξ2−k)χDv(ξ),ifk=−2jψˆ2(2jξ2ξ1−k−1)χDh(ξ)+ψˆ2(2jξ1ξ2−k)χDv(ξ),ifk=2j−1ψˆ2(2jξ1ξ2−k),otherwise
and Ah=4002,Sh=1101 and Av=2004, Sv=ShT. In (13), it is required to compute fˆ(ξ) in discrete form, so given an N×N image f∈ℓ2(ZN2), the 2D discrete Fourier transform (DFT) of f will be:
fˆ[k1,k2]=1N∑n1,n2=0N−1f[n1,n2]e−2πi(n1Nk1+n2Nk2),
where −N2⩽k1,k2⩽N2 and brackets [.,.] denote the indices of fˆ in Fourier domain. To compute the integrand of (13), in the space domain and at the resolution level j, firstly, the Laplacian-pyramid (LP) algorithm (Burt and Adelson, 1983) associated with the pseudo-polar Fourier transform (Averbuch et al., 2008) will be utilized. This will achieve the multiscale partition illustrated in Fig. 4, by decomposing faj−1[n1,n2], 0⩽n1,n2⩽Nj−1, into a low pass filtered image faj[n1,n2], a quarter of the size of faj−1[n1,n2], and a high pass filtered image fdj[n1,n2]. Consequently, to compute the shearlet coefficients ⟨f,ψj,k,md⟩ given by (13), in the discrete domain, it is enough to compute the inverse pseudo-polar DFT and apply the inverse two-dimensional fast Fourier transform (FFT) on each decomposition level.
The figure illustrates the succession of Laplacian-pyramid and directional filtering.
We refer to Labate et al. (2005), Guo and Labate (2007, 2012) for additional information about shearlet and its applications in various sciences and engineering.
Non-Subsampled Shearlet Transform (NSST)
The main idea of shearlet transform is to filter signals in pseudo-polar grid (Averbuch et al., 2008), and then utilize a bandpass filter in frequency domain directly without sampling operations. Therefore, the directional filtering is kept away from distortion and lead to invariance in shearlet transform. Also, as described in the last section, the discrete form of shearlet transform is achieved by combination of Laplacian-pyramid (Burt and Adelson, 1983) algorithm and directional filter. To improve the computational efficiency and also to reduce the effect of Gibbs phenomena, usually the directional filter is so designed that has a small size support.
The non-subsampled form of Laplacian-pyramid filter (NSLP) is proposed by Cunha et al. (2006). By substituting NSLP for LP and combining it with discrete shearlet transform, non-subsampled shearlet transform (NSST) is designed to improve effectiveness of discrete shearlet transform (Hou et al., 2012). The NSST is known as the shift-invariant version of the shearlet transform. Since the NSST is a fully shift-invariant, multi-scale and multi-directional expansion in comparing to shearlet transform, it can diminish the effect of pseudo-Gibbs phenomena and shearlet-like artifacts in the related processing. The analysis of NSLP can be done as the following iterative processing (Hou et al., 2012):
NSLPj+1f=(Fhj1∏k=1j−1Fhk0)f,
where f is an image, NSLPj+1 is the detail coefficients at scale j+1, and Fhk0 and Fhj1 are low pass and high pass filters of NSLP at scale j and k, respectively. Therefore, according to Fig. 4, given an N×N image f∈ℓ2(ZN2), the procedure of the non-subsampled shearlet transform associated with non-subsampled Laplacian-pyramid at a fixed resolution scale j and in the number of direction Dj can be summarized in Algorithm 1.
Non-subsampled shearlet transform
Proposed Algorithm and Experimental Results
In this section, firstly, we have proposed a denoising algorithm based on non-subsampled shearlet transform associated with log-transform method and secondly evaluated its performance for reducing the speckle noise of the ultrasound images. The structure of present method is similar to those described in Abazari and Lakestani (2018a), Hou et al. (2012). Consider the speckle noisy problem u=vns, where u is observed image, v is noise free image and ns is the multiplicative speckle noise which is independent of noise free image v. According to relationships (3)–(6), after applying the log-transform, we have
g=f+Υ,
where g=log(u), f=log(v) and Υ=log(ns). The additive noise Υ has properties similar to additive Gaussian noise. Our goal is to obtain an estimatation of f, namely f˜, from the noisy data g by applying a classical soft thresholding scheme (Labate et al., 2005; Guo and Labate, 2007) on the shearlet coefficients of g. The threshold levels are given by τj,k=cjσΥj,k, as in Labate et al. (2005), Guo and Labate (2007, 2012, 2013), where σΥj,k is the standard deviation of noise at scale j and shear directional band k and cj is the scaling parameter. Here the standard deviation σΥj,k is estimated by using the MATLAB function of std. By using Laplacian-pyramid decomposition, we used five levels of the NSLP decomposition, and we applied a directional decomposition on four of the five scales. According to Fig. 4, we used eight shear filters of size 32×32 for the first two scales (coarser scales), and sixteen shear filters of size 16×16 for the third and forth levels (fine scales) and so on. Finally, by using the exp-transformation, the estimated image can be obtained as f˜=exp(f˜). Let f be noise free image of size N×N and f˜ denotes the estimated image, to test our algorithm and to assess its performance, we used two measurements from the following:
PSNR: as peak signal-to-noise ratio, measured in decibels (dB), defined by
PSNR(f,f˜)=10log10255N‖f−f˜‖F,
where ‖.‖F is the Frobenius norm.
SSIM: for calculating the structural similarity (SSIM) index between denoised image f˜ and original image f, defined by Wang et al. (2004),
SSIM(f,f˜)=(2μfμf˜+c1)(2σff˜+c2)(μf2+μf˜2+c1)(σf2+σf˜2+c2),
where μf,μf˜ and σf2,σf˜2 are the average and variance of f,f˜, respectively, σff˜ is the covariance of f and f˜ and c1=(k1L)2, c2=(k2L)2 are two variables to stabilize the division with weak denominator. Here, we set L=255 as dynamic range of the pixel values and k1=0.01, k2=0.03.
The proposed scheme, which is briefly mentioned in Algorithm 2, is implemented using MATLAB 2012. The scours of ShearLab (2008) are used to construct a discrete form of shearlet transform and its non-subsampled form is erected by open source files of NSCT (2008) proposed by Minh. N. Do which initially applied for contourlet transform.
Proposed method
Step 1:
Input a speckled noisy ultrasound image u,
Step 2:
Take the logarithmic transform on the speckled noisy image u to obtain f.
Step 3:
Computes the discrete shearlet transform of f by applying eight shear filters of size 32×32 for the first two scales (coarser scales), and sixteen shear filters of size 16×16 for the third and forth levels (fine scales) and so on, associated with 5 scale, i.e. j=1,2,…,5, see Algorithm 1.
Step 4:
Compute the standard deviation σΥj,k in each scale j and shear direction k.
Step 5:
Applying thresholding τj,k=cjσΥj,k in the scale j and shear direction k.
Step 6:
Apply the inverse discrete shearlet transform to obtain image f˜.
Step 7:
Apply the exp-transform, f˜:=exp(f˜), to obtain estimate image f˜.
Sample test images. (a) Cartoon-like image. (b) Real ultrasound image of a tumour. (c) Real ultrasound image of a fetus.
Different methods of noise removal on a synthetic image. (a) Original image. (b) Image polluted by multiplicative speckle noise with Rayleigh distribution (variance is υ=0.8) (PSNR = 18.41, SSIM = 0.2372). (c) The denoising result using SRAD (Yu and Acton, 2002) (PSNR = 24.35, SSIM = 0.8551). (d) The result using VO method (Vese and Osher, 2003) (PSNR = 23.04, SSIM = 0.8044). (e) The denoised result using framelet regularization method without backward diffusion (Wang et al., 2014) (PSNR = 23.95, SSIM = 0.8164). (f) The denoised result using framelet regularization method and backward diffusion (Wang et al., 2014) (PSNR = 24.23, SSIM = 0.8314). (g) The result of proposed method (PSNR = 27.24, SSIM = 0.9258).
Visual comparison of various speckle suppressing methods on a ultrasound image of a tumour. (a) Original image. (b) The denoising result using SRAD (Yu and Acton, 2002). (c) The result using VO method (Vese and Osher, 2003). (d) The denoised result using framelet regularization method without backward diffusion (Wang et al., 2014). (e) The denoised result using framelet regularization method and backward diffusion (Wang et al., 2014). (f) The result of the proposed method.
Visual comparison of various speckle suppressing methods on a ultrasound image of a fetus. (a) Original image. (b) The denoising result using SRAD (Yu and Acton, 2002). (c) The result using VO method (Vese and Osher, 2003). (d) The denoised result using framelet regularization method without backward diffusion (Wang et al., 2014). (e) The denoised result using framelet regularization method and backward diffusion (Wang et al., 2014). (f) The result of the proposed method.
Visual comparison of various speckle suppressing methods on two real ultrasound images. (a) and (b) Real ultrasound images. (c) and (d) The result of the proposed method.
In the first set, we have considered three images listed in Fig. 5 (one sample cartoon-like image and two real ultrasound images same as considered in Yu and Acton (2002), Vese and Osher (2003), Wang et al. (2014), and have compared the results of the proposed method with some related medical ultrasound despeckling techniques from the recent literature such as Yu and Acton (2002), Vese and Osher (2003), Wang et al. (2014). Firstly, we degraded a noise free sample cartoon-like image, (Fig. 5(a)), by multiplicative speckle noise with Rayleigh distribution by variance of υ=0.8 and then applied the proposed method on this image. To evaluate the efficiency of the proposed method, we used peak signal to noise ratio (PSNR) as a quantity to measure the quality of reconstruction of noisy images and SSIM as a structural similarity index between denoised image and original image. The despeckled image of our proposed method, (Fig. 6(g)), is compared with related techniques of Yu and Acton (2002), Vese and Osher (2003), Wang et al. (2014) with respect to PSNR and SSIM and are shown in Figs. 6(b)–6(f). The measurements of PSNR and SSIM of the proposed method in comparison to those methods mentioned in Yu and Acton (2002), Vese and Osher (2003), Wang et al. (2014) show that the proposed approach is better than those methods, in addition, the proposed method preserves the texture and edges of images while the methods of Yu and Acton (2002), Vese and Osher (2003), Wang et al. (2014) lead to smooth edges of images. Also, to evaluate the visual quality of the proposed approach, similarly to the methods mentioned in Yu and Acton (2002), Vese and Osher (2003), Wang et al. (2014), we employed the proposed method to two real noisy medical ultrasound images (see Figs. 5(b) and 5(c) same as in Yu and Acton (2002), Vese and Osher (2003), Wang et al. (2014)). The results of our proposed method in comparison with the results of state-of-the-art methods mentioned in Yu and Acton (2002), Vese and Osher (2003), Wang et al. (2014) are listed in Figs. 7 and 8.
In the second set, two other real ultrasound images, Fig. 9(a) and 9(b), from Siemens Healthcare (Siemens Healthcare GmbH, 2019), are considered and the proposed method is applied to reduce their speckle noise. The despeckled result of these images are shown in Fig. 9(c) and 9(d), respectively, which show that our proposed method can effectively suppress the speckle noises.
Generally, experimental results of the first and the second set of selected images illustrate that the proposed approach can obtain better performance in terms of PSNR and SSIM for ultrasound image denoising. From visual comparison, it is easy to see that our proposed method gets smooth effect while preserving the edges of images, which lead to maintain the useful texture information of test images.
Conclusions
In this paper, we have proposed a denoising method based on discrete shearlet transform and log-transform methods for speckle suppression in ultrasound images. Our numerical tests are done in MATLAB experiments. The experimental results on a sample cartoon-like image and two real ultrasound images polluted by multiplicative speckle noise illustrate the efficiency of the proposed method in terms of qualitative visual evaluation. Experimental results illustrate that the proposed approach can obtain better performance in terms of peak signal to noise ratio (PSNR) and structural similarity index (SSIM) between despeckled images and original images. This approach can be helpful to assist radiologists in their quest and diagnostics.
ReferencesAbazari, R., Lakestani, M. (2018a). Fourier based discrete shearlet transform for speckle noise reduction in medical ultrasound images. , 14(3), 477–483.Abazari, R., Lakestani, M. (2018b). A hybrid denoising algorithm based on shearlet transform method and Yaroslavsky’s filter. , 77(14), 17829–17851.Abbott, J.G., Thurstone, F.L. (1979). Acoustic speckle: theory and experimental analysis. , 1, 303–324.Averbuch, A., Coifman, R.R., Donoho, D.L., Israeli, M., Shkolnisky, Y. (2008). A framework for discrete integral transformations I – the pseudo-polar Fourier transform. , 30(2), 764–784.Bhuiyan, M.I.H., Ahmad, M.O., Swamy, M.N.S. (2009). Spatially adaptive thresholding in wavelet domain for despeckling of ultrasound images. , 3, 147–162.Burt, P.J., Adelson, E. (1983). The Laplacian pyramid as a compact image code. , 31(4), 532–540.Candes, E.J. (1998). Ridgelets: Theory and Applications. PhD thesis, Department of Statistics, Stanford University.Candes, E.J., Donoho, D.L. (2000). Curvelets: a surprisingly effective nonadaptive representation for objects with edges. Tech. report, Department of Statistics, Stanford University.Colonna, F., Easley, G.R., Guo, K., Labate, D. (2010). Radon transform inversion using the shearlet representation. , 29(2), 232–250.Cunha, A.L., Zhou, J., Do, M.N. (2006). The nonsubsampled contourlet transform: theory, design, and applications. , 15(10), 3089–3101.Deka, B., Bora, P.K. (2013). Wavelet-based despeckling of medical ultrasound images. , 59(2), 97–108.Do, M.N., Vetterli, M. (2005). The contourlet transform: an efficient directional multiresolution image representation. , 14(12), 2091–2106.Donoho, D.L., Huo, X. (2001). Beamlets and multiscale image analysis. Tech. report, Stanford University.Elmoniem, K., Youssef, A., Kadah, A. (2002). Real-time speckle reduction and coherence enhancement in ultrasound imaging via nonlinear anisotropic diffusion. , 49, 997–1014.Elyasi, I., Pourmina, M.A. (2016). Reduction of speckle noise ultrasound images based on TV regularization and modified bayes shrink techniques. , 127(24), 11732–11744.Goodman, J.W. (1976). Some fundamental properties of Speckle. , 66, 1145–1150.Guo, K., Labate, D. (2007). Optimally sparse multidimensional representation using shearlets. , 39, 298–318.Guo, K., Labate, D. (2012). Optimally sparse representations of 3d data with c2 surface singularities using parseval frames of shearlets. , 44, 851–886.Guo, K., Labate, D. (2013). The construction of smooth parseval frames of shearlets. , 8(1), 82–105.Gupta, S., Chauhan, R.C., Saxena, S.C. (2004). Wavelet-based statistical approach for speckle reduction in medical ultrasound images. , 42, 189–192.Hazarika, D., Nath, V.K., Bhuyan, M. (2015). A lapped transform domain enhanced lee filter with edge detection for speckle noise reduction in sar images. In: , pp. 243–248.Hiremath, P.S., Akkasaligar, P.T., Badiger, S. (2013). Speckle noise reduction in medical ultrasound images, advancements and breakthroughs in ultrasound imaging. . https://doi.org/10.5772/56519.Hou, B., Zhang, X., Bu, X. (2012). SAR image despeckling based on nonsubsampled shearlet transform. , 5(3), 809–823.Huang, Y., Hu, B., Zhu, J. (2016). Study on the use of quantitative ultrasound evaluation of diabetic neuropathy in the rat sciatic nerve. , 39(4), 997–1005.Kabir, S.M., Bhuiyan, M.I.H. (2015). Modeling of log-transformed speckle noise in the contourlet transform domain for medical ultrasound images. , 24(2), 1540004(35 pp.).Khare, A., Khare, M., Jeong, Y., Kim, H., Jeon, M. (2010). Despeckling of medical ultrasound images using Daubechies complex wavelet transform. , 90(2), 428–439.Krissian, K., Westin, C., Kikinis, R., Vosburgh, K. (2007). Oriented speckle reducing anisotropic diffusion. , 16, 1412–1424.Kutyniok, G., Lim, W. (2011). Image separation using wavelets and shearlets. In: , Avignon, France, 2010, Lecture Notes in Computer Science, Vol. 6920.Labate, D., Lim, W.-Q., Kutyniok, G., Weiss, G. (2005). Sparse multidimensional representation using shearlets. In: , SPIE Proc., Vol. 5914. SPIE, Bellingham, WA, pp. 254–262.Lakestani, M., Razzaghi, M., Mousavi, Z. (2016). Combined shearlet shrinkage and total variation minimization for image denoising. https://doi.org/10.22099/IJSTS.2016.3677.Loupas, T., Mcdicken, W., Allan, P.L. (1989). An adaptive weighted median filter for speckle suppression in medical ultrasonic images. , 36(1), 129–135.Meyer, F.G., Ronald, R.C. (1997). Brushlets: a tool for directional image analysis and image compression. , 4(2), 147–187.Non-subsampled contourlet transform (2008). MATLAB source, Retrieved from http://http://minhdo.ece.illinois.edu/software/.Patel, V.M., Easley, G., Healy, D.M. (2009). Shearlet-based deconvolution. , 18, 2673–2685.Ragesh, N.K., Anil, A.R., Rajesh, R. (2011). Digital image denoising in medical ultrasound images: a survey. In: , Dubai, UAE, pp. 12–14.Rajeshwar, D. (2018). Speckle noise reduction of ultrasound images using BFO cascaded with wiener filter and discrete wavelet transform in homomorphic region. , 132, 1543–1551.Ritenour, E.R., Nelson, T.R., Raff, U. (1984). Application of median filter to digital readiographic images. In: , Vol. 23, pp. 1–4.ShearLab (2008). . Retrieved from http://www.shearlet.org.Siemens Healthcare GmbH (2019). Available: http://www.healthcare.siemens.com/ultrasound.Vard, A., Jamshidi, K., Movahhedinia, N. (2012). An automated approach for segmentation of intravascular ultrasound images based on parametric active contour models. , 35(2), 135–150.Vese, L., Osher, S. (2003). Modeling textures with total variation minimization and oscillating patterns in image processing. , 19(11), 553–572.Vishwakarma, A., Bhuyan, M.K., Iwahori, Y. (2018). Non-subsampled shearlet transform-based image fusion using modified weighted saliency and local difference. , 77(24), 32013–32040.Wang, G., Xu, J., Pan, Z., Diao, Z. (2014). Ultrasound image denoising using backward diffusion and framelet regularization. , 13, 212–217.Wang, Z., Simoncelli, E.P., Bovik, A.C. (2004). Multiscale structural similarity for image quality assessment. In: , Vol. 2, pp. 1398–1402.Yu, Y., Acton, S. (2002). Speckle reducing anisotropic diffusion. , 11, 1260–1270. 2002.Zong, X., Laine, A.F., Geiser, E.A. (1998). Speckle reduction and contrast enhancement of echocardiograms via multiscale nonlinear processing. , 17, 532–540.Zhang, Y., Wang, Y., Wang, W., Liu, B. (2001). Doppler ultrasound signal denoising based on wavelet frames. , 48, 709–716.