Visual enhancement of digital ultrasound images: wavelet versus Gauss-Laplace contrast pyramid

1

Visual enhancement of digital ultrasound images: Wavelet versus Gauss-Laplace contrast pyramid

Ali S. Saad,

King Saud University

College of applied medical sciences-

Dept. of Biomedical Technology,

P.O. Box 10219, Riyadh 11433,

Kingdom of Saudi Arabia

Phone:966-1-4355010 (307), mobile: 966508975969.

Emails: [email protected], [email protected]

Abstract words count : 205

Ms words count : 4,750

Number of references : 43

2

Abstract:

Purpose

Noise is the principal factor which hampers the visual quality of ultrasound images, sometimes leading to

misdiagnosis. Speckle noise in ultrasound images can be modeled as a random multiplicative process.

Speckle reduction techniques were applied to digital ultrasound images to suppress noise and improve

visual quality.

Rationale

Previous reports indicate that wavelet filtering performs best for speckle reduction in digital ultrasound

images. Reportes on x-ray images compared wavelet filtering with Laplace-Gauss contrast enhancement

(LGCE) showed that the LCGE performed better. As LGCE was never been applied to Ultrasound

images, this study compared two filtering approaches for speckle reduction on digital ultrasound images.

Methods

Two methods were implemented and compared. The first method uses the wavelet soft threshold (WST)

approach for enhancement. The second method is based on multi-scale Laplacian-Gaussian contrast

enhancement (LGCE).

LGCE is derived from the combination of a Gaussian pyramid and a

Laplacian one. Contrast enhancement is applied on local scale by using

varying sizes of median filter.

Results

The two methods were applied to synthetic and real ultrasound images. A comparison between WST and

LGCE methods was performed based on noise level, artifacts and subjective image quality.

Conclusion

WST visual enhancement provided better results than LGCE for selected ultrasound images.

Keywords: Speckle reduction, wavelet filtering, Laplacian pyramid, ultrasound images

and contrast enhancement.

I. Introduction

Ultrasound imaging techniques are widely used in medical diagnosis. Its noninvasive

nature, low cost, portability, and real-time image formation make ultrasound imaging

an attractive means for medical diagnosis. One of the limitations of ultrasound images

is poor image quality affected by speckle noise. Speckle reduction remains a difficult

problem due to the lack of reliable models to estimate noise.

Ultrasound images are very difficult to diagnose because of the existence of speckle

which hampers the perception and the extraction of fine details from the image. The

speckle is a characteristic phenomenon in different fields such as in Laser, Synthetic

Aperture Radar (SAR) images or in ultrasound images. Its effect is a granular aspect in

the image. Speckle is caused by interference between coherent waves which are

3

backscattered by targeted surfaces and arrive out of phase at the sensor [1]. Speckle can

be modeled as random multiplicative noise [1,2]. Filtering (visual enhancement)

techniques are used for enhancing the visual quality of the image. Furthermore, they are

used as preliminary treatment before segmentation and classification. Several filtering

methods have been proposed for speckle reduction. They can be divided roughly into

two categories. The first category improves the image by summing several observations

of the same object, the assumption that no change or motion of the object occurred

during the reception. The second category enhances the image using statistical models

for both noise and signal. Some techniques are applied in the spectral domain [3]. Other

techniques, such as the Kalman recursive filter (2D Bayesian estimation), are applied in

real domain. In reality, it is very difficult to represent a natural area by a simple

statistical model; moreover, these techniques are costly from a computational

viewpoint. Within this second category, several methods have been developed in order

to reduce speckle in a variety of imaging areas, with some using adaptive techniques

[4]. In [5] a review of wavelet applications in biomedical signals is summarized. In [6]

a multi-scale speckle reduction method was compared to a large number of filtering

methods for multiplicative noise reduction on SAR images which use the same

multiplicative noise as ultrasound [7]. The methods used include the Frost method [8] –

which is based on a multiplicative model of noise and gamma distribution for signal

and χ2 for the noise – and a modified version of the Frost method [9]. Other methods

used include Lee's speckle reduction method which is based on local statistics [10,11],

Lee's non-biased sigma filtering[12], adaptive local ranking filter (based on first and

second order local statistics[13]), mean filter, median filter, multi-resolution filtering by

contrast modification using the median filter and an adaptive ranking method with

entropy calculation for local disorder[6]. The study in [6] shows that multi-resolution

filtering by contrast modification using the median filter gives the best compromise

between edge preservation and multiplicative noise reduction.

Another iterative algorithm which uses wavelet maxima [14], was employed for

contrast enhancement of mammograms [15]. This algorithm is especially suitable for

compression. In this method the positions and the magnitudes of the edges in the details

images of the wavelet decomposition are retained, it is a very good tool for noise

reduction but has the drawback that the algorithm is an iterative process. In addition,

the convergence speed of the algorithm is quite slow. Wavelet filtering was used for X-

ray images and mammogram contrast enhancement [16]. A Fast Wavelet Transform

(FWT) for radiography enhancement versus Lapalcian pyramid is proposed in [17]

which deal only with x-ray images and mammography. This study concluded that

contrast enhancement using the LGCE method gave better results than with the wavelet

method. However, the study did not deal with ultrasound images. The result of [17]

motivated the study on ultrasound images, which is similar to X-ray if a log-transform

on the ultrasound data was performed.

Wavelet speckle reduction in ultrasound was tackled in [18,19]. Speckle reduction and

contrast enhancement using multi-scale non linear processing of echocardiograms was

proposed in [20]. These methods used statistical models which though costly (from a

computational and modeling estimation viewpoint) produced excellent results.

In [21] a comparative study of speckle filtering in ultrasound imaging of the carotid

artery found that, in this specific type of image, filtering using local statistics and

geometric filtering showed better results than other filters including WST. Manual

estimation of standard deviations (STDV) of noise was performed from different

4

selected regions and threshold estimation was performed as in [22]. The multi-scale

contrast enhancement method was omitted from the study as were other interesting

filtering methods such as the Frost method.

In [23] multi-resolution filtering by contrast modification using the median filter was

applied on ultrasound images and proved to give a very good visual improvement.

Recently in [7] a study of speckle filtering of medical ultrasound images suggested that

WST and modified WST speckle reduction method perform the best.

In [17] the study shows that contrast enhancement using the LGCE method gives better

results than wavelet filtering for x-ray images. In [6] the LGCE pyramidal method

yielded better results than large filtering methods with the sole exception of WST

filtering. In [7,19,20] WST proved to be among the best for despeckling in ultrasound

images. From the above studies we can see that the LGCE filter and the WST filter

have not been compared to each other on ultrasound images. For the above reasons

these methods were chosen in this paper to be compared to each other on ultrasound

images. The purpose of this paper is to apply these methods and compare them to each

other. The first method is based on wavelet filtering using soft-thresholding [22] of the

wavelet details components. The second method is based on enhancing of the image

contrast on a Laplace-Gauss pyramidal multi-scale scheme [6, 23]. Contrast is modified

at each level in the pyramid using an adaptive filter. The filtered image is in fact the

one reconstructed from the modified contrast process. Results of these techniques are

presented and compared to each other, using synthetic and real ultrasound images.

In [24] some generalized criteria of image quality measurement are introduced. In a

recent publication [25] several criteria of image assessment were used, 3 groups of

criteria of assessment were used. The first group is used in texture analysis, the second

group is a general quantitative assessment such as the SNR, MSE, RMSE, and other

types of error using Minkowski distance. The second group also includes the so called

universal criteria of [24]. The third group is an interesting one, using visual inspection

of the images by two experts. The problem of these types of assessment using so many

criteria on the whole image is so confusing and difficult to analyze without any

weighting among those parameters and focus the analysis to a certain areas of interest.

In this study, three criteria for speckle reduction assessment are used for ultrasound

images, three quantitative criteria SNR and MSE were used for synthetic images and

standard deviations in homogenous regions (selected regions of interest) were used for

real images. Two qualitative assessments – one is realized by an expert in ultrasound

diagnosis and other showing the general pattern of error images between filtered

ultrasound image and unfiltered one for both real and synthetic images.

The structure of the paper is as follows. Section 2, presents the methodology used here

including the wavelets decomposition, wavelet Basis selection for ultrasound images,

soft-thresholding method and the pyramidal Laplace-Gauss contrast enhancement. In

section 3, a comparison of the effects of contrast enhancement in the frame work of

these different decomposition methods is presented. In section 4, a discussion of the

merits and problems of these methods follows the influence of multi-resolution

properties on the perception of the ultrasonic images. Finally a conclusion will be

drawn from this study.

5

II. Methods

II.1. Model of Speckled Ultrasound Image

A universally agreed upon definition of a model still seems to be lacking. Nevertheless,

a number of possible formulations, whose feasibility was verified via their practical

use, exist. A generalized model of the speckle imaging as proposed in [26] and used in

[27], [28] is given by:

g(x,y) = f(x,y) u(x,y) + α(x,y), (1)

where g, f, u, and α stand for the observed envelope image, original image,

multiplicative and additive components of the speckle noise, respectively. Here the

indices x and y denote the axial and lateral indices of the image samples. Despite its

possible theoretical shortcomings [2], the model (1) has been successfully used both in

ultrasound and SAR imaging. Moreover, evidence exists that, when applied to

ultrasound images, only the multiplicative component u of the noise is considered,

hence, (1) can be considerably simplified by disregarding the additive noise term. This

leads to the following simplified model:

g(x,y) = f(x,y)u(x,y) (2)

This model was developed in [1,2,9] and used in [6,7,20,23], the wave reaching any

point of the ultrasound sensor is the sum of several waves which are reflected from the

target surface. Waves arrive at a sensor point out of phase and interference among these

waves causes the granular aspect termed speckle. The mathematical expression for a

signal observed at point p, whose coordinates are x,y in the image, is as follows :

∑∑ −−= ),(),(),( jiji yyxxhyxeyxo (3)

Where e(x , y) : is the signal scattered from tissue and contributing to O(x,y), h is the

impulse response of the acquisition system. The intensity I(x ,y) at this point can be

stated in a multiplicative form as :

2 2( , ) ( , ) ( , ) ( , )I x y O x y e x y u x y= = (4)

Where u(x,y) is noise, independent from the useful signal. Within homogenous regions

this model offers a good approximation for ultrasound images.

Note that there exists an alternative model, as proposed in [29] and used in [30,31],

describing the speckle noise as an additive noise, with its amplitude proportional to

square root of the true image. However, this model was proposed to account for the

speckle pattern after a sequence of standard processing steps performed by a typical

ultrasound scanner.

6

II.2 Filtering Based Wavelet

II.2.1 Choice of wavelet Base

The choice of wavelet filter bases depends on the signal. Signals coming from different

sources have different characteristics.

For audio, speech, image and video signals the best choices of wavelet bases are

known. The best choice for ultrasound images is not clear. The problem is to represent

typical signals with a small number of convenient computable functions.

An investigation to choose the best wavelet bases for ultrasound images was performed

on a simulated ultrasound image (image accessible at http://telin.rug.ac.be/~sanja). The

majority of the wavelet bases which exist in the Matlab-7 software [32- 36] were tested.

In [37] a study about wavelet bases choice for image denoising was proposed using

PSNR (peak signal to noise ratio) as criterion for discrimination between bases. In [38]

a best wavelet bases selection for MRI (magnetic resonance imaging) was studied using

PSNR.

The criterion used to determine the best wavelet basis was the one which optimized the

PSNR in homogenous areas and preserved edges at the boundaries. The bi-orthogonal

wavelets basis [39-41] especially the 3.5 basis in Matlab-7 yielded the best average

PSNR among others. This base was chosen in this paper to be applied to ultrasound

image.

II.2.2 Wavelet decomposition

The principle of the wavelet decomposition is to transform the original raw particle

image into several components: one low-resolution component called “approximation”,

and the other components called “details” (Fig.1).

The approximation component is obtained after applying a bi-orthogonal low-pass

wavelet filter in each direction (horizontal and vertical) followed by a sub-sampling of

each image by a factor of 2 for each dimension. The details are obtained with the

application of a low-pass filter in one direction and a high-pass filter in the other, or a

high-pass filter in both directions. The noise is mainly present in the details

components. A higher level of decomposition is obtained by repeating the same

filtering operations on the approximation. D1i, (i varies from 1 to 6) are the details

(Fig. 1). This type of wavelet is a Fast Wavelet Transform (FWT) [36]. The FWT is not

the only type of wavelet transform used in the filtering process. Some used both a

redundant wavelet and reconstruction from wavelet maxima [33,16]. In the redundant

wavelet the scaling is not achieved by sub-sampling of the image in each step as in the

method presented above, but rather by a scaling of the filter.

II.2.3. Wavelet denoising by Soft-Thresholding

In [22], a simple thresholding procedure for the recovery of functions from noisy data

was proposed. It consists of three steps: the signal is transformed into an orthogonal

domain, using a discrete wavelet transform producing empirical wavelet coefficients.

The empirical wavelet coefficients are subjected to nonlinear soft-thresholding

ηt(y) = sign(y) (|y| − t), (5)

7

with a threshold

t = (2 log(n)).5σ, (6)

where σ is the standard deviation of the white noise and n is the data length. This

method is known as wavelet denoising, with uniform soft thresholding.

The processed wavelet coefficients are inversely transformed, supplying an estimation

of the actual image. As a result, denoised images generally are much less over-

smoothed, in comparison with the images denoised by, e.g., linear filtering. Note that

uniform soft thresholding is not the only way to suppress the wavelet coefficients of the

noise, and a multitude of methods have been proposed based, for example, on

principles of Bayesian estimation and detection theory [42]. In [22] another threshold

selection based on a decreasing factor for each level and noise STDV was used in

[20,21].

In this paper the first step in the processing was a log-transformation applied to the

original ultrasound image in order to transform the multiplicative noise to an additive

one. The method used in this paper differs slightly from the methods cited above in the

way of estimating the noise standard deviation. And the threshold chosen in this paper

is as follows:

Ts,l=(2/l)σs,l (7)

Where Ts,l is the threshold and σs,l is the standard deviation of the noise at the level l

and for the sub-band s.

The noise variance needed for definition of the threshold was estimated by assuming,

that most empirical wavelet coefficients at each detail level of the decomposition are

induced by the noise, and, thus, the median absolute deviation of wavelet coefficients at

this level accurately reflects the noise size [22]. In this paper the noise is chosen from

the detail components of the wavelet transformation, and also from each level. The

histogram of the wavelet coefficients plays an important role in estimating the noise

characteristics, because it is known from wavelet theory that this histogram is

composed of two parts details information (mainly edges and noise which is mainly

grouped around the center of the histogram because of their low gray value which is

due to their random distribution in the image). If the tails of the histogram are removed

which represent the useful details the remaining information in the histogram are

mostly noise and the calculation of the standard deviation from this remaining data are

a good approximation of the existing noise in each sub-band. In this paper the details

and outlier pixels which represent 30% of the total image pixels were removed, and the

standard deviation of the remaining data estimates the one of the noise. This estimation

is much easier and more robust than manually choosing few regions from the image

and calculates σ, also less dependent of the ultrasound scanner and the type of image

treated.

II.3 Filtering Based on Laplace-Gauss pyramid

The method described here for speckle reduction is based on multi-resolution contrast

enhancement. This method enables an adaptive approach regardless of the shape and

size of the homogenous area in the image. At low resolutions the homogenous area is

strongly smoothed (loss of small details). As the resolution increases, this method adds

8

meaningful details to the resulting image. Contrast enhancement can be applied on a

local scale. The pyramidal representation of an image enables different levels of

resolution to be obtained. A multi-scale representation of contrast is called a contrast

pyramid. This pyramid is built by using the Gaussian pyramid and the Laplacian one. A

detailed description of the construction of the contrast pyramid is presented in the next

paragraph.

II.3.1 Construction of the Gaussian pyramid

Because of the multiplicative nature of noise in ultrasound images a log-transformation is

realized as first step in order to transform the noise to the additive form. The construction

of the Gaussian pyramid is achieved by low pass Gaussian filtering followed by sub-

sampling by a factor of 2 in each direction, horizontal and vertical.

The 2D Gaussian filter used here is separable (the same filter is used for both directions).

The filter used here has the following digital impulse response: [1/16, 4/16, 6/16, 4/16,

1/16].

II.3.2. Construction of the Contrast Pyramid

Contrast at any point p in the image can be defined at a given resolution level as follows

[6]:

( ) ( )( )

( ) ( )

G p V pC p

G p V p

−=

+

(8)

Where G(p) is the gray level at point p and V(p) is the measure of the variation of the

gray level of the area surrounding p. This definition enables us to define a multi-

resolution contrast scheme by approximating V(p) by the Laplacian L(p) at point p. The

new contrast definition at level k in the pyramid for a given point p becomes:

( ) ( )( )

( ) ( )

k kk

k k

G p L pC p

G p L p

−=

+ (9)

With k =K-1,….,0 where K is the highest level of the pyramid in our application K=3.

The Laplacian Lk(p) is approximated as the difference between Gk(p) and a smoothed

Gk(p).

II.3.3 Contrast Enhancement

Contrast enhancement involves smoothing contrast in homogenous areas and enhancing

it at edges.

Non linear functions were used and shown to give good results. An adaptive median

filter of different sizes is applied to modify the contrast before the reconstruction.

9

II.3.4 Reconstruction of the Filtered Image

Reconstruction of the visually enhanced image is performed from the top of the

pyramid to the bottom of it. The idea of the reconstruction is based on the following

expression:

Gk(p)=Ck(p) expand[G k+1(p)] (10)

Where expand is an operation which doubles the size of the columns and the rows by

inserting first, lines and columns of zeros and then applying a low pass Gaussian filter.

In the case where the contrast is modified the expression becomes:

Gmk(p)=Cmk(p) expand[Gmk+1(p)] where Cm represents the modified contrast and Gm

represents the modified Gaussian pyramid. For the highest level of the pyramid

GmK=GK is assumed, and Gm0 is the resulting filtered image.

II.4 Synthetic image reconstruction

In order to investigate the quantitative performance of the selected methods, an image

with artificial speckle noise was employed. A synthetic speckled image used in [19]

and modeled as in [18,43] d = f.v is commonly used where f is a reference noise-free

image, and v is a unit mean random field. Realistic spatially correlated speckle noise

vk in ultrasound images can be simulated by low-pass filtering a complex Gaussian

random field and taking the magnitude of the filtered output [43]. The low-pass

filtering was realized by averaging the complex values in a 3x3 sliding window. Such a

short-term correlation was found sufficient [17] to model the realistic images well. By

changing the variance of the underlying complex Gaussian random field, an image with

different levels of speckle noise was generated. a purely synthetic image in Fig. 2,

which consists of regions with uniform intensity, sharp edges, and strong scatters. As a

quantitative performance measure, a signal to noise ratio in dB was used – defined as

SNR = 10 log(Ps/Pn), (11)

Where Ps is the variance of the noise-free reference image, and Pn is the noise

variance.

III. Results

In figure 2a a synthetic image without any noise is presented. In 2b a generated noise

with SNR of 8.3 db is shown. In 2c the result obtained with WST filter with proposed

thresolding an SNR of 15.8 db was obtained from this image. In 2d the result from

LGCE method the SNR calculated from this image is 11.3 dB. Another criterion which

was evaluated is the absolute difference (error) between the noise free image and the

other three. This error image (map) shows us the location of error in the filtered image,

if the resulting error map is all black (zeros) the implication is that the filtering is ideal.

The smallest error map indicates best filtering. The image in 2.e shows resulting image

of the difference between noisy and noise-free image, this difference shows the error

image or the noise distribution and intensity.

10

Figure 2.f shows error map between the filtered image using SWT method and the

noise free one. Figure 2.g shows the error map between the LGCE filtered image and

the noise-free one. The table 1 summarizes the quantitative measure used for all three

error maps.

Table 1: quantitative criteria used for phantom image.

criteria Noisy-

image

WST LPGC

MSE 53 24 38

SNR (dB) 8.3 15.8 11.3

The methods described in the methods section were applied on some ultrasound images

with some selected areas of interest. First, a WST filtering was applied on six

ultrasound images, an example of real ultrasound image is shown in figure 3. Wavelet

decomposition of level two with biorthogonal 3.5 basis was used during enhancement

process.

A soft-thresholding of two standard deviations(7) was applied on the first level of

detail components and one SD on the second level of components details, the resulting

image is shown in figure 4.

Figure 5 shows a filtered image using a LGCE method. A two level contrast pyramid

was used. A median filter of size 5x5 was applied on the first level of contrast and a

median filter of size 3x3 at the second level. Then, subsequently, a reconstruction was

used in order to obtain the enhanced image. It is difficult for non-expert in the

diagnoses of ultrasound images to notice the difference between the two enhanced

images. For the purpose of improving the visual differences between the two images,

two things were necessary. First a difference between the filtered image and the

original one was accomplished, Figure 6 shows the difference between the original

image and the image filtered with the LGCE method. All the values in this image are

multiplied by a factor of 10 in order to improve the perception of errors. Similarly, the

difference for the WST method with the original (unfiltered) is shown in figure 7.

Next, In order to focus the study on one part of the image, a region of interest was

selected manually by an expert in ultrasound diagnosis for both filtered images. Figure

4 shows the selected white boundary of the area in the WST method. A calculation of

the standard deviation inside some ROI from 5 images gave an average value of 10.3.

The same area in the image enhanced by LGCE method using median filters is shown

in figure 5 encircled by white boundary. A calculation of the standard deviation inside

the same ROI from the same 5 images gave an average value of 18.7

IV. Discussions

A recent study used the WST method in ultrasound imaging of carotid artery[21]. In

this study (restricted to this specific data type) filtering using local statistics and

geometric filtering gives better results than wavelet filtering. This is maybe due to the

specific nature of data or other unknown causes, such as the choice of wavelet basis or

thresholding techniques used. In [17] the superiority of LGCE contrast enhancement

over wavelet filtering (on X-ray images using a specific non-linear modification of the

wavelet coefficient) is demonstrated. The wavelet filtering used in [17] is different than

11

soft-thresholding techniques this may explain the result. A large amount of work in

despeckling ultrasound images [7,19,20] show the superiority of wavelet techniques

over others methods except the LGCE method which was not compared to WST

filtering on ultrasound images. Recently in [7] a study on ultrasound images confirmed

the superiority of WST over some other well known multiplicative noise filtering

methods. Here in this paper the study focus on comparing SWT and LGCE on

ultrasound images. It is very difficult to compare two methods for removing speckle

from real ultrasound images, because it dependent of what we need from such an

image. For small details it is not obvious to a non-expert in the diagnoses of ultrasound

images to know what we need to eliminate or to preserve and enhance. Some of the

statistical measurements such as SNR or RMSE (root mean square error) are not very

significant in the case of real ultrasound images if they are applied to the whole image

without a prior knowledge of the image. Such measurements could be efficient in the

case of simulated images.

In this paper a region of interest was chosen inside the real image in order to aid in the

assessment of filtering for the purpose of visual quality assessment. The best filtering

should give the best homogeneity inside the region; the internal variation should be the

minimum, while preserving the boundary. Another way to analyze the effect of filters is

to study the error image between the original image and the filtered one. The error

image is calculated here by the absolute difference between the two images. The

distribution of information inside the error image provides a qualitative indication about

the filtering method. If the information is randomly distributed, generally similar to

speckle distribution, this indicates that the extracted information from the original

image is non-structured and may be considered to be mainly noise. If there is some

structured information present in the error image, this indicates that some structured

information (boundaries or edges) is higher in one image than other. If the difference

between the original image and the filtered one proves that edges in the original are

higher in amplitude than the filtered one, this indicates that a smoothing of the edges

was obtained. If the opposite case happens, the indication is that a boosting of the edges

is present.

The error image concerning WST in figure 7, shows clearly that a random distribution

of information is present in this image, which is very similar to speckle distribution.

This error image presents a map distribution on every pixel in the image which much

better representation than mean square error or other, generally calculated on the whole

image, which is of quite small interest in this type of highly noisy data. This type of

error representation proves visually that the edges neither amplified nor damped and the

random distribution of noise shows that the filtering reduces mainly the speckle.

The error image in figure 6, concerning the LGCE, shows that some structures (edges)

and non-structures (mainly noise) information are amplified in this image. This is due

to the boosting of local strong variation without taking into account the directionality as

in wavelets. This qualitative assessment shows that WST is better adapted to ultrasound

images.

Another criterion used in order to assess the quality of enhancement is the standard

deviation inside the selected ROI. This criterion shows that the variation inside the ROI

is lower in the WST enhanced image than the variation in the LGCE enhanced image.

Furthermore, the quality of contours is better in the wavelet method as experts notice.

The results show that enhancement based on wavelets are superior to those based on the

LGCE.

12

V. Conclusions In this paper two multi-scale methods were tested and compared for enhancing visual

quality of ultrasound images. Those methods were proven to be very efficient to treat

multiplicative noise and reduce speckle in ultrasound and in SAR images. Several

studies prove that WST was among the best methods for ultrasound images. The LGCE

method using median filtering was applied recently on ultrasound images and proves to

give very good results. To date, WST and LGCE have not been compared to each other

using ultrasound images where a lot of variation exists in the background. The overall

quantitative assessment in terms of homogeneity inside the ROI, contour preservation

and error distribution indicated that WST provided a better overall improvement for

visual display.

A qualitative assessment of few ultrasound images performed by a diagnosis expert

(using some areas of interest) confirms that the WST method shows a much better

visual enhancement than LGCE one. This conclusion demonstrates that results

obtained in ultrasound images are not valid for X-ray and vice versa.

This study proves that wavelet filtering using soft-thresholding is an excellent tool for

visual enhancement of digital ultrasound images, which certainly helps for better

diagnosis. Further work is needed to evaluate the suggested method as a preprocessing

step to unsupervised segmentation of ultrasound images.

References

1.Godman J.W. (1976) Some fundamental properties of speckle, JOSA, 66, 11.

2. Tur and al. (1982) When Speckle is multiplicative, Applied Optics vol. 21, No. 7.

3a. Hillery D. et al. (1991) Iterative wiener filters for images restoration. IEEE Transaction on Signal

Processing, 39, pp.1892-1899.

4. Frost VS et al. (1982) A model for radar images and its application to adaptive digital filtering of

multiplicative noise. IEEE Transaction Pattern Analysis and Machine Intelligence, 4-2. pp. 157-166.

5.Unser M and Aldroubi (1996) A review of wavelets in biomedical applications. Proceeding IEEE, 84,

pp. 626-638.

6. Saad A , Elassad S. Barba D. (1996) Speckle filtering in SAR images by contrast modification,

comparison with a large class of filters. Annals de telecommunication, 51, 5, pp. 233-244.

7. Michailovich O., Tannenbaum A., (2006) Despeckling of ultrasound images, IEEE Trans.

Ultrasounics Ferroelectrics and Frequency Control, Vol. 53, no. 1 pp. 64-78.

8. Frost V.S, (1982) A model for radar images and its application to adaptive digital filtering of

multiplicative noise, IEEE Trans. On PAMI Vol. 4, No. 2.

9.Touzi R. , Analyse d'image radar en teledetection: Amelioration radiometriques filtrage de speckle et

detections des contours, Ph.D Thesis University Paul Sabatier Toulouse France Mars 1988.

10. J. S. Lee (1981) Speckle analysis and filtering smoothing of synthetic aperture radar image",

computer Graphics and image processing, No. 17, pp24-32,

11. J. S. Lee, (1980), Digital image enhancement and noise filtering by use of local statistics, IEEE

Trans. On PAMI Vol. 2, No 2,pp 165-168 .

12. J. S. Lee, (1983), A simple smoothing algorithm for SAR images, IEEE Trans. On Systems, Man

Cybernetics, Vol. SMC-13, No. 1. pp. 85-89.

13. P. Zamperoni, (1993), An adaptive rank order filter for image enhancement depending on a measure

of local spatial order, GRETSI , France, pp. 543-547.

14 . Mallat S and Zhang S. (1992) Characterization of signals from multiscale edges. IEEE Transaction

on Pattern Analysis Machine Intelligence, 14, 710-732.

13

15. Vuylsteke P and Schoeters E (1994) Multiscale image contrast amplification (MUSICATM).

Proceeding SPIE image processing, vol. 2167, pp.551-560.

16. Laine A, Fan J. and Yang W. (1995) Wavelets for contrast enhancement of digital mammography.

IEEE Engineering on Medical Biology and Magnetism, 14, pp.536-574.

17. Dippel S, Stahl M, Wiemker R, Blaffert T. (2002) Multiscale contrast Enhancement for

radiographies: Lapacian pyramid versus Fast wavelet Transform. IEEE Transaction on Medical Imaging

21, 4, pp: 343-353.

18. Achim A, Bezerianos A, and Tsakalides P. (2001) Novel Bayesian Multiscale Method for Speckle

Removal in Medical Ultrasound Images. IEEE Transaction on Medical Imaging, 20, 8.

19. Pizurica A, Philips W, Lemahieu I, and Acheroy M (2003) A Versatile Wavelet Domain Noise

Filtration Technique for Medical Imaging. IEEE Transaction on Medical Imaging, 22, 3, pp.323–331.

20. Zong X., Laine A., and Geiser E. (1998) Speckle Reduction and Contrast Enhancement of

echocardiograms via Multiscale Nonlinear Processing, Transaction on Medical Imaging.,17, 4, pp.532-

540.

21. Loizou CP, Pattichis CS, Christodoulou CI, Istepanian RSH, Pantziaris M, Nicolaides A (2005)

Comparative evaluation of despeckle filtering in ultrasound imaging of the carotid artery. IEEE Trans

Ultras Ferroel Freq Control 52(10):653–669.

22a. Donoho D. (1995) De-noising by soft thresholding. IEEE Transcation on Information Theory, 41,

pp. 613-621.

23. Saad A, Zaart A. and Al-Mejrad A., (2004) Speckle Reduction in Digital Ultrasound Images by

Multi-Resolution Contrast Enhancement. International Conference on Biomedical Engineering 2004,

IFMBE, BioMed, Kuala Lumpur, Malaysia.

24. Wang Z, Bovik A, Sheikh H, Simoncelli E (2004) Image quality assessment: from error measurement

to structural similarity. IEEE Trans Imag Process 13(4):600–612.

25. C.P. Loizou, C.S. Pattichis, M. Pantziaris, T. Tyllis, A. Nicolaides, (2006) Quantitative quality

evaluation of ultrasound imaging in the carotid artery," Med. Biol. Eng.& Comput., vol. 44, no. 5, pp.

414-426.

26. Jain A. K., (1989) Fundamental of Digital Image Processing. Englewood Cliffs, NJ: Prentice-Hall.

27. Zong X., A. F. Laine, and E. A. Geiser, (1998)Speckle reduction and contrast enhancement of

echocardiograms via multiscale nonlinear processing,” IEEE Trans. Med. Imag., vol. 17, pp. 532–540.

28. Achim A., Bezerianos A., and Tsakalides P., (2001) Novel Bayesian multiscale method for speckle

removal in medical ultrasound images,” IEEE Trans. Med. Imag., vol. 20, pp. 772–783.

29. Loupas T, McDicken WN, Allan PL (1989) An adaptive weighted median filter for speckle

suppression in medical ultrasonic images. IEEE Trans Circuits Syst 36:129–135.

30. Hao X., Gao S., and Gao X., (1999) A novel multiscale nonlinear thresholding method for ultrasound

speckle suppressing,” IEEE Trans. Med. Imag., vol. 18, pp. 787–794.

31. A. Rakotomamonjy, P. Deforge, and P. Marche, (2000) Wavelet-based speckle noise reduction in

ultrasound B-scan images,” Ultrason. Imag., vol. 22, pp. 73–94.

32. Strang G, Nguyen T (1997) Wavelets and Filter Banks. Wellesley - Cambridge Press, UK.

33.Meyer Y (1993) Wavelets and operators. Cambridge press, UK.

34. Daubechies I. (1992), Ten lectures on wavelets: SIAM,.

35.Vetterli M. and Herely C., (1992) Wavelets and filter banks: theory and design, Transactions on

signal processing, vol. 40, pp. 2207-2232.

36. Beylkin G., Coifman R., and Rokhlin V. (1991) Fast Wavelet transforms and numerical algorithms,

Communication of Pure Application. Math., vol. 44, pp. 141-183.

37. F. Luisiera, T. Blua, B. Forsterb and M, Unser, (2005) Which wavelet bases are the best for image

denoising ?, Proceedings of the SPIE Optics and Photonics, Conference on Mathematical Methods:

Wavelet XI, San Diego CA, USA, July 31-August 3, , vol. 5914, pp. 1-12.

38. Abu-Rezq A.N, Tolba A. S, Khuwaja G A, S.G. Foda (1999) Best Parameters selection for Wavelet-

packet-Based compression of Magnetic Resonance Images , Computer and biomedical research 32, pp:

449-469.

39. Cohen A, Daubechies I, Feauveau JC (1992) Bi-orthogonal Bases of Compactly supported Wavelets.

Communication on Pure Application Mathematic, 45, pp. 485-560.

40. Phoong S-M, Kim CW, Vaidyanathan PP and Ansari R. (1997) A new class of two-channel bi-

orthogonal filter banks and wavelets bases. IEEE Trans. on Signal Processing, 43, pp. 649-665.

41. Turcajova J. and Kautsky R. (1995) Discrete Biorthogonal Wavelet transform as block circulant

matrices, Linear Algebra and its Application, vol. 223, pp. 393-413.

14

42. Muller P.and Vidakovic B., (1999) Eds. Bayesian Inference in Wavelet Based Models, Lecture

Notes in Statistics. vol. 141, New York: Springer-Verlag,.

43. F. Sattar, L. Floreby, G. Salomonsson and B. Lovstrom, (1997) Image enhancement based on a

nonlinear multiscale method,” IEEE Trans. Image Proc., vol. 6, no. 6, pp. 888–895.

Figure 1. An example showing wavelet decomposition of level 2 using bi-orthogonal bases.

Figure 2. Synthetic image used to show the quality of filters used in this study.

2a. a noise free synthetic image.

2b. synthetic image after noise multiplication.

2c. filtering of 2b by WST method.

2d. filtering of 2b by LGCE method.

2e. error map between the noisy image and the noise free one.

2f. shows the error map between 2a and 2c.

2g. shows the error map between 2a and 2d.

Figure 3. Example of real ultrasound image used for processing.

Figure 4. Result of the wavelet soft thresholding method applied to a

Real ultrasound image.

Figure 5. Result of the multi-resolution contrast enhancement method applied to a real ultrasound image.

Figure 6. This figure shows the image difference between Laplace-Gauss contrast image and the

original.

Figure 7. This figure shows the image difference between the wavelet

filtered image and the original.