Beruflich Dokumente
Kultur Dokumente
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
enhancement,
I. INTRODUCTION
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
ultrasound images cleaner with clearer boundaries, but the
images suffer from degraded spatial resolution and increased
system complexity. In contrast, rather than compounding
multiple sub-images, in SR-PF, a specially designed spatial
filter is applied to a single image [22-32]. The SR-PF method
can be classified as a single-scale linear or nonlinear spatial
filtering method and as a spatial filtering method based on
multiscale analysis (i.e., SR-PFSC and SR-PFMS, respectively).
In SR-PFSC, various spatial filters, e.g., smoothing [22],
Wiener [23], unsharp masking [24], region growing [25-27],
slope-facet modeling [28] and the stochastic approach [29],
have been proposed to reduce speckle. Though these spatial
filters can indeed reduce speckle, they suffer from excessive
suppression, blurred boundaries and obscured details. To more
effectively suppress speckle, diffusion-based spatial filtering
methods utilizing partial difference equations (PDE) have been
proposed.
In the speckle reducing anisotropic diffusion (SRAD)
method [30], an edge-sensitive anisotropic diffusion process
that utilizes the instantaneous coefficient of variation, which is
a function of the local gradient magnitude and Laplacian
operators, is applied. Similarly, an advanced SRAD (OSRAD)
method has been proposed, in which directional filtering of the
image along structures is performed by a diffusion matrix to
preserve boundaries [31]. Similarly, in nonlinear coherent
diffusion (NCD), three different diffusion filters are
dynamically applied, depending on the various speckle regions
[32]. The OSRAD and NCD methods show enhanced speckle
reduction and edge perseverance, but they still have limitations
in retaining fine features such as small lesions or distinguishing
small structures from tiny cysts.
Such nonlinear iterative diffusion methods based on gradient
operators have difficulty in accurately differentiating signal and
noise. Thus, a multi-scale analysis approach, which has the
potential to provide effective signal and noise separation, can
be utilized for speckle reduction, i.e., SR-PFMS. The
wavelet-based speckle reduction method, in which soft
thresholding of the logarithmically transformed image is used,
has been proposed [33]. However, it is difficult to determine a
suitable threshold level for the wavelet coefficients. The
recently proposed nonlinear diffusion filtering and edge
enhancement method based on dyadic wavelet transform
(NMWD) shows favorable denoising properties. These
properties are supported by the integration of nonlinear
diffusion and wavelet shrinkage, but the approach is lacking in
terms of its ability to preserve edges and structural similarities
[34].
In the same manner, the Laplacian pyramid transform
method combined with nonlinear diffusion (LPND) has been
proposed. In the pyramid domain, a signal is decomposed into
decimated lowpass and bandpass signals on a finer scale and is
then reconstructed into a single interpolated signal from a
coarse scale [35]. The pyramid decomposition and
reconstruction of a two dimensional (2D) image has the
advantage of not requiring a quadrature mirror filter (QMF)
pair, as the wavelet-transform methods do [33, 34]. However,
although the LPND method successfully suppresses speckle, it
2
cannot preserve the edges and boundaries of lesions. To
compensate for the blurring caused by the LPND method, a
shock filter was adopted, i.e., the Laplacian pyramid-based
nonlinear diffusion and shock filter (LPNDSF) [36]. The
LPNDSF method utilizes nonlinear diffusion and a shock filter
in a coupled partial differential equation (PDE) process in the
Laplacian pyramid domain. However, the edges and structures
of lesions in the processed image tend to appear too artificial or
unnatural for use in ultrasound imaging.
In addition, diverse non-local patch based algorithms with a
Bayesian framework have recently shown competitive
performance [37-39]. Among these methods, the optimized
Bayesian non-local means with blockwise approach (OBNLM)
method [37] is performed to reduce the complexity burden
while suppressing speckle noise and preserving structural
details. Nevertheless, its patch comparison procedure for
selecting a relevant criterion still leads to high computational
complexity.
In this paper, a new feature-enhanced speckle reduction
(FESR) method based on multiscale analysis and feature
enhancement filtering is proposed to more effectively reduce
speckle noise while preserving tissue structures and enhancing
low echogenicity with improved border definition and
continuity. The performance of the proposed FESR method is
quantitatively compared with other speckle reduction methods
(i.e., OSRAD, NMWD, LPNDSF and OBNLM) in terms of
edge preservation, mean structure similarity, speckle
signal-to-noise ratio (SSNR) and contrast-noise ratio (CNR).
Moreover, a qualitative analysis conducted by three
experienced radiologists that demonstrates their diagnostic
preferences is presented in this paper.
This paper is structured as follows. In Section II, a speckle
model for medical ultrasound imaging is presented. In Section
III, the nonlinear anisotropic diffusion and pyramid transform
that are the foundation of the proposed FESR method are
reviewed. In Section IV, the proposed FESR method is detailed.
In Section V, the experimental setup and evaluation metrics are
presented. In Section VI, the experimental results are discussed
and compared with other speckle reduction methods in
simulation, phantom, and in vivo studies with both quantitative
and qualitative evaluation. Section VII concludes the paper.
II. SPECKLE MODELING
Variations in tissue attenuation, propagation and scattering
properties cause ultrasound echoes to interfere in complex
ways. Echoes from reflectors spaced closer than the resolution
limit will interfere, thus creating artificially large (i.e.,
constructive interference) or small (destructive interference)
signals. Therefore, constructive and destructive coherent
interferences from backscatterer echoes that are typically much
smaller than the wavelength of an ultrasound incident wave
create phase summations or phase cancellations. Thus, the
speckle pattern can be defined as a multiplicative noise, rather
than a type of normal additive noise. The speckle noise model
for a medical ultrasound image can be approximated as
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
( x, y) f ( x, y)wm ( x, y) wa ( x, y)
(1)
( x, y) f ( x, y)wm ( x, y)
(2)
(3)
t
I (t 0) I 0
(4)
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
G0 (n) I (n)
Gl (n) R Gl 1 (n)
(5)
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
x,
0,
otherwise
(6)
ut sign u u , u x, y, t | t 0
(8)
(9)
' ( x) c( x) x .
to increase robustness.
u
u
div | u | | G u |
t
| u |
(7)
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
(10)
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
K x, y (2 2 ) 1 exp(
x2 y 2
)
2 2
(11)
J (u ) (1 | 2 ) 1
0
0 1T
2 2T
(12)
( )2
H 1 exp c 2
(13)
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
The FESR method utilizes (6) to suppress speckle noise, and
the scale factor k was set to 1.95. The edge preservation
parameter e in (9) and coherence enhancing factors c in (11)
and c in (12) were set as described in Table I. The illumination
coefficient and reflectance coefficient for contrast
enhancement were automatically estimated as shown in Fig. 6.
All of the methods were implemented in the Matlab
(Mathworks Inc., Natick, MA, USA) environment, and the
computational time in Table I was measured in simulation
studies for a frame of the echogenicity map of 460 447 pixels.
A. Simulation study
To generate B-mode ultrasound RF data, we conducted
pseudo B-mode pulse-echo image generation as reported in [30,
51]. Based on the convolution of the point spread function and
tissue scattering function, we observed the generated speckle
image changes in relation to the axial sound pulse width x and
the lateral beam width y , as illustrated in Fig. 7.
7
B. Phantom study
To apply the five speckle reduction methods to the phantom
experiment, we scanned a commercial gray scale phantom (i.e.,
040 GSE, CIRS INC., Norfolk, VA, USA) with -5 dB/MHz
attenuation and obtained RF beamformed data containing 30
consecutive frames using a commercial ultrasound system (i.e.,
Accuvix V10, Samsung Medison., Korea) with a L5-12 linear
array probe. The RF data of the 30 frames were enveloped and
logarithmically compressed in the post-processing stage to
simulate the B-mode images.
C. In vivo study
The original images of 28 consecutive frames were captured
by a commercial ultrasound system (iU22 xMATRIX, Philips
Healthcare, Bothell, WA, USA) using a C5-1 convex probe and
a L9-3 linear array probe for the abdominal and thyroid areas of
a volunteer. To acquire the original data, we deactivated the
real-time compounding, time-averaging (persistence) and
speckle reduction filtering modes at that time. Then, the images
were transferred onto an external PC and processed offline.
Quantitative and qualitative assessments were performed to
evaluate the improvement in image quality and clinical
feasibility of the five speckle reduction methods.
D. Evaluation metrics
In the simulation study, the visibility errors between a
distorted image and a reference image, as shown in Fig. 10,
were quantified by measuring the speckle regions
signal-to-noise ratio (SSNR), contrast-to-noise ratio (CNR),
edge preservation factor , and mean structural similarity
quality measurement (MSSIM). The SSNR was measured in a
fully formed speckle region (in a homogeneous region) without
any relative structures and can be defined by
y 0.5 ~ 4.5 .
SSNR
(14)
| 1 2 |
(15)
12 2 2
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
(I I , I I)
(16)
(I I , I I ) (I I, I I)
M N
(( I (i, j ) I (i, j ))
i , j 1
(17)
1
M
(2 x y C1 )(2 xy C2 )
i 1
2
x
y2 C1 )( x2 y2 C2 )
(18)
S,D
1
(u i, j, t u (i, j , t 1)) 2
S D i , j 1
(19)
Fig. 10. Relative intensity profiles of a nonlinear filter from noisy input signal
for the result of (a) 1, 20, 100 iterations on single scale and (b) 2, 3, 4 and 5
scales with an iteration.
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
a diffusion filtering threshold is selected manually, as in [34].
In contrast, iteration and multiscale processing are
performed in the same direction in terms of local operations,
leading to global effects [54]. Figure 10 demonstrates that the
four scales were sufficient to produce an effective result while
100 iterations of a single scale were needed. Thus, multiscale
iterated filtering maintains the high degree of nonlinearity and
adaptability found in the iterated method while simultaneously
using a smaller number of iterations on a single scale. As such,
the decreased computational complexity and cost of the
reduced number of iterations offers substantial advantages in
real-time systems.
VI. EXPERIMENTAL RESULTS AND DISCUSSION
A. Simulation study
9
Table I, its computational time shows its complexity burden
compared with the other methods.
The SSNR, CNR, , and MSSIM values were measured for
30 simulated images with the same speckle level, with the mean
and standard deviation of the measurement values summarized
in Table II. As shown in Table II, all of the speckle reduction
methods are observed to improve the results in terms of all
parameters. Consistent with the visual assessments, all of the
parameter values were higher in the proposed FESR method
than in other speckle reduction methods, i.e., 23.362.55,
1.700.15, 0.160.01 and 0.570.01. Furthermore, the SSNR
value from the FESR method was considerably higher than for
all other methods, i.e., 23.362.55 vs. 18.311.76, 15.521.62,
17.481.55 and 21.423.86, whereas the obtained values of
CNR, and MSSIM were comparable. It must be noted that
the OBNLM method outperforms the NCD, NMWD, and
LPND methods.
SSNR, CNR,
TABLE II
and MSSIM values of 30 simulated images with the same
speckle level
Method
Fig. 11. Simulated pseudo B-mode image and filtered results. (a) Speckled
noisy image. (b)-(f) Images filtered by OSRAD, NMWD, LPNDSF, OBNLM
and FESR, respectively.
Parameters
SSNR
CNR
MSSIM
Noisy
8.190.79
1.020.08
0.040.00
0.270.01
OSRAD
18.311.76
1.610.13
0.140.01
0.530.01
NMWD
15.521.62
1.570.13
0.120.00
0.490.01
LPNDSF
17.481.55
1.590.14
0.140.01
0.550.01
OBNLM
21.423.86
1.630.15
0.150.01
0.560.01
FESR
23.362.55
1.700.15
0.160.01
0.570.01
Table III
CNR values for phantom images containing 30 consecutive frames
Method
CNR
ROI 1
ROI 2
ROI 3
ROI 4
Noisy
6.310.03
0.430.00
2.300.01
3.160.01
OSRAD
9.000.06
0.690.01
3.480.01
4.510.02
NMWD
9.780.06
0.680.01
3.610.01
4.580.02
LPNDSF
8.670.04
0.860.01
3.540.02
4.530.01
OBNLM
9.220.06
0.880.01
3.900.01
5.170.01
FESR
10.700.06
0.990.01
4.250.02
5.490.03
B. Phantom study
Fig. 12. Phantom image and its filtered results. (a) Original noisy image. (b)-(f)
Image filtered by OSRAD, NMWD, LPNDSF, OBNLM and FESR methods.
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
10
Fig. 13. Ultrasound in vivo abdominal liver image and its filtered results. (a) Conventional noisy image. (b)-(f) Image filtered by OSRAD, NMWD, LPNDSF,
OBNLM and FESR, respectively.
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
the FESR method shows the highest CNR values compared
with the other four speckle reduction methods for all four ROIs,
i.e., 10.700.06, 0.990.01, 4.250.02 and 5.490.03,
respectively. Among the remaining four methods, the OBNLM
method shows the highest CNR values for all ROIs.
C. In vivo study
Figure 13(a) shows the original ultrasound B-mode image
that was acquired from the liver of a volunteer, and the
processed images using the five speckle reduction methods are
shown in Figs. 13(b)-13(f). The second row of Fig. 13
represents zoom-in images of a solid box in Fig. 13(a). Under
visual assessment, the FESR method shows uniform reduction
in speckle in the liver parenchyma areas with improved tissue
differentiation. In addition, the borders and boundaries of the
hepatic vessels (i.e., A in Fig. 13(a)) in the FESR image are
more clearly depicted than in the images processed by the other
speckle reduction methods. Additionally, the FESR method
significantly suppresses acoustic clutter, which produces
spurious echoes within the gall bladder (i.e., region B in Fig.
13(a)) near the liver. The other cystic contents and tissue
margins in the FESR image are shown to be a clearer
visualization without a loss of underlying diagnostic details
(e.g., the region C of Fig. 13(a)).
In contrast, the NMWD method shows higher speckle
reduction but produces blurred edges and boundaries in the
hepatic vessels, similar to the results found in the simulation
and phantom experiments. The LPNDSF and OBNLM
methods also suffer from artificial speckle patterns in the liver
parenchyma areas, although they preserve local coherent edges.
The OSRAD method substantially suppresses speckle noise
and maintains the structures of the hepatic vessel and the gall
bladder. However, its image suffers from a faded appearance.
11
standard deviations of the OBNLM and FESR methods are
greater than those for the OSRAD, NMWD and LPNDSF
methods because a relatively large speckle pattern can also be
processed as texture and can thus be erroneously enhanced by
the feature enhancement filtering used in the FESR method. In
addition, the rejected outlier is shown in all five results due to
the presence of unwanted signals, which was emphasized in the
filtering process.
In the case of the abdominal kidney, as represented by Fig.
15, FESR substantially reduces the appearance of speckle and
increases the conspicuity of low-contrast tissue compared with
the other speckle reduction methods. Additionally, the FESR
image of the thyroid case shows an improved delineation of
linear striation patterns from fat and muscle with continuity, as
shown in Fig. 16. Furthermore, the internal architecture of the
benign nodule is depicted with clear visualization and thus
provides better detectability.
The subjective preference scores for five speckle reduction
methods given by the radiologists are shown in Table IV. For
the overall parameters, the three radiologists mostly indicated a
score of better for the FESR method in three different cases,
which were compared with worse or normal for other
methods. In particular, the abdominal liver obtained
significantly different scores in the FESR method compared to
other methods, as shown in Table IV.
For margin sharpness, the FESR method shows no
statistically significant difference for the abdominal kidney (i.e.,
p=0.065) and thyroid (i.e., p=1) compared with the OBNLM
method, though all three radiologists rated it with a higher than
average score. For artificiality, the radiologists scored the
LPNDSF and OBNLM methods relatively low due to the
unnatural speckle patterns, whereas they showed considerable
results for quantitative assessment. Additionally, FESR did not
show statistical significance for thyroid images compared with
the OSRAD method. For tissue contrast, all three radiologists
indicated a score of better for the FESR method in three
different cases, which were compared with worse or normal
for the other methods. It must be noted that, in the present study,
inter- or intra-rater agreement between radiologists was not
evaluated.
VII. CONCLUSION
Fig. 14. The box plots show the distribution of SSNR and CNR in the
conventional noisy image and the filtered images processed by the OSRAD,
NMWD, LPNDSF, OBNLM and FESR methods. (a) SSNR values. (b) CNR
values. The * represents outliers.
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
12
Fig. 15. Ultrasound image of in vivo abdominal kidney and its filtered results. (a) Conventional noisy image. (b)-(f) Image filtered by OSRAD, NMWD, LPNDSF,
OBNLM and FESR methods, respectively.
Fig. 16. Ultrasound image of in vivo abdominal kidney and its filtered results. (a) Conventional noisy image. (b)-(f) Image filtered by OSRAD, NMWD, LPNDSF,
OBNLM and FESR methods, respectively.
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
13
Table IV
Subjective preference scores of five speckle reduction methods in three different tissues
[3]
REFERENCES
[1]
[2]
[4]
[5]
[6]
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TBME.2015.2486042, IEEE Transactions on Biomedical Engineering
> TBME-00795-2015.R1<
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
14
[34] Y. Yue et al., Nonlinear multiscale wavelet diffusion for speckle
suppression and edge enhancement in ultrasound images, IEEE Trans.
Med. Imag., vol. 25, no. 3, pp. 297-311, 2006.
[35] F. Zhang et al., Nonlinear diffusion in laplacian pyramid domain for
ultrasonic speckle reduction, IEEE Trans. Med. Imag., vol. 26, no. 2, pp.
200-211, 2007.
[36] F. Zhang et al., Multiscale nonlinear diffusion and shock filter for
ultrasound image enhancement, in Proc. 2006 IEEE Computer Society.
Computer Vision and Pattern Recognition, pp. 1972-1977, 2006.
[37] P. Coupe et al., Nonlocal means-based speckle filtering for ultrasound
images, IEEE Trans. Image Process., vol. 18, no. 10, pp. 2221-2229,
2009.
[38] Y. Gu et al., Ultrasound echocardiography despeckling with non-local
means time series filter, Neurocomputing, vol. 124, pp. 120-130, 2013.
[39] Y. Zhan et al., Nonlocal means method using weight refining for
despeckling of ultrasound images, Signal Processing, vol. 103, pp.
201-203, 2014.
[40] V. Dutt, Statistical analysis of ultrasound echo envelope, Ph. D.
dissertation, Mayo Graduate School, Rochester, MN, 1995.
[41] P. Perona and J. Malik, Scale-space and edge detection using anisotropic
diffusion, IEEE Trans. on Pattern Anal. Mach. Intell., vol. 12, no. 7, pp.
629-639, Jul 1990.
[42] P. J. Burt, and E. H. Adelson, The Laplacian pyramid as a compact
image code, IEEE Trans. on Commun., vol. COM-31, No. 4, Apr 1983.
[43] F. Sattar et al., Image enhancement based on a nonlinear multiscale
method, IEEE Trans. Image Process., vol. 6, no. 6, pp. 888-895, 1997.
[44] M. J. Black et al., Robust anisotropic diffusion, IEEE Trans. Image
Process., vol. 7, pp. 421-432, Mar 1998.
[45] R. W. Brockett and P. Maragos, Evolution equations for
continuous-scale morphology, in Proc. IEEE ICASSP, pp. 125-128,
1992.
[46] J. Weickert, Multiscale texture enhancement, Computer analysis of
images and patterns, Lecture notes in Computer science, vol. 970, pp.
230-237, 1995.
[47] A. Oppenheim et al., Nonlinear filtering of multiplied and convolved
signals, IEEE Trans. Audio and Electro., vol. 56, issue 8, pp. 1264-1291,
Sep. 1968.
[48] R. Fattal et al., Gradient domain high dynamic range compression,
Proceedings of the 29th annual conference on computer graphics and
interactive techniques, vol. 21, no. 3, pp. 249-256, 2002.
[49] N. Otsu, A threshold selection method from gray-level histogram, IEEE
Trans. Systems, Man and Cybem., vol. 9, issue 1, pp. 62-66, 1979.
[50] S. Finn et al., Echocardiographic speckle reduction comparison, IEEE
Trans. Ultrason. Ferroelect. Freq. Contr., vol. 58, no. 1, pp. 82-101, Jan.
2011.
[51] J. C. Bamber and R. J. Dickinson, Ultrasonic B-scanning: a computer
simulation, Phys. Med. Biol., vol. 25, no. 3, pp. 463-479, 1980.
[52] Z. Wang et al., Image quality assessment: From error visibility to
structural similarity, IEEE Trans. Image Process., vol. 13, no. 4, pp.
1-14, Apr 2004.
[53] J. Weickert et al., Efficient and reliable schemes for nonlinear diffusion
filtering, IEEE Trans. Image Process., vol. 7, no. 3, pp. 398-410, Mar
1998.
[54] P. Mrzek et al., On iterations and scales of nonlinear filters, in Proc.
Computer Vision Winter Workshop, O. Drbohlav, Ed., Valtice, Czech
Republic, pp. 61-66, Feb. 2003.
0018-9294 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.