Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Introduction

Ultrasound is a real-time, non-invasive, portable, versatile and relatively low cost, compared with other medical imaging techniques that allow the imaging of the internal body in real-time. Ultrasound has been widely accepted in the medical area since its emergence as a tool to help in the surgical guidance, preoperative planning and diagnose. This technique is preferred over techniques such as Magnetic Resonance Imaging (MRI) or the X-rays [1, 2], especially for the pregnant women, patients with arrhythmias or MR contraindications. For example, during a MRI study, the patient must remain very still to acquire a clear image and the powerful X-rays have the ability to create birth defects, diseases and can alter the DNA.

Ultrasound imaging technique uses ultrasonic waves produced from the transducer and travel through the patient’s body. When the wavefront hits a discontinuity, scattered waves are produced and echoes are bounced back. These echoes are detected by the same transducer, then processed and displayed by the equipment [3]. The noise is introduced by particles of tissues with size less than the wavelength applied and corrupts the ultrasound images. The interference of the sound waves scattered, received from different points in the imaged organ distorts important details during the acquisition stage. The resulting speckle noise pattern is visible in the image as dark and light spots of different sizes [4].

The speckle noise is one of the most important issues in ultrasound images because the poor signal to noise ratio may drastically affect the final diagnose. Therefore, the need of a denoising step may be vital to avoid errors in the diagnostic. The noise can be additive or multiplicative, depending on the modality used to acquire the image [5, 6]. Additive noise is easier to reduce and more tractable because the multiplicative noise is signal dependent.

The current methods to denoise make use of a trade-off between the reduction of unwanted data that distort the shape of the object of interest in the image and the image itself. The noise follows different probability distributions depending on the modality used. For example, in the ultrasound modality the model is multiplicative and follows a Rayleigh distribution. In other modalities, such as MRI and X-ray, the noise is Rician and Poisson distributed respectively [7].

In this chapter, an evaluation of denoising methods used to remove speckle noise, in the spatial domain, in ultrasound medical images, is carried out, with the aim of comparing their performance. The remaining of the chapter is organized as follows. In Sect. 13.2, a brief review of current methods in the state-of-the-art is presented and discussed. In Sect. 13.3, the multiplicative noise model of the ultrasound medical images is discussed. In Sect. 13.4, the filters used as test benches to remove speckle noise, against which all other methods compare their performance are discussed. In Sect. 13.5, the metrics used are explained. In Sect. 13.6, the phantom, the visual and the quantitative results of the filters shown in Sect. 13.5 are presented. Finally, the chapter concludes in Sect. 13.7.

2 Review of Speckle Noise Reduction Methods for Ultrasound Medical Images

The speckle noise can be reduced by increasing the ultrasound frequency. However, the higher the frequency contents the lower the penetration in the tissue [8, 9]. Consequently, most of the major commercial ultrasound systems software such as, Siemens, Kostec and GE ViewPoint 6, include Computer Aided Diagnosis (CAD) techniques for speckle noise reduction. CAD helps in the screening and avoids unnecessary biopsies [10,11,12,13,14,15]. Usually, texture features are used to discriminate normal, benign and malignant tissues in prostate analysis. Then, speckle filtering is an important pre-processing step in the analysis of ultrasound medical imaging. The speckle reduction methods can be classified into two main categories, those that process the image in the transform domain and those that process in the spatial domain (image). Even though in both categories the denoised image has good quality, the second category is preferred because is faster and can be used for real-time applications. However, extensive research has been done in both categories to yield good quality images in a reduced timeframe. Following, the most important works carried out in these two categories are presented.

2.1 Transform Domain Methods

Transform methods change the image to the frequency domain by using transforms such as, Fourier, wavelets, shearlet, curvelet and contourlet among others. After a filtering or a thresholding process, applied to the transform coefficients to reduce the noise, the denoised image is recovered by applying the inverse transform.

These methods have a high computational complexity due to the transformation and anti-transformation steps, and may insert artificial frequencies to the recovered image. Following, the most important methods found in the literature are explained.

In the Fast Fourier Transform (FFT) method, the image is enhanced, before transformation, to better detect the noise. Then, the FFT is applied to the image. The frequencies that represent dark and light spots (speckle) are searched, located and masked with a special function before recovering the image [16]. Notice that, besides the time introduced by the transformation, the searching operation is computationally expensive.

The wavelet transform is a multiscale and multi-resolution tool widely used in image processing [17]. Before transformation, the multiplicative noise model is transformed into an additive model by computing the logarithm of the image. Then thresholding is applied to shrink the coefficients. The main strength of wavelet thresholding is its capability to process the different frequency components separately [18,19,20,21]. Similarly, the Wiener filtering has been proposed as alternative to denoise the detail subbands [22]. Framelet algorithms, based on wavelet frames, combined with regularization terms, such as the total variation (TV) [23], have the advantages of multi-resolution analysis, remove noise and preserve edges [24, 25]. However, the use of TV may suppress the texture features of the image. The combination of wavelets, principal component analysis (PCA) and thresholding operations [26] has shown some improvements. Nevertheless, the calculation of the singular value decomposition of the covariance matrix, to compute the principal components, is computationally extensive. Furthermore, wavelets and non-linear diffusion [27] based on the iterative edge enhancement feature of nonlinear diffusion have shown interesting results in removing speckle noise while preserving the edges. Most of the efforts to despeckle ultrasound images have been done in the wavelet domain [28,29,30,31,32,33,34,35]. Nevertheless, wavelet-based methods require converting the image to the Cartesian space. This contributes to extra computational time and reduces solution accuracy.

Shearlet is multiscale transform, for signal representation, that provides directional selectivity, localization, anisotropy and shift invariance [36, 37]. Shearlet is an extension of wavelets and include the concept that the subbands capture all the anisotropic features such as edges. Wavelets cannot represent well the edges because they are isotropic objects [38]. Based on this, some works have been proposed to remove speckle noise in the shearlet domain. Most of them filter the transform coefficients to estimate the denoised shearlet coefficient. For example, an anisotropic diffusion method is applied to the noisy coefficients to reduce the noise and preserve the edge [39, 40].

Curvelets is another multiscale transform for representing very general linear symmetric systems of hyperbolic differential equations [41, 42]. Researchers have paid much attention to this transform [43,44,45]. The works estimate the unknown curvelet coefficients using similar filters as the used in the wavelet domain. For example, in [44], a threshold is calculated using SURE-LET [45] strategy and applied to the detail subbands. The results of shrinkage are further processed by using a nonlinear diffusion technique. In [46] the curvelet coefficients are modeled and processed with the Perona and Malik Anisotropic Diffusion filter [47]. A maximum a posteriori threshold is calculated to further process the coefficients and to avoid artifacts and to recover the edges.

The contourlet transform can deal effectively with smooth contours [48]. The pyramidal filter bank structure of the contourlet transform has very little redundancy. However, this transform is not shift-invariant. Conversely, the Non-Subsampled Contourlet Transform (NSCT) is fully shift-invariant, multi-scale and multi-direction expansion; most of the wavelet-based denoising methods have been extended to the NSCT for multiplicative noise removal assuming that the NSCT coefficients follow a generalized Gaussian distribution (GGD). However, this extension is not trivial [49,50,51,52,53,54].

2.2 Spatial Domain Methods

In the spatial domain denoising methods, no transformation is carried out on the original image. All the process is performed in the spatial domain (image domain). These methods are preferred for real-time application because they are less time-consuming [54]. Following, the most important methods found in the literature are explained.

In the non-local means (NLM) methods, the estimated pixel considers the weighted average of all pixels in the image. The weight depends on the similarity of the pixel under estimation with respect to another pixel usually inside a window. The weight depends on the noise deviation and the weighted Euclidean distance of the two pixels [55,56,57]. The algorithms are computationally expensive to be widely used in real-time applications. However, the recovered images show good removal of speckle noise and edge preservation.

The Oriented SRAD [58] is an extension of the Speckle Reducing Anisotropic Diffusion (SRAD) method [59] and is based on matrix anisotropic diffusion to obtain different diffusions across the principal curvature directions [60].

Methods based on the Total Variation (TV) [23] use a similarity term to measure the amount of smoothing in the image and a regularization term as a prior knowledge of the noise distribution. In ultrasound, the later term follows a Rayleigh distribution. The shortcoming is that noise samples can be large and an edge could be confused with a noise sample and vice versa. That is, if the noise is removed the edges can be blurred and if the edge is preserved, the noise could not be well-removed [61, 62].

The time-domain deconvolution methods yield good results. These methods rely on the Point Spread Function (PSF) calculation, for example the algorithms based on phase unwrapping and a noise-robust procedure to estimate the pulse in the complex cepstrum domain, then a robust estimate of the PSF can obtained to reduce the noise level linearly with the number of pulses estimates. The Wiener filter is used for subsequent deconvolution with sharper images than the original image because of the deconvolution. Other deconvolution method acts on the envelope of the acquired radio-frequency signals. The measured data is used to estimate a point spread function (PSF) the image is reconstructed in a non-blind way. These methods represent a major problem when the PSF has to be estimation [63,64,65,66,67,68,69].

3 Multiplicative Noise Model

The multiplicative noise, known as speckle, occurs in images acquired by coherent imaging systems such as Laser, SAR, Optical and ultrasound systems. This noise causes serious problems to represent an image. The intensity values of the pixels are multiplied by random values. The probability density function of the speckle noise follows a gamma distribution of the form:

$$P(x) = \frac{{x^{\alpha - 1} }}{{(\alpha - 1)!a^{\alpha } }}e^{{ - \tfrac{x}{a}}} ,$$
(3.1)

where x is the intensity of the pixel and the variance is \(a^{2} \alpha\).

The observed samples are the output of the ultrasound imaging system and can be represented by the vector \({\mathbf{g}} = \left[ {g_{1} , \ldots ,g_{n} } \right]\). The samples of the speckle noise and the noise free image are the vectors \({\mathbf{n}} = \left[ {n_{1} , \ldots ,n_{n} } \right]\) and \({\mathbf{f}} = \left[ {f_{1} , \ldots ,f_{n} } \right]\) respectively. The additive noise vector is \({\varvec{\upeta}} = \left[ {\eta_{1} , \ldots ,\eta_{n} } \right]\). Therefore, the speckled image is commonly modeled as

$${\mathbf{g}} = {\mathbf{fn}} + {\varvec{\upeta}}.$$
(3.2)

In general, it is widely accepted that the additive noise (from sensor) be considered small as compared to the multiplicative part. Hence, Eq. (3.2) can be reduced to:

$${\mathbf{g}} = {\mathbf{fn}}.$$
(3.3)

Despeckling algorithms that use the model of Eq. (3.3) require two important conditions: preservation of the mean intensity value over homogeneous areas and preservation of important details such as texture and edges while removing the noise. The multiplicative speckle model was the base to development the minimum mean-square error (MMSE) Lee [70], Kuan et al. [71], Gamma MAP [72] and Frost et al. [73] filters. The Speckle Reducing Anisotropic Reduction (SRAD) [59] is based on the analysis of the Frost et al. and Lee filters as isotropic filters and the Perona and Malik model [47].

4 Speckle Denoising Filters Compared in This Chapter

Following, the comparison of filters that use the multiplicative model to remove speckle noise is described. We implemented the original version of every method As it can be seen in Sect. 13.2; many of the proposed methods could outperform the compared methods. However, they are hybrid methods (i.e. wavelets and thresholding or wavelets and TV or TV plus more terms), their performance depends on the input image and the selected transform and the implementation is computationally expensive. The following speckle filters are considered the test benches against which most of the methods are compared.

4.1 Average Filter

The image is processed using a sliding window of size 2k + 1. It is also called neighborhood average method. It does not remove the speckle noise but integrates it incoherently into the mean value. In this procedure, the central pixel \(\hat{f}\) of the sliding window is replaced by the average intensity, yielding a smooth image with blurred edges. The average filer is defined as:

$$\hat{f}_{n} = \frac{1}{2k + 1}\sum\limits_{i = - k}^{k} {g_{i} } .$$
(4.1)

The filter is optimal for images with additive Gaussian noise. In other words, it achieves the Kramer-Rao lower bound. However, it is not optimal for the case of a multiplicative model.

4.2 Median Filter

The median filter is a non-linear local procedure used to reduce speckle noise and to retain steps and ramp functions. The filter is robust to impulsive type noise with the particularity of preserving the edges of the image. Hence, it produces a less blurred image. The image is analyzed using a sliding window on the image in order to replace the central value of the window by its median intensity. In other words, the noisy pixel is replaced by its median value \(\hat{f}_{n}\). Hence, the noise is reduced without blurring the edges. If the window length is 2k + 1 the filtering is given by:

$$\hat{f}_{n} = med\,\left[ {g_{n - k} , \ldots ,g_{n} , \ldots ,g_{n + k} } \right] ,$$
(4.2)

where med[∙] is the median operator. To find the median it is necessary to sort all the intensities in a neighborhood into numerical order. This is a computationally complex process due to the time needed to sort pixels to find the median value of the window. Yet, the filter is good at preserving edges.

4.3 Frost Filter

The frost filter [73] is an adaptive Wiener filter that reduces the multiplicative noise while preserving edges. It replaces the central pixel of a window of size 2k + 1 by the sum of weighted exponential terms. The weighting factors depend on the distance to the central pixel, the damping factor, and the local variance. The more far the pixel from the central pixel the less the weight. Also, the weighting factors increase as variance in the window increases. The filter convolves the pixel values within the window with the exponential impulse response:

$$h_{i} = e^{{ - K\alpha_{g} (i_{0} )\left| i \right|}} ,$$
(4.3)

where K is the filter parameter, i 0 is the location of the processed pixel, |i| is the distance measured from the pixel i 0. The coefficient of variation is defined as \(\alpha_{g} = \sigma_{g} /\bar{g}\); where \(\bar{g}\) and \(\sigma_{g}\) are the local mean and standard deviation of the window. The optima MMSE Wiener filer in a window with normalization constant K 1 becomes:

$$\hat{f}_{n} = \sum\limits_{i = - k}^{k} {K_{1} } h_{i} g_{i} .$$
(4.4)

In order to preserve the mean value, K 1 is computed by using:

$$K_{1} = \frac{1}{2k + 1}\,\,\sum\limits_{i = - k}^{k} {h_{i} } .$$
(4.5)

4.4 Kuan et al. Filter

The Kuan et al. filter [71] is a local minimum mean square error method used to restore images with signal-dependent noise. The filter considers the multiplicative noise model of Eq. (3.3) as an additive model of the form:

$${\mathbf{g}} = {\mathbf{f}} + ({\mathbf{n}} - 1){\mathbf{f}}.$$
(4.6)

Assuming unit-mean noise, the estimate pixel value \(\hat{f}_{n}\) in the local window is:

$$\hat{f}_{n} = \bar{g} + \frac{{\sigma_{f}^{2} (g_{n} - \bar{g})}}{{\sigma_{f}^{2} (ENL + 1)\bar{g}^{2} }}ENL,$$
(4.7)

with

$$\sigma_{f}^{2} = \frac{{ENL\sigma_{g}^{2} - \bar{g}^{2} }}{ENL + 1} ,$$
(4.8)

and

$$ENL = \left( {\frac{Mean}{StDev}} \right)^{2} = \left( {\frac{{\bar{g}}}{{\sigma_{g}^{2} }}} \right)^{2} .$$
(4.9)

The Equivalent Number of Looks (ENL) estimates the noise level and is calculated in a uniform region of the image. One shortcoming of this filter is that the ENL parameter needs to be computed beforehand.

4.5 Lee Filter

The Lee filter [70] uses the Digital Number (DN) values calculated in the window to estimate the pixel under processing. Unlike a typical low-pass smoothing filter, the Lee filter and other similar sigma filters preserve image sharpness and details while reducing the noise. The pixel being filtered is replaced by a value that is calculated using the surrounding pixels. If the variance is low, smoothing will be performed. In the other hand, if the variance is high, assuming an edge, the smoothing will not be performed. Therefore, Eq. (4.7) can be simplified as:

$$\hat{f}_{n} = \bar{g} + k(g_{n} - \bar{g}),$$
(4.10)

where k is an adaptive filter coefficient. The Lee filter is a particular case of the Kuan et al. filter without the term \(\sigma_{f}^{2} /ENL\).

4.6 Gamma MAP Filter

The Gamma Maximum A Posteriori (MAP) filter [72] is based on a Bayesian analysis of a multiplicative noise model and assumes that the image and the noise follow a Gamma distribution. The pixel being filtered is replaced with a value that is calculated based on the local statistics. The filter takes into consideration not only the noise but also the image. The Gamma MAP filter of the pixel under processing is given by:

$$\hat{f}_{n} = \frac{{\bar{g}(\alpha - ENL - 1) + \sqrt {4\alpha \,ENL\,g\,\bar{g} + \bar{g}^{2} (\alpha - ENL - 1)^{2} } }}{2\alpha },$$
(4.11)

where

$$\alpha = \frac{ENL + 1}{{ENL(\sigma_{g} /\bar{g})^{2} - 1}}.$$
(4.12)

4.7 Anisotropic Diffusion

Diffusion is a physical phenomenon that is used in the image-processing field to denoise images and to detect edges. The phenomenon aims at minimizing the spatial concentration g(x; t) of a substance. In other words, the goal is to minimize the differences in values of the pixels belonging to similar regions. This process is described by the Fick’s law that states that concentration differences induce a flow j of a substance in direction of the negative concentration gradient. Therefore, the flow can be expressed as:

$$j = - c{\nabla }g,$$
(4.13)

where c is the diffusivity that describes the speed of the diffusion process from one point to another and \({\nabla }\) (nabla) is the gradient operator. Once the flow is described, the continuity equation is used to observe the change in time of the concentration g and is expressed by the negative of the divergent of the flow:

$$\partial_{t} g = - div\,j .$$
(4.14)

The diffusion equation is obtained by replacing Eq. (4.13) into Eq. (4.14) as:

$$\partial_{t} g = div(c{\nabla }g) .$$
(4.15)

If the diffusivity c is a constant, i.e. c = 1, the process is linear isotropic and homogeneous. If c depends on the concentration g, c = c(g), and the process becomes a nonlinear diffusion. However, if c is a matrix-valued diffusivity, the process is called anisotropic and it will lead to a process where the diffusion is different for different directions. In image processing, the goal is to use the anisotropic process to obtain less diffusion in edges. In other words, the diffusivity should decrease with strong gradients. Perona and Malik [47] proposed a generalization of the diffusion equation to denoise images without blurring the edges. Base on this, the speckle reducing anisotropic diffusion (SRAD) was later proposed by Yongjian and Acton [59]. The approach is based on the minimum mean square error (MMSE) approach of the Lee and Frost filters and extended the Perona and Malik algorithm for images corrupted by speckle noise.

4.7.1 Speckle Reducing Anisotropic Diffusion (SRAD)

Yongjian and Acton [59] rearranged the Eq. (4.7) as:

$$\hat{f} = g + (1 - k)(\bar{g} - g) .$$
(4.16)

The term \((\bar{g} - g)\) can be seen as an approximation to the Laplacian operator (with c = 1). Then, Eq. (4.16) can be expressed as:

$$\hat{f} = g + k^{\prime}\,div\,(\nabla g) .$$
(4.17)

Equation (4.17) is an isotropic process. Hence, Eq. (4.15) can be easily transformed into an anisotropic version by including only the c factor:

$$\partial_{t} g = div(c{\nabla }g) = c\,div\,({\nabla }g) + {\nabla }c{\nabla }g .$$
(4.18)

The SRAD filter can be stated as follows. Given an initial condition that indicates the amount of concentration at time t = 0, with finite energy and no zero values over the image domain Ω. The output image g(x, y; t) is evolved according to the following Partial Derivative Equation (PDE).

$$\left\{ {\begin{array}{*{20}c} {\frac{\partial g(x,y;t)}{\partial t} = div\,\left[ {c(q){\nabla }g(x,y;t)} \right]} \\ {g(x,y;0) = g(x,y)} \\ {\left. {\frac{\partial g(x,y;t)}{{\partial \vec{n}}}} \right|_{\partial \varOmega } = 0} \\ \end{array} } \right.$$
(4.19)

Observe that c(q) is the diffusion coefficient and g(x, y; t) is the instantaneous coefficient of variation. In other words, it is the edge detector. The last boundary condition states that, the derivative of the function along the outer normal, at the image boundary, must vanish. This assures that no concentration (brightness) will leave or enter the image, i.e. the average brightness will be preserved.

5 Metrics

Besides the mean and variance of the reconstructed images, the filters were evaluated on the synthetic image using the mean square error (MSE) and in the cyst phantom using the contrast to noise ratio (CNR) and the lesion to background contrast (CLB) [74]. The metrics are defined as follows:

$$MSE = \frac{1}{M \times N}\sum\limits_{m = 0}^{M - 1} {\sum\limits_{n = 0}^{N - 1} {\left[ {I(m,n) - \hat{I}(m,n)} \right]} }^{2} .$$
(5.1)

M × N is the image size, I is the clean or reference image and \(\hat{I}\) is the filtered image.

$$CNR = \frac{{\left| {\mu_{L} - \mu_{B} } \right|}}{{\sqrt {\sigma_{B}^{2} + \sigma_{L}^{2} } }} .$$
(5.2)
$$CLB = \frac{{\mu_{L} - \mu_{B} }}{{\mu_{B} }} .$$
(5.3)

Notice that \(\mu_{L}\) and \(\sigma_{L}^{2}\) are the mean and variance in the lesion (cyst) respectively. \(\mu_{L}\) and \(\sigma_{B}^{2}\) are the mean and variance of intensities of pixels in the background region.

6 Results

The synthetic datum is a 256 × 256 image, gray scale, 8 bits per pixel. The image was contaminated with speckle noise of variance \(\sigma_{N} = 0.02\) and \(\sigma_{N} = 0.02\). The variance, mean and MSE of the denoised image was calculated with respect to the clean image (true image). To evaluate the CNR and the CLB a cyst from a B-mode image cyst Phantom-Field II ultrasound simulation [75, 76] was used. The experiments were carried out in an Intel core i5 1.60 GHz processor with the visual C# software.

Figures 13.1 and 13.2 show the resulting images for two different noise variances. Table 13.1 shows the quantitative results. We can observe that the Frost, the Lee, the Kuan, the Gamma MAP and the SRAD filters yield sharper recovered images sharper images and in the case of the Median and the Frost filters most of the edges are preserved but the images are blurred.

Fig. 13.1
figure 1

Synthetic image used to compare the algorithms. a Clean image. b Noisy image \(\sigma_{\text{N}} = 0.02\). Images filtered by using the c mean, d median, e Frost et al., f Lee, g Kuan et al., h Gamma MAP and i SRAD filters

Full size image
Fig. 13.2
figure 2

Synthetic image used to compare the algorithms. a Clean image. b Noisy image. \(\sigma_{\text{N}} = 0.2\) Images filtered by using the c mean, d median, e Frost et al., f Lee, g Kuan et al., h Gamma MAP and i SRAD filtered images

Full size image
Table 13.1 Mean, standard deviation and MSE of the recovered synthetic images after filtering
Full size table

In Table 13.1 we can see that the Median, Frost et al., Lee, Kuan et al., Gamma MAP and SRAD filters increased the intensity value up to a certain point but the mean is not well preserved. The Average filter decreased the mean and produced a more blurred image. One reason for this is the multiplicative nature of speckle noise, which relates the amount of noise to the signal intensity. The other reason is that the filter is not adaptive in the sense that do not account for the particular speckle properties of the image. Notice that for low noise power the Gamma MAP yielded the least MSE. However, for a high noise power the MSE increased. In the case of high variance noise the Median filtering yielded the least MSE. In the SRAD filter preserves the mean well and yields a good visual result.

Figure 13.3 shows a 3-D mesh of the simulated lesion (cyst) of each method. The black arrow points to the lesion. Note that the SRAD method yields the best result because it reduces the noise and preserves the lesion yielding the best CLB as shown in Table 13.2. However, SRAD took 200 iterations to obtain the final result. Also, the Kuan et al., and Gamma MAP filters yield better results than the remaining of the filters.

Fig. 13.3
figure 3

3-D Mesh of the simulated B-mode cyst Phantom Field-II used to compute the CNR and CLB. a Original image. Mesh after processing with the b mean, c median, d Frost et al., e Lee, f Kuan et al., g Gamma MAP and h SRAD-200 iterations filters

Full size image
Table 13.2 Results of Contrast to Noise Ratio (CNR) and Lesion to Background Contrast (CLB) fort the simulated B-mode cyst Phantom Field-II
Full size table

The quantitative results for the cyst phantom are shown in Table 13.2. It can be seen that the SRAD yields the best result in 200 iterations.

The results of filtering a real ultrasound image with an ovarian cyst are shown in Fig. 13.4. It can be observed how SRAD filter performs well however the image texture is lost.

Fig. 13.4
figure 4

Ultrasound images of a real ovarian cyst. a Original image. Filtered images by b mean, c median, d Frost et al., e Lee, f Kuan et al., g Gamma MAP and h SRAD–200 iterations filters

Full size image

7 Conclusions

Ultrasound is becoming more useful for medical diagnostics. Speckle introduced during the image acquisition distorts the image and can lead to erroneous diagnostics. Medical application software’s include post-processing techniques to reduce the speckle that contaminates the ultrasound images and improve the shape of the images to help in the diagnostic. In this chapter, several types of filters, applied in the spatial domain of the image, have been evaluated. It was seen that the Median filter can preserve the edges; however, the mean of the image is altered. The Lee, the Frost, and the Gamma Map filter seem similar in reducing speckle noise. Nevertheless, Lee, Kuan and Frost yield better results in reducing speckle in texture areas. Gamma MAP preserves more texture while SRAD removes more noise and texture. The denoising methods discussed have been integrated in a medical tool, developed in visual C# to enhance images and help in the analysis of medical image in the patient’s diagnostic stage.