Real-time B-mode ultrasound quality control test procedures Report of
High Quality Ultrasound B-mode Image Generation Using 2-D ...
Transcript of High Quality Ultrasound B-mode Image Generation Using 2-D ...
High Quality Ultrasound B-mode Image
Generation Using 2-D Multichannel-based
Deconvolution and Multiframe-based
Adaptive Despeckling Algorithms
A thesis submitted to the Department of Electrical and Electronic Engineering
of
Bangladesh University of Engineering and Technology
in partial fulfillment of the requirement for the degree of
MASTER OF SCIENCE IN ELECTRICAL AND ELECTRONIC ENGINEERING
byJayanta Dey
Student ID: 0417062214 P
DEPARTMENT OF ELECTRICAL AND ELECTRONIC ENGINEERING
BANGLADESH UNIVERSITY OF ENGINEERING AND TECHNOLOGY
May 2019
Scanned by CamScanner
iii
Dedication
To my parents.
iv
Acknowledgements
First and foremost, I am grateful to the Almighty for giving me the opportunity andstrength to carry out this thesis work.
I would like to extend my heart-felt gratitude to my supervisor, Dr. Md. KamrulHasan for giving me the opportunity to work on this topic. I am grateful for his un-conditional support, constant guidance and supervision throughout this research work.Working under his supervision is a reminder of the discipline, patience, and persever-ance required to carry out high-level research.
I am also grateful to the Head of the Department of Electrical and Electronic Engi-neering, BUET, for the research lab facilities. I would also like to thank the presentand past members of the DSP research lab, without whose help this thesis may nothave been possible, especially, Dr. Sharmin Rowshan Ara, Mr. Nabid Ibtehaz Nizam,Mr. Md. Hadiur Rahman Khan and Mr. Md. Shifat-E-Rabbi. Moreover, I would liketo thank my colleagues at the Department of Electrical Engineering, AUST, who werea constant source of support and inspiration and my friends, respected seniors, andbeloved juniors at BUET, who are all shining examples to follow.
This work has been supported by Higher Education Quality Enhancement Project, Uni-versity Grants Commission (CPSF#96/BUET/Win-2/ST(EEE)/2017), Bangladesh.The in vivo breast data were acquired at BUET Medical Center by Dr. Farzana Alam,Assitant Professor, Department of Radiology and Imaging, Bangabandhu Sheikh MujibMedical University, Dhaka-1000, Bangladesh.
Contents
Approval Certificate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iCandidate’s Declaration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Abstract xii
1 Introduction 1
1.1 Motivation of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Objectives of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . 21.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.4 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Deconvolution of Ultrasound Images 5
2.1 Deconvolution of RF Echo Data: Literature Review . . . . . . . . . . . 62.2 Signal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.1 Fundamentals and Frequency Domain Approach . . . . . . . . . 92.3.2 bMCFLMS Algorithm for TRF estimation . . . . . . . . . . . . 142.3.3 Effect of Noise on the Convergence of the Algorithm . . . . . . . 20
2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252.4.1 Simulation Phantom Results . . . . . . . . . . . . . . . . . . . . 272.4.2 In-Vivo Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3 B-mode Image Generation Framework 35
3.1 Despeckling of the Envelope of Deconvolved Data: Literature Review . 363.2 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2.1 Deconvolution of RF Echo Data . . . . . . . . . . . . . . . . . . 373.2.2 Proposed Despeckling Algorithm . . . . . . . . . . . . . . . . . 40
3.3 Post-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
v
CONTENTS vi
4 Results 51
4.0.1 Simulation Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 524.0.2 In-Vivo Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5 Discussion and Conclusion 62
5.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
List of Publications 65
List of Figures
2.1 Illustration of ultrasonic data acquisition system as SIMO model, show-ing the relationship between the backscattered RF data xi(n) and thepoint spread function, s(n). . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Block diagram of the proposed bMCFLMS algorithm. The upper partshows the modeling of the RF data and RF image from the transducersegmented into B blocks. The lower part shows the sequential estimationof the TRF from the RF segments using the proposed algorithm. . . . . 11
2.3 Effect of noise on the convergence of the bMCFLMS algorithm (a) Be-havior of the NPM curve around the misconvergence point for the firstblock with SNR = 30 dB. (b) Behavior of the correlation cost functionaround the misconvergence point for the first block. (c) Behavior of theNPM curve around the misconvergence point for the first block withcorrelation constraint (⇠ = 3e-7, ⇢ = 2.55, � = 2.3). (d) Misconvergencephenomenon of the second block of TRFs for B = 2 for with no addi-tive white noise in the data. (e) Misconvergence Problem Solved for theSecond Block of TRFs for B = 2 using correlation constraint. . . . . . . 21
2.4 Convergence of the bMCFLMS algorithm with the proposed correlationconstraint for the simulation phantom data at different SNR. . . . . . . 23
2.5 Deconvolution performed on a simulation phantom with 20 scatterersper resolution cell and SNR = 30 dB. The RF data size is 1038 ⇥ 128.The darker circular inclusion with radius 5 mm is created by placingscatterers with relatively lower strength than the surroundings. Log-envelope image of the (a) true TRF, (b) backscattered standard RF data,and (c) deconvolved TRF by the bMCFLMS algorithm, (d) spectra ofthe true and estimated RF data of a single scan-line marked by verticalgreen line in the log envelope images of (b) and (c). (e) The NPM curvebetween the deconvolved TRF and true TRF of the first block. . . . . . 26
vii
LIST OF FIGURES viii
2.6 Performance analysis of the bMCFLMS algorithm for the in-vivo backscat-tered RF data of a breast cyst. Spectrum of the (a) deconvolved TRF,(b) R-MINT estimated PSF, and (c) spectra of the true and estimatedRF data of a single scan-line marked by vertical red line in log envelopeimages. Standard log envelope images of (d) the backscattered RF data,(e) deconvolved TRFs, and (f-g) zoomed-in views of (d-e). . . . . . . . 27
2.7 Performance analysis of the bMCFLMS algorithm for the in-vivo backscat-tered RF data for a left carotid artery. Spectrum of the (a) deconvolvedTRF, (b) R-MINT estimated PSF, and (c) spectra of the true and es-timated RF data of a single scan-line marked by vertical red line in logenvelope images. Standard log envelope images of (d) the backscatteredRF data, (e) deconvolved TRFs, and (f-g) zoomed-in views of (d-e). . . 28
2.8 Time required per iteration by the time- and the frequency-domain al-gorithms for the simulation phantom data of size Na ⇥ 128 where Na ismade variable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1 A SIMO model for backscattered RF signal. . . . . . . . . . . . . . . . 383.2 A new SIMO model for deconvolved 2-D RF data. . . . . . . . . . . . . 413.3 The block diagram for the estimation procedure of true ultrasound image. 463.4 Block diagram of the proposed ultrasound image reconstruction method
from the raw RF data. . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.1 Effect of the number of frames in despeckling the modified Shepp-Loganphantom image using the proposed algorithm: (a) clean phantom, (b)noisy phantom (� = 0.4), (c)-(f) despeckled using 5,10, 15 and 20 frames,respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2 Despeckling of Shepp-Logan phantom image corrupted by synthetic specklenoise. (a) clean phantom, (b) noisy phantom (� = 0.4), (c) true specklenoise in the 5th frame; despeckled image using (d) SRAD, (e) OBNLM,(f) proposed MSNE; (g) extracted noise from the 5th frame using MSNE;(h) NPM measure between the true and estimated noise using MSNEwithout constraint; (i) NPM measure between the true and estimatednoise using MSNE with constraint. . . . . . . . . . . . . . . . . . . . . 53
4.3 Deconvolution of ultrasound images using adaptive bMCFLMS algo-rithm. (a) Raw RF image, (b) 1-D deconvolved image, (c) 2-D de-convolved image, (d)-(f) zoomed-in views of image segments of (a)-(c),respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
LIST OF FIGURES ix
4.4 Despeckling of breast ultrasound image- 1. (a) Deconvolved image, im-ages obtaind using (b) SRAD, (c) OBNLM, (d) proposed algorithm, (e)machine B-mode image, (f) estimated speckle pattern of the 5-th frame. 58
4.5 Despeckling of breast ultrasound image- 2. (a) Deconvolved image, im-ages obtaind using (b) SRAD, (c) OBNLM, (d) proposed algorithm, (e)machine B-mode image, (f) estimated speckle pattern of the 5-th frame. 59
4.6 Estimated B-mode image with the deconvolution step as (a) bMCFLMS,and (b) cepstrum. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
List of Tables
2.1 Symbols and Description . . . . . . . . . . . . . . . . . . . . . . . . . . 322.2 Constrained bMCFLMS Algorithm . . . . . . . . . . . . . . . . . . . . 332.3 Performance of different algorithms on simulation phantom data with
additive noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342.4 Performance of different algorithms on in-vivo data . . . . . . . . . . . 34
3.1 bMCFLMS algorithm for 2-D deconvolution of ultrasound RF image . . 41
4.1 Simulation results on the estimation accuracy in terms of NPM (dB) ofspeckle pattern using the proposed method for different noise levels . . 55
4.2 Performance measures computed for the simulation study with differentnoise level (�) using diffetrent despeckling approaches . . . . . . . . . . 56
4.3 Axial and lateral correlation energy for raw RF, 1-D and 2-D decon-volved data using the b-MCFLMS algorithm. . . . . . . . . . . . . . . . 57
4.4 NIQE measure for images despeckled with different algorithms . . . . . 60
x
Glossary
VSS Variable step-size
RF Radio-frequency
PSF Point spread function
TRF Tissue reflectivity function
LMS Least mean square
MADS Multiframe-based Adaptive Despeckling
MSNE Multiframe-based Adaptive Speckle Noise Estimation
SNC Speckle noise cancellation
xi
Abstract
Improving resolution and removing speckle noise from medical ultrasound images whilepreserving tissue texture, small details, and edges without introducing artifact and dis-tortion is a major challenge in ultrasound image restoration. The underlying physicalphenomena related with US image acquisition and imperfection of US imaging systemdesign give rise to low resolution and speckle noise that tend to reduce the image con-trast, obscure and blur image details such as inclusion and small structure boundary,tissue texture and thereby, decrease the quality and reliability of medical ultrasound.In this thesis, a complete framework of signal processing approaches comprising of de-convolution to enhance resolution, despeckling, and post-processing for the generationof ultrasound B-mode image with superior edge, details and tissue texture has beenestablished. In the first step, we propose a correlation constrained blind multichan-nel frequency-domain least-mean-squares (bMCFLMS) algorithm to undo the effectof point spread function (PSF) on the ultrasound radio-frequency (RF) data. ThebMCFLMS algorithm, however, shows misconvergence due to both channel noise andpropagation of TRF estimation error from the previous blocks. This phenomenon ismore intense in the case of md-bMCFLMS algorithm because of increased estimationerror. To address this problem, a novel constraint based on the correlation between themeasured RF data and estimated TRF is proposed in this thesis. Then in the secondstep, based on a multiple input single output (MISO) model over the consecutive de-convolved ultrasound image frames, a multiframe-based adaptive despeckling (MADS)algorithm to reconstruct a high-resolution B-mode image from raw radio-frequency(RF) data has been proposed. It utilizes the speckle patterns estimated using a novelmultiframe-based adaptive approach for ultrasonic speckle noise estimation (MSNE)based on a single input multiple output (SIMO) modeling of consecutive deconvolvedultrasound image frames to estimate the despeckled ultrasound image as single outputfrom the deconvolved image frames as multiple input. The elegance of the proposed al-gorithms is that it addresses the deconvolution and despeckling problem by completelyfollowing the signal generation model rather than the existing ad-hoc smoothening orfiltering approach described in the literature, and therefore, it is likely to maximally
xii
ABSTRACT xiii
preserve the image features. The efficacy of our proposed blind deconvolution algorithmis measured using simulation phantom and in-vivo data. The proposed md-bMCFLMSalgorithm shows normalised projection misalignment (NPM) improvement of about2.12 ⇠ 16 dB and resolution gain (RG) improvement of 1.14 ⇠ 6.4 dB compared toother techniques in the literature. Moreover, because of the frequency-domain im-plementation it is computationally more efficient, fast converging and robust than itstime-domain counterpart l1-bMCLMS algorithm reported in the literature. Again, theefficacy of the proposed despeckling algorithm is evaluated both visually and quantita-tively on the simulation and in-vivo data. The results show 8.55�15.91 dB, 8.24�14.94
dB, 0.57�7.03 improvement in terms of SNR, PSNR, and NIQE, respectively, for sim-ulation data and 2.22� 3.17 improvement in terms of NIQE for in-vivo data comparedto the traditional despeckling algorithms.
Chapter 1
Introduction
In this Chapter, we discuss the motivation behind the development of a complete
framework of novel signal processing approaches, including deconvolution, despeckling,
and post-processing to generate a high quality B-mode image from raw radio-frequency
(RF) image. Next, the primary objectives of this thesis are highlighted. Finally, the
organization of the thesis is described.
1.1 Motivation of the Thesis
Ultrasound (US) imaging system being non-invasive, non-ionizing, portable, and cost
effective has become the most prevalent diagnostic tool among all the currently avail-
able imaging modalities, e.g., X-ray, magnetic resonance imaging, and computed to-
mography. However, imperfection of US imaging system design and the underlying
physical phenomena related with US image acquisition give rise to low resolution and
speckle noise that tend to reduce the image contrast, obscure and blur image details
such as inclusion and small structure boundary, tissue texture and thereby, decrease
the quality and reliability of medical ultrasound [1]. The low resolution of US image
can be modelled as the convolution of point spread function (PSF) of ultrasound imag-
ing system with the tissue reflectivity function (TRF). The removal of PSF effect from
the measured backscattered RF images can restore the resolution of images and thus
improve the diagnostic quality of ultrasound imaging. On the other hand, speckle noise
1
CHAPTER 1. INTRODUCTION 2
is a granular pattern inherent in any coherent imaging modalities [2], [3] similar to ul-
trasound imaging. It results from the constructive and destructive interferences of the
reflected echos with different phases and amplitudes from the target at the receiving
transducer. Removal of speckle noise from US images is difficult due to its multiplica-
tive nature and the challenge of maintaining the precise texture of the image [4]. As
traditional despeckling filters distort the original image texture and introduce artifacts
like blurring edges, changing the shape of structures present in the image by smoothen-
ing the noise corrupted image, the original noise affected images are sometimes more
preferred than the noise-removed ones in the analysis where the image details have high
importance [4]. Therefore, an effective signal processing approach for removing speckle
noise from the image while preserving the original tissue texture and small details of
the image is vital to increase the diagnostic potential of medical ultrasound.
1.2 Objectives of the Thesis
The objectives of this work are:
1. To develop a new algorithm for the deconvolution of radio frequency (RF) ultra-
sound data.
2. To develop a novel model and estimation algorithm for despeckling of deconvolved
ultrasound RF data.
3. To establish a complete framework for high quality ultrasound B-mode image
generation.
4. To compare the performance of the proposed techniques with some well-known
methods using data from simulation phantom, experimental phantom, and pa-
tient (in vivo).
CHAPTER 1. INTRODUCTION 3
1.3 Problem Formulation
Low resolution and speckle noise are the major issues related with US imaging. How-
ever, an overall US image enhancement can be achieved by addressing these two issues
in a sequential two-step process as described in [1]. First, the correlation between the
image samples is to be minimized to increase the image resolution, and second, speckle
noise has to be removed from the decorrelated image to improve the image contrast
and better visualize the tissue texture. To achieve the aforementioned two objectives,
a suitable model representing the US imaging system is necessary. With this in view,
considering linear wave propagation through the tissue, and the scattering of the ultra-
sound pulse in the tissue as weak, we can use the first order Born approximation and
consider the tissue scattering system as a linear system [1]. Therefore, the blurring, i.e.,
low resolution of an RF-image can be modeled as the result of convolution between the
point-spread function (PSF) s(m,n) of the imaging system with the tissue reflectivity
function (TRF) h(m,n) [1], [5], [6]. Mathematically, this can be written as
x(m,n) = s(m,n) ⇤ h(m,n) + v(m,n) (1.1)
where x(m,n) is the backscattered ultrasound image data from the n-th A-line at
discrete time m and v(m,n) is the additive noise associated with measurement error
and other physical phenomena not accounted by the convolution model. In US imaging,
h(m,n) is corrupted by speckle noise, and as described in [1], [7], [8], he(m,n), i.e., the
envelope of h(m,n) can be modeled as multiplicative with the true image as
he(m,n) = r(m,n)u(m,n) + ⇣(m,n) (1.2)
where r(m,n), u(m,n), and ⇣(m,n) are true image, speckle noise, and additive noise
resulting from the deconvolution process and the portion not accounted by the mul-
tiplicative process, respectively. Therefore, in ultrasound imaging, we are given the
back-scattered data x(m,n), and our objective is to design algorithms to estimate the
TRF h(m,n) from x(m,n), and thereafter, obtain the despeckled image r(m,n) from
the envelope of h(m,n).
CHAPTER 1. INTRODUCTION 4
1.4 Organization of the Thesis
This thesis consists of five chapters. Chapter 1 gives a brief discussion about the
motivation of the thesis and the limitations of the existing techniques. Chapter 2 has
the detailed description of the deconvolution algorithm (bMCFLMS) along with its
performance analysis. Chapter 3 establishes and proposes the framework for a high
quality B-mode image generation. Chapter 4 tests our proposed approaches on the
simulation and the in-vivo breast data. It also represents RF image processing and
the information of the presets of the ultrasound instrument. Finally, in Chapter 5, a
conclusion has drawn showing the outcomes and drawbacks of the thesis and indicating
the thoughts and topics which require further research for improvements.
Chapter 2
Deconvolution of Ultrasound Images
In this chapter, we propose a blind deconvolution algorithm in the frequency-domain
with correlation constraint for noise-corrupted ultrasound RF data. Due to smaller
eigenvalue spread, a frequency-domain approach facilitates faster convergence of the
adaptive algorithms compared to the time-domain ones [9], [10]. Moreover, the variable
step-size (VSS) MCFLMS algorithm is in general known to be more noise robust com-
pared to the time-domain ones at the same noise level [11], [12]. Unlike the blocking
procedure described in [13], in this thesis, we introduce a new blocking technique that
facilitates the use of FFT and make the algorithm computationally efficient. However,
as with the other reported cross-relation based blind adaptive algorithms [14], [15], the
proposed algorithm also suffers from misconvergence in the presence of channel noise
and estimation error from the previous estimated blocks of the TRFs. To overcome this
problem, we propose a novel constraint based on the correlation between the measured
RF data and estimated TRFs that can compensate the effect of noise and estimation
error stated above. The performance of the proposed blind deconvolution techniques
is evaluated on the simulation phantom and in-vivo data.
5
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 6
2.1 Deconvolution of RF Echo Data: Literature Re-
view
A number of algorithms have been proposed in the literature to deconvolve the PSF
from the received RF data and thereby restore the image resolution. These algorithms
are mainly of two groups. The first group estimates the PSF first and then a classical
deconvolution algorithm is applied to estimate the TRFs. These methods are based on
homomorphic filtering [16], [17], [18], [19] which involve filtering out the PSF either in
the cepstrum domain [16], [18] or in the log magnitude domain [19]. These algorithms
are elegant in the sense that they are simple and can be implemented in real-time.
Filtering the wavelet coefficients of the log magnitude spectrum gives better result in
terms of mean square error (MSE) than the cepstrum based methods [18]. However,
tuning the length of the filter in cepstrum domain and selecting the decomposition
level of wavlet decomposition determines the smoothness of the estimated PSF [20]
and hence the overall deconvolution accuracy. In addition, this category of techniques
assume that the spectrum of the PSF and the TRF lie in separate spectral band
which is not completely true [16]. Moreover, inaccuracy in phase unwrapping poses
another problem for these algorithms [18]. To solve the problem of phase unwrapping,
a recent hybrid parametric inverse filtering (HYPIF) algorithm [18] has been proposed
which estimates the PSF in two steps. First, partial information of the PSF, i.e.,
power spectrum is estimated using the homomorphic filtering. This partial information
obtained is used to constraint the shape of the inverse filter. Then linearity of the
inverse filter is exploited to recover the phase of the PSF. However, in this method the
energy of the inverse filter should be regularized to avoid instability where the PSF has
zero or very low magnitude.
The second group of algorithms estimates the PSF and the TRFs simultaneously.
Among them the blind deconvolution method described in [21] improves the conver-
gence speed and reduces the computational load by projecting the TRF into the null
space of the correlation matrix and the PSF onto the space spanned by the third-order
B-spline wavelet basis. However, in the presence of noise thresholding is necessary to
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 7
determine the null space bases. Another classical algorithm of iterative blind deconvo-
lution described in [22] suffers from poor convergence property. Again, the parametric
methods [23], [24] of this group models the imaging process using a low-order autore-
gressive moving average (ARMA) system. However, these methods are only applicable
for fairly smooth RF images only. In addition, due to extremely complex composition
of most biological tissues, derivation of a convenient and accurate parametric model
for the imaging system is not possible [25].
The algorithms discussed so far assume spatially invariant PSF. However, while
propagating through the tissue the PSF suffers from attenuation dependent on the
depth of penetration. In order to address this problem, the spatially variant meth-
ods reported in the literature divides the RF data into several overlapping or non-
overlapping blocks so that the PSF may be considered stationary for that block. The
algorithm proposed in [25] assumes cyclic convolution of the PSF with the TRF blocks
for each of these RF data blocks instead of linear convolution between the PSF and the
total TRF. Lately, a block-based blind deconvolution method in the time-domain using
the multichannel LMS (MCLMS) algorithm has been presented in [13] which estimates
the TRFs block by block to account for the nonstationarity of the PSF. The blocking
procedure described in [13] uses convolution matrix segmentation and thereby restricts
the use of fast Fourier transform (FFT) for implementing convolution. Therefore, the
algorithm is computationally inefficient due to the implementation of time-domain con-
volution using matrix multiplication. In addition, the size of the matrix increases as
the number of blocks increases. Furthermore, it suffers from misconvergence in the
presence of additive white noise and propagation of estimation error from block to
block. An attempt was made to solve this problem using a damped variable step-
size [26], gradient averaging and l1-norm constraint. In order to apply the l1-norm
regularization in the algorithms described in both [25] and [13], the data is assumed
to be sparse which is not generally true for the case of in-vivo data. The noise effect
minimizing methods described in these algorithms are adopted on ad hoc basis and the
misconvergence may not be completely stopped even after applying all the aforesaid
noise effect minimization techniques.
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 8
2.2 Signal Model
In the ultrasound imaging system, an array of piezoelectric elements sequentially emit
the same ultrasound pulse in the tissue and receive echo signals from multiple A-lines.
This system can be modeled as a SIMO model with the ultrasound pulse as the single
s(n)
h1(n)
h2(n)
hM
(n)
x1(n)
x2(n)
xM
(n)
v1(n)
v2(n)
vM
(n)
PSF
TRFBackscattered
RF Data
Additive
Noise
Figure 2.1: Illustration of ultrasonic data acquisition system as SIMO model, show-
ing the relationship between the backscattered RF data xi(n) and the point spread
function, s(n).
input, the measured echo signal lines as multiple outputs, and the TRFs along the
axial direction as multiple system channels [21]. Figure 2.1 shows the SIMO model of
the backscattered ultrasound RF data, where the ultrasound pulse or the PSF s(n)
convolves with the i-th channel transfer function or TRF hi(n). With additive noise
vi(n), the measured RF data are given by
xi(n) = s(n) ⇤ hi(n) + vi(n), i = 1, 2, · · · ,M (2.1)
where xi(n) denotes the backscattered RF data of the i-th scan line and M is the
number of total scan lines. If the length of xi(n) is L and that of s(n) is Ls with
Ls << L, we can write (2.1) in matrix form as
xi(n) = S(n)hi + vi(n) (2.2)
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 9
where S(n) is the (L+ Ls � 1)⇥ L convolution matrix constructed from s(n), and
hi =hhi(n) hi(n� 1) · · · hi(n� L+ 1)
iT
xi(n) =hxi(n) xi(n� 1) · · · xi(n� L� Ls + 2)
iT
vi(n) =hvi(n) vi(n� 1) · · · vi(n� L� Ls + 2)
iT
However, as each data sample of the ultrasound RF signal results from the reflection
of the ultrasound pulse from a scatterer of the tissue, in reality, we have RF echo data
equal to the TRF length L instead of L + Ls � 1. Moreover, (2.1) assumes that
the PSF is stationary, i.e., remains constant while penetrating the tissue. But, if
the acquired RF data is long in the axial direction, the PSF suffers from the depth
dependent attenuation while traveling through the tissue. Then this nonstationary PSF
restricts the direct use of the cross-relation based MCFLMS algorithm for blind SIMO
model identification in the deconvolution of ultrasound images [13]. Therefore, the
main challenge in ultrasound deconvolution process is to estimate the TRFs, hi(n), i =
1, 2, · · · ,M , using the truncated and nonstationary RF data corrupted by additive
white noise. Here the deconvolution process is carried out in the frequency-domain to
obtain more robustness to noise than in the time-domain along with faster convergence
facilitated by smaller eigenvalue spread and reduced computational complexity.
2.3 Method
2.3.1 Fundamentals and Frequency Domain Approach
For noise-free case, the cross-relation error defined in the following can be exploited to
estimate the TRFs:
eij(n) = xi(n) ⇤ hj(n)� xj(n) ⇤ hi(n), i, j = 1, 2, · · · ,M,
i 6= j (2.3)
Note that the error function in (2.3) becomes zero when xi(n) in (2.1) for noiseless
condition is substituted into it. However, as shown in [13], due to incomplete data
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 10
acquisition in ultrasound imaging only the first L samples of the error function will be
equal to zero. Taking this into account, the truncated error function in matrix form is
written as
eij = D(Cxihj �Cxjhi) (2.4)
where,
xi(n) = xi(n), n = 0, 1, 2, · · · , L� 1,
Cxi is the convolution matrix formed with the truncated RF data xi(n) of length L
and
D =hIL⇥L 0L⇥(L�1)
i
is the truncation matrix, I is the identity matrix of size L⇥ L, 0 is the null matrix of
size L⇥ (L� 1),
hi =hhi(n) hi(n� 1) · · · hi(n� L+ 1)
iT
Due to nonstationarity of the PSF in ultrasound imaging, (2.4) cannot be directly
used to estimate the TRFs. An appropriate solution to this problem is to estimate
the TRFs block by block from the error blocks formed from (2.4) as shown in [13].
Then we can consider the PSF as quasi-stationary within that particular block. How-
ever, the blocking approach described in [13] requires matrix multiplication between
segments of a convolution matrix formed from the RF data and estimated TRF. Its
frequency-domain implementation will again require multiplication of an error block
with a DFT matrix making the algorithm computationally inefficient. Unlike in [13],
here we propose a different blocking approach that uses smaller and fixed size vector
operations with FFT applicability and thus a faster approach.
The cross-relation error in (2.3) is basically the difference between two convolution
operations. Therefore, if we can implement a full convolution as a summation of
convolutions between smaller blocks of signals, (2.4) can be implemented efficiently by
eliminating the need of using the truncation operator D as well as the unnecessary
computation of the full convolution (the bracketed part in (2.4)). Before we go into
details, we first show how a convolution can be implemented block by block. In the
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 11
con
volu
tion
PSF
TRF
Block #1
. . .
Block #B
RF
seg
men
ts
L
Ls
=1
...
B
Not
ava
ilab
le
from
tra
nsd
uce
r
}
LL
L -1s
bMCFLMS
RF image from transducer
b = 1,2,..,B
TR
F s
egm
ents
RF
seg
men
ts
convoluted RF image
TR
F s
egm
ent
))))))))
Block #1
. . .
Block #b-1
Block #1
. . .
Block #b
Figure 2.2: Block diagram of the proposed bMCFLMS algorithm. The upper part
shows the modeling of the RF data and RF image from the transducer segmented into
B blocks. The lower part shows the sequential estimation of the TRF from the RF
segments using the proposed algorithm.
subsequent discussion, ‘˜ ’ is used to denote truncated data, channel number is placed
as subscript and block number is presented in the superscript. Now consider the
convolution between the truncated RF data xi(n) of an arbitrary channel i and the
estimated TRF data hj(n) of the j-th channel given by
yij(n) = xi(n) ⇤ hj(n) (2.5)
The z-transform of hj(n), denoted as Hj(z) can be expressed as
Hj(z) = hj(0) + hj(1)z�1 + · · ·+ hj(Lb � 1)z�(Lb�1)
+ hj(Lb)z�Lb + · · ·+ hj(BLb)z
�(BLb�1)
= H1j(z) + z
�LbH2j(z) + · · ·+ z
�(B�1)LbHB
j(z) (2.6)
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 12
where Hb
j(z) can be written as
Hb
j(z) = hj
�(b� 1)Lb) + hj
�(b� 1)Lb + 1)z�1
+ · · ·+ hj
�(bLb � 1))z�(Lb�1)
, b = 1, 2, · · · , B (2.7)
Here B is the total number of blocks and Lb = floor(L/B).
Similarly, for the i-th channel RF data xi(n), we can write
Xi(z) = X1i(z) + z
�LbX2i(z) + · · ·+ z
�(B�1)LbXB
i(z) (2.8)
where
Xb
i(z) = xi
�(b� 1)Lb) + xi
�(b� 1)Lb + 1)z�1
+ · · ·+ xi
�(bLb � 1))z�(Lb�1)
, b = 1, 2, · · · (2.9)
Using (2.7) and (2.8) , (2.5) can be written in the z-domain as
Yij(z) = Xi(z)Hj(z)
= X1i(z)H1
j(z) + z
�LbX1i(z)H2
j(z)+
z�LbX
2i(z)H1
j(z) + z
�2LbX2i(z)H2
j(z)
+ · · ·+ z�2(B�1)LbX
B
i(z)HB
j(z) (2.10)
Taking the inverse z-transform of (2.10), we get
yij(n) = x1i(n) ⇤ h1
j(n) + z
�Lbx1i(n) ⇤ h2
j(n)
+ z�Lbx
2i(n) ⇤ h1
j(n) + z
�2Lbx2i(n) ⇤ h2
j(n)
+ · · ·+ z�2(B�1)Lbx
B
i(n) ⇤ hB
j(n) (2.11)
Here multiplication by z�1 refers to unit sample delay, and x
b
i(n) and h
b
j(n) represent
the b-th block of the i-th channel RF and j-th channel TRF, respectively. As evident
from (2.11), a convolution operation can be splitted into a sum of smaller convolution
blocks of identical length. For reasons explained in (2.4), we will consider the first L
samples of total 2L� 1 samples of the convolution in (2.11). Now modifying (2.11) we
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 13
can write:
yij(n) = x1i(n) ⇤ h1
j(n) + z
�Lbx1i(n) ⇤ h2
j(n)
+ z�Lbx
2i(n) ⇤ h1
j(n) + z
�2Lbx2i(n) ⇤ h2
j(n)+
· · ·+ z(B�1)Lbx
B
i(n)h1
j(n) + z
(B�1)Lbx1i(n)hB
j(n)
= y11ij(n) + z
�Lby21ij(n) + z
�Lby12ij(n) + · · ·
+ z(B�1)Lby
1Bij(n) + z
(B�1)LbyB1ij(n) (2.12)
where
ypq
ij(n) = x
p
i(n) ⇤ hq
j(n)
To account for the nonstationarity problem of ultrasound PSF, the convolution must
be evaluated in smaller blocks of Lb samples. Here each of the smaller convolutions
in (2.12) is of length 2Lb � 1. Now if we observe (2.12), it is evident that only the
first Lb samples of y11ij(n) contributes to the first convolution block y
1ij
and its last
Lb � 1 samples contributes to the next convolution block y2ij. Adding to this the first
Lb samples of y12ij(n) and y
21ij(n), we get the second convolution block y
2ij
and so on.
As the first block does not represent the general idea behind the blocking technique,
we explain the mathematical operations on the second convolution block of length Lb.
Here two truncation matrices A1 and A2 are used to take the last Lb � 1 and the first
Lb samples of a convolution, respectively. Now,
1. y11ij
= A1Cx1ih1j= A1y
11ij
where A1 =
2
40(Lb�1)⇥LbI(Lb�1)⇥(Lb�1)
01⇥Lb01⇥(Lb�1)
3
5
2. y12ij
= A2Cx1ih2j= A2y
12ij
where A2 =hILb⇥Lb
0Lb⇥(Lb�1)
i
3. y21ij
= A2Cx2ih1j= A2y
21ij
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 14
Therefore, the second block of yij described in (2.12) is given by
y2ij= y
11ij+ y
12ij+ y
21ij
= A1y11ij+A2y
12ij+A2y
21ij
=1X
p=1
A1yp(2�p)ij
+2X
p=1
A2yp(2�p+1)ij
In general, for any block b, the last Lb�1 length of the following convolutions contribute
to the b-th block convolution yb
ij:
yp(b�p)ij
= Cxpih(b�p)j
, p = 1, 2, · · · , b� 1 (2.13)
And the first Lb length of the following convolutions contribute to yb
ij:
yp(b�p+1)ij
= Cxpih(b�p+1)j
, p = 1, 2, · · · , b (2.14)
The convolution between xi(n) and hj(n) for the b-th block is then given by
yb
ij=
b�1X
p=1
A1yp(b�p)ij
+bX
p=1
A2yp(b�p+1)ij
(2.15)
2.3.2 bMCFLMS Algorithm for TRF estimation
As the truncated cross-relation error described in (2.4) is the difference between two
convolutions, a similar approach as described in (2.13), (2.14) and (2.15) can be adopted
to evaluate the cross-relation error block by block. In a similar way to (2.13), the last
Lb�1 samples of the following error function contributes to the b-th block cross-relation
error:
ep(b�p)ij
= Cxpih(b�p)j
�Cxpjh(b�p)i
, p = 1, 2, · · · , b� 1 (2.16)
where ep(b�p)ij
is the cross-relation error considering the p-th block of RF data and
(b� p)-th block of TRF data. Notice that now the true TRF has been replaced by the
estimated TRF. In the same manner, according to (2.14), the first Lb samples of the
following error function contributes to the b-th block cross-relation error:
ep(b�p+1)ij
= Cxpih(b�p+1)j
�Cxpjh(b�p+1)i
, p = 1, 2, · · · , b (2.17)
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 15
Therefore, the cross-relation error for the b-th block is
eb
ij=
b�1X
p=1
A1ep(b�p)ij
+bX
p=1
A2ep(b�p+1)ij
=b�1X
p=1
A1ep(b�p)ij
+bX
p=2
A2ep(b�p+1)ij
+A2e1bij
=b�1X
p=1
ep(b�p)ij
+bX
p=2
ep(b�p+1)ij
+ e1bij
(2.18)
Notice that the error components ep(b�p)ij
and ep(b�p+1)ij
that constitute the b-th block
error function eb
ij, can be computed parallely. Here only the third term of the right
side of (2.18) depends on the b-th block TRF, hb and the first two terms depend on
h1, h
2, · · · , and h
b�1. While estimating the TRF of the b-th block, all the TRF hq
for q = 1, 2, · · · , b � 1 are already known and therefore, may be treated as constant.
Therefore, we can write (2.18) as
eb
ij= e
1bij+ c (2.19)
where c is a constant defined as
c =b�1X
p=1
ep(b�p)ij
+bX
p=2
ep(b�p+1)ij
Taking Fourier transform of (2.19), we get the Fourier transformed error ebij
as
eb
ij= F1e
1bij+ F1c
= F1.A2(Cx1ihb
j�Cx
1jhb
i) + c
= F1A2F�12 F2(Cx
1ihb
j�Cx
1jhb
i) + c (2.20)
where F1 and F2 denote the DFT matrix of size Lb ⇥ Lb and (2Lb � 1) ⇥ (2Lb � 1),
respectively. Rewriting (2.20), we get
eb
ij= F1A2F
�12 (x1
i. ⇤ h
b
j� x
1j. ⇤ h
b
i) + c
= Be1bij+ c (2.21)
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 16
where B = F1A2F�12 , x1
iis the Fourier transform of the first block of the i-th channel
TRF data x1i
and hb
idenotes the Fourier transform of hb
iof length 2Lb�1. As convolu-
tion in the time-domain becomes multiplication in the frequency-domain, we can write,
for example, F2(Cx1ihb
j) as x
1i. ⇤ h
b
j, where ‘.⇤’ denotes the element-wise multiplication
operation. Subsequently, underbar with any quantity will define its Fourier transform.
Now, the cost function for the b-th block for estimating hb
can be defined as
Jb =
M�1X
i=1
MX
j=i+1
ebH
ijeb
ij(2.22)
Here ‘H’ denotes the hermitian operation. An estimate of the b-th block TRF, hb
can
be obtained by minimizing the cost function Jb as
hb
= arghb min J b, subject to ||h|| = 1 (2.23)
where ‘|| · ||’ denotes the l2-norm and
h(m) =hh1T(m) h
2T(m) · · · h
bT
(m)iT
(2.24)
with
hb
(m) =hhb
1(m) hb
2(m) · · · hb
M(m)
i(2.25)
In what follows, we derive a variable step-size multichannel FLMS algorithm for the
solution of (3.6).
Substituting (2.21) into (2.22), we get
Jb =
M�1X
i=1
MX
j=i+1
�Be
1bij+ c�H �
Be1bij+ c�
=M�1X
i=1
MX
j=i+1
�e1bHij
BHBe
1bij+ e
1bHij
BHc+ c
HBe
1bij+ a�
(2.26)
where a = cHc is a constant. Taking the gradient of (2.26) with respect to the b-th
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 17
block TRF of the k-th channel, we get
@Jb
@hb⇤k
=@
@hb⇤k
M�1X
i=1
MX
j=i+1
(e1bHij
BHBe
1bij+ e
1bHij
BHc+
cHBe
1bij+ a)
�
=k�1X
i=1
�x1⇤i. ⇤BH
Be1bik+ x
1⇤i. ⇤BH
c�
�MX
j=k+1
�x1⇤j. ⇤BH
Be1bkj+ x
1⇤j. ⇤BH
c�
=MX
i=1
x1⇤i. ⇤BH(Be
1bik+ c)
=MX
i=1
x1⇤i. ⇤BH
eb
ik, k = 1, 2, · · · ,M (2.27)
Here ‘*’ denotes the conjugate operation. The update equation of the bMCFLMS
algorithm for the b-th block of the RF data is given by
hb
(m+ 1) = hb
(m)� µb(m)rbJ
b(m)|h=h(m), b = 1, 2, · · · , B (2.28)
where,
rbJb(m) =
@Jb(m)
@hb⇤(m)
=h
@Jb(m)
@hb⇤1 (m)
@Jb(m)
@hb⇤2 (m)
· · · @Jb(m)
@hb⇤M (m)
i(2.29)
Here, hb
(m) denotes the m-th iteration estimate of hb and µb(m) is the variable step-
size (VSS) for the b-th block. The step-size is adapted so that the distance between
hb
(m + 1) and hb
(m) becomes minimum at each iteration and for noise-free case it is
given by (see [12] for details)
µb(m) =
hbT
(m)
||rbJb(m)||2rbJ
b(m)� hbT (m)
||rbJb(m)||2rbJ
b(m) (2.30)
Here ‘T ’ denotes the transpose operation. The problem with this equation is that we
need to know the true TRF of the b-th block to calculate µb(m) which is not known
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 18
beforehand. Unlike in [26], hb and rbJb(m) are not orthogonal because rbJ
b(m) is
not only a function of hb, but also of other blocks up to the b-th block. Therefore, we
use the TRFs and the gradient of the cost function up to the b-th block to calculate
the step-size as
µb(m) =
hT (m)
||rJ b(m)||2rJb(m)� h
T (m)
||rJ b(m)||2rJb(m) (2.31)
where,
rJb(m) =
@Jb(m)
@h⇤(m)
=h( @J
b(m)
@h1⇤
(m))T ( @J
b(m)
@h2⇤
(m))T · · · ( @J
b(m)
@hb⇤(m)
)TiT
(2.32)
Now, the second term of (2.31) becomes zero as the true TRFs vector, h formed by
concatenating the true TRFs of all the blocks up to the b-th block is orthogonal to
rJb(m). The step-size in (2.31) then becomes
µb(m) =
hT
(m)
||rJ b(m)||2rJb(m) (2.33)
To evaluate (2.32) and (2.33), we also need to derive the gradients of (2.22) with respect
to other blocks q, where q 6= b. To this end, as we are taking gradient with respect to
hq
, we can consider other parts of (2.18) which do not depend on hq
as constant. From
(2.18), we can write
eb
ij= e
(b�q)q + e(b�q+1)q +
b�1X
p=1,p 6=b�q
ep(b�p)
+bX
p=1,p 6=b�q+1
ep(b�p+1)
= e(b�q)q + e
(b�q+1)q + c1 (2.34)
where the constant c1 is defined as
c1 =b�1X
p=1,p 6=b�q
ep(b�p) +
bX
p=1,p 6=b�q+1
ep(b�p+1)
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 19
Taking the Fourier transform of (2.34), we get
eb
ij= B1e
(b�q)q +Be(b�q+1)q + c1 (2.35)
where B1 = F1A1F�12 .
Now the cost function in (2.22) becomes
Jb =
M�1X
i=1
MX
j=i+1
ebH
ijeb
ij
=M�1X
i=1
MX
j=i+1
�B1e
(b�q)qij
+Be(b�q+1)qij
+ c1)H
�B1e
(b�q)qij
+Be(b�q+1)qij
+ c1)
Taking gradient with respect to the conjugate of hq
k, where q 6= b, we get
@Jb
@hq⇤k
=MX
i=1
(x(b�q)⇤i
. ⇤BH
1 eb
ik+ x
(b�q+1)⇤i
. ⇤BHeb
ik) (2.36)
Using (2.36) with (2.32), the VSS µb(m) in (2.33) can now be calculated and the TRFs
are updated using (2.28). The TRF up to the b-th block is normalized to avoid the
trivial zero solution after each update, i.e.,
h(m+ 1) =h(m+ 1)
||h(m+ 1)||(2.37)
Here unity norm constraint is applied on all the blocks of TRFs up to the b-th block
among the total B blocks of TRFs, because the cost function Jb contains all the TRFs
up to hb
. Finally, the inverse Fourier transform of the estimate obtained using (2.28)
will result in an estimate of the b-th block of the TRF. Executing (2.28) for all the
blocks b = 1, 2, · · · , B, an estimate of the full TRF can be obtained from (3.7).
The overall idea of the algorithm is depicted in Fig. 2.2 where the RF image is
divided into B segments. As reported in the literature [25], the relative scaling factor
between two adjacent estimated TRF segments may result in visible discontinuities at
the segment boundaries known as blocking artifacts. However, as shown in (2.26), the
cost function for the b-th block TRF includes all the TRFs up to the (b� 1)-th block.
Therefore, optimization of (2.26) to update the b-th block TRF ensures the continuity
between the blocks of TRFs up to the b-th block.
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 20
2.3.3 Effect of Noise on the Convergence of the Algorithm
So far we have assumed a noise-free case for the adaptive algorithm. However, the
presence of noise may not be avoided in practice. It is well-known that noise has a
significant impact on the convergence of the cross-relation based adaptive algorithms
[14], [15], [27], [28], [29]. To stop misconvergence of this type of algorithms due to
the effect of noise, a spectral constraint is proposed in [14] where it is assumed that
the acoustic channel impulse response is spectrally flat. This assumption, therefore, is
not valid for TRFs. The time-domain approach described in [13] uses damped variable
step-size as described in [26], gradient averaging, and l1-norm constraint to prevent
misconvergence. The l1-norm constraint gives some sort of robustness against noise
when the RF data is sparse. But unfortunately sparsity of in-vivo data cannot be
guaranteed. In [15] a modified cost function has been proposed where it is assumed that
the additive noise in different channels have the same variance, a condition that may
not be satisfied in practice. Thus, none of these techniques are generalized to address
the problem of misconvergence. In this work, we develop a generalized approach to get
rid of this problem and thereby make the adaptive algorithm robust to noise.
As described in the previous section, we attempt to make the error function de-
scribed in (2.3) zero block by block in the proposed algorithm. In noisy case, (2.3)
becomes
eij(n) = [xi(n) + ni(n)] ⇤ hj(n)� [xj(n) + nj(n)] ⇤ hi(n)
= [xi(n) ⇤ hj(n)� xj(n) ⇤ hi(n)]+
[ni(n) ⇤ hj(n)� nj(n) ⇤ hi(n)]
= es
ij(n) + e
n
ij(n) (2.38)
where ni(n) denotes the additive noise in the i-th channel, es
ij(n) is the error due
to noiseless data and en
ij(n) is the error due to noise. Therefore, in noisy case the
gradient in (2.27) will also have two components – one due to es
ij(n) and the other
is for en
ij(n). However, as the noise power is generally lower than the desired signal
power, in the beginning, the signal gradient will be higher than the noise gradient.
Thus esij(n) will reduce to a lower value faster than e
n
ij(n). And at a certain instant of
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 21
iteration, the two error values will be comparable. After this point, the noise gradient
dominates and causes the solution to misconverge. In order to prevent the algorithm
from misconverging, we have to introduce a constraint in the estimation process that
somehow restrains the noisy gradient.
-17.2
-17.6
-18
-18.4
-17
-17.2
-17.4
-17.6
-17.8
-18
-18.2
-18.4
-18.6
-18.8
500 1000 1500 150010005000
a0
-2
-4
-6
-8
-100 1 2 3 4 5 6 7 8 9 10
-2
-4
-6
-8
-10
-120 1 2 3 4 5 6 7 8
x 104
x 104
No. of iterations No. of iterations No. of iterations
0
320
318
316
314
312
310
Jc
orr
-17
-17.2
-17.4
-17.6
-17.8
-18
-18.2
-18.4
-18.6
-18.8
0 500 1000 1500N
PM
(dB
)
NP
M(d
B)
NP
M(d
B)
NP
M(d
B)
a
b
c d
e
Figure 2.3: Effect of noise on the convergence of the bMCFLMS algorithm (a) Behavior
of the NPM curve around the misconvergence point for the first block with SNR = 30
dB. (b) Behavior of the correlation cost function around the misconvergence point for
the first block. (c) Behavior of the NPM curve around the misconvergence point for the
first block with correlation constraint (⇠ = 3e-7, ⇢ = 2.55, � = 2.3). (d) Misconvergence
phenomenon of the second block of TRFs for B = 2 for with no additive white noise in
the data. (e) Misconvergence Problem Solved for the Second Block of TRFs for B = 2
using correlation constraint.
Consider the following correlation between the RF data xi(n) and estimated TRF
hi(n):
r0i(n) = xi(n) ⇤ hi(�n) (2.39)
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 22
Using (2.1), (2.43) can be expressed as
r0i(n) = [s(n) ⇤ hi(n) + vi(n)] ⇤ hi(�n)
= s(n) ⇤ hi(n) ⇤ hi(�n) + vi(n) ⇤ hi(�n)
= s(n) ⇤ hi(n) ⇤ hi(�n)
= s(n) ⇤ rhi(n) (2.40)
where the correlation of noise with TRF, i.e., vi(n) ⇤ hi(�n) is assumed zero and
rhi(n) = hi(n) ⇤ hi(�n) (2.41)
From filtering point of view, (2.40) can be viewed as if hi(�n) is filtered by hi(n)
and the filter output is further filtered by s(n). Here only hi(n) is changing with
iteration. Up to the misconvergence point hi(n) is getting closer in shape to the true
TRF hi(n) and after the misconvergence point, it deviates from the shape of hi(n).
From the concept of matched filter [30], we know that a filter passes maximum energy
at its output if the input signal is of the same shape of the filter impulse response.
Therefore, the energy in rhi(n) increases as hi(n) gets closer to the misconvergence
point, but it decreases after the misconvergence point. As a result, the energy in r0i(n)
reaches its maximum value at the misconvergence point. In this thesis, we exploit this
phenomenon to prevent misconvergence. With a little modification of (2.39), consider
the following convolution:
ri(n) = xi(n) ⇤ hi(n) (2.42)
It is obvious that r0i(n) and ri(n) both have the same power spectrum but different
phase spectrum. Therefore, we can use (2.42) instead of (2.39) while using its energy
as constraint to prevent misconvergence. Since we estimate TRFs block by block, we
cannot use the full length convolution as described in (2.42), rather we will follow the
block-based convolution as described in the previous section. In a similar manner to
(2.15), for the b-th block of the total convolution length defined in (2.42), we can write
rb
i=
b�1X
p=1
A1Cxpih(b�p)i
+bX
p=1
A2Cxpih(b�p+1)i
(2.43)
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 23
0 500 1000 1500-20
-18
-16
-14
-12
-10
-8
-6
-4
-2
0
30 dB25 dB20 dB10 dB
No. of iterations
NP
M(d
B)
30 dB25 dB
20 dB
10 dB
Figure 2.4: Convergence of the bMCFLMS algorithm with the proposed correlation
constraint for the simulation phantom data at different SNR.
To use the energy in rb
ias constraint, consider the following cost function for the
b-th block of data for i = 1, 2, · · · ,M :
Jb
corr=
MX
i=1
rbH
irb
i(2.44)
In order to show that J b
corrbecomes maximum at the misconvergence point graphically,
the simulation phantom data as described later in the result section was used. In Fig.
2.3(a) the misconvergence phenomenon is shown on the simulation data with 30 dB
SNR. As shown in Fig. 2.3(b), the first block correlation cost function J1corr
is maximum
at the misconvergence point and then it decreases. Therefore, the misconvergence due
to noise can be avoided if we minimize the b-th block cost function Jb in (2.22) while
at the same time maximize Jb
corror equivalently minimize �J
b
corr. Adding (2.44) as
constraint to our previous cost function in (2.22) gives
Jbt(m) = J
b(m)� (m)J b
corr(m) (2.45)
where (m) is the Lagrange multiplier, also known as the coupling factor. In general,
for any block b, the gradient of the cost function with respect to hb⇤k
can be written as
@Jbt
@hb⇤k
=MX
i=1
x1⇤i. ⇤BH
eb
ik� (m)x1⇤
k. ⇤BH
rb
k(2.46)
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 24
Considering the correlation constraint with (2.36), we get
@Jbt
@hq⇤k
=MX
i=1
(x(b�q)⇤i
. ⇤BH
1 eb
ik+ x
(b�q+1)⇤i
. ⇤BHeb
ik)
� (m)(x(b�q)⇤k
. ⇤BH
1 rb
k+ x
(b�q+1)⇤k
. ⇤BHrb
k) (2.47)
Here the update process of the TRFs is the same as described in the previous section
with a change in the gradient. Now using the gradient from (2.46) and (2.47) in (2.32),
(3.7) and (2.33), we can calculate µb(m). Next, we update the b-th block TRFs using
(2.28).
The coupling factor (m) in (2.45) should be so chosen that it gives a smaller value
for a higher value of Jb(m) and a larger value for a lower value of J
b(m). This is
due to the fact that initially the value of J b(m) will be high and the gradient due to
noiseless data is dominant. Therefore, initially the coupling factor should be of a small
value to facilitate unconstrained update of the TRF due to the dominant gradient for
the noiseless data. The update equation described in (2.28) makes the cost function
decrease even after the misconvergence point and at this point the gradient due to noise
becomes comparable to signal gradient. For this reason, (m) should be so chosen that
it increases with the decrease of J b(m) for a small value of J b(m). As it crosses the
misconvergence point, the higher value of (m) makes the noise effect compensation
stronger. A suitable expression for the coupling factor is empirically obtained as
(m) = ⇠(|⇢ log10(Jb(m))|)� (2.48)
Here � determines the sensitivity of (m) to decrease in the value of J b(m). A higher
value of � means that (m) will increase highly for a small decrease of J b(m). Fig.
2.3(c) shows that the misconvergence problem is solved after adding the constraint
described in (2.45). Misconvergence phenomenon can also emerge from estimation
noise. For example, as we are using the estimated TRFs of the first block instead
of true TRFs while estimating the TRFs of the second block, actually we are adding
estimation noise to the process. This phenomenon is shown in Fig. 2.3(d). The
misconvergence becomes stronger as the number of blocks increases. Fig. 2.3(e) shows
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 25
that our proposed constraint also works against the estimation noise. Figs. 2.3(c) and
(e) also show that the proposed constraint helps to achieve better NPM than that of
the misconverging point. Fig. 2.4 shows the efficacy of the proposed constraint against
the misconvergence of the bMCFLMS algorithm at different noise level. However, the
final NPM level degrades for decreasing SNR.
The proposed bMCFLMS algorithm is summarized in Table II.
2.4 Results
In this section, the performance of the proposed algorithms is measured on simu-
lation phantom and in-vivo RF data. The results obtained are compared with the
time-domain method described in [13], cepstrum method [16] and CR based method
[21]. Here the performance is measured using the normalized projection misalignment
(NPM) and the resolution gain (RG) [31]. The NPM is defined as
NPM(m) = 20log10
✓k⇣(m)kkhk
◆(2.49)
⇣(m) = h� hTh(m)
hT (m)h(m)h(m) (2.50)
where h and h(m) represent the true and estimated TRFs, respectively. Measurement
of NPM requires the true TRF and hence it can be calculated only for the simulation
phantom data where the true TRF is known. The RG is defined as
Gd =R
o
d
Rd
d
, d = 5 dB, 10 dB (2.51)
where Ro
dand R
d
drepresent resolutions before and after the deconvolution, respectively.
To calculate Ro
dand R
d
d, the normalized 2-D autocovariance function of the RF and
the TRF data are calculated. Then the axial slice through the peak is considered and
the width of the slice at a level d dB is measured which represents Rd
d.
In all the subsequent figures the reference images were generated from the RF data
which have the blurring effect in it introduced by the PSF. To produce the ultrasound
images the envelope of the data was taken from the absolute value of the Hilbert
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 26
Width (mm)
De
pth
(m
m)
0 10 20 30 40
0
10
20
30
40
True TRF Envelope Deconvolved TRF Envelope
De
pth
(m
m)
Width (mm)
0 10 20 30 40
0
10
20
30
40
De
pth
(m
m)
Width (mm)
0 10 20 30 40
0
10
20
30
40
-6
−10
−14
−18
−22
0 3 6 9 12
x 13
0
NP
M (
dB
)
No. of iterations
True
Deconvolved with constraint
d
x 107
-2 e
Frequency (MHz)
No
rma
lize
d M
ag
nitu
de
1
0.8
0.6
0.4
0.2
00 0.40.2 0.6 0.8 1.41.21 1.6 1.8 2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Standard RF Envelope
b ca
Figure 2.5: Deconvolution performed on a simulation phantom with 20 scatterers per
resolution cell and SNR = 30 dB. The RF data size is 1038 ⇥ 128. The darker circu-
lar inclusion with radius 5 mm is created by placing scatterers with relatively lower
strength than the surroundings. Log-envelope image of the (a) true TRF, (b) backscat-
tered standard RF data, and (c) deconvolved TRF by the bMCFLMS algorithm, (d)
spectra of the true and estimated RF data of a single scan-line marked by vertical
green line in the log envelope images of (b) and (c). (e) The NPM curve between the
deconvolved TRF and true TRF of the first block.
transform of the data. Then a log compression was performed on the normalized
envelope data where the dynamic range was set as 35 dB for all the images. The
data used in the result section have axial depth of 40 mm and for this depth a total
block number of 2 was found reasonable. However, the number of total blocks may be
increased which will increase the propagation of estimation errors from the previous
block to the subsequent blocks. Therefore, for all data the total number of block was
set at B = 2 and the parameters of noise effect compensating constraint were set to
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 27
0 5 10 15
x 106
0
0.5
1
Frequency (Hz)
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
7
0
0.5
Frequency (Hz)
No
rm.
ma
gn
itu
de
No
rma
lize
d
ma
gn
itu
de
Zoomed-in view
0 0.5 1 1.5 2
x 107
0
0.5
1
Frequency (Hz)
No
rm.
ma
gn
itu
de
0 10 20 30
0
5
10
15
20
25
30
35
40
Standard Log Envelope
Dep
th, m
m
Deconvolved Log Enveope
Lateral distance, mm
g
300
Lateral distance, mm
10 20
20
x 10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
d
TRF spectruma b
Original RF spectrum
Estimated RF spectrum
c
40 40
e
f
g
Figure 2.6: Performance analysis of the bMCFLMS algorithm for the in-vivo backscat-
tered RF data of a breast cyst. Spectrum of the (a) deconvolved TRF, (b) R-MINT
estimated PSF, and (c) spectra of the true and estimated RF data of a single scan-line
marked by vertical red line in log envelope images. Standard log envelope images of
(d) the backscattered RF data, (e) deconvolved TRFs, and (f-g) zoomed-in views of
(d-e).
⇠ = 1e� 4, ⇢ = 2.55 and � = 2.4.
2.4.1 Simulation Phantom Results
The ultrasound simulation was done using the FIELD-II [32] where the transducer
element height was chosen to be 5 mm. Ultrasound simulation was done on a 3-D
simulation phantom (40 mm ⇥10 mm ⇥40 mm) with a scatterer density of 20 scatterers
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 28
0 10 20 30
0
5
10
15
20
25
30
35
40
0 5 10 15
x 106
0
0.5
Frequency (Hz)
No
rm.
ma
gn
itu
de
0 0.5 1 1.5 2
x 107
0
0.5
1
Frequency (Hz)
No
rm.
ma
gn
itu
de
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x 107
0
0.5
1
Frequency (Hz)
No
rma
lize
d
ma
gn
itu
de
Zoomed-in view
g
i
0 10 20 30
b PSF Spectrum
a TRF spectrum
Original RF spectrum
Estimated RF spectrum
c
Standard Log Envelope
Lateral distance, mm Lateral distance, mm
Deconvolved Log Enveope
1
20
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
4040
d e
g
f
Figure 2.7: Performance analysis of the bMCFLMS algorithm for the in-vivo backscat-
tered RF data for a left carotid artery. Spectrum of the (a) deconvolved TRF, (b)
R-MINT estimated PSF, and (c) spectra of the true and estimated RF data of a single
scan-line marked by vertical red line in log envelope images. Standard log envelope
images of (d) the backscattered RF data, (e) deconvolved TRFs, and (f-g) zoomed-in
views of (d-e).
per resolution cell. A circular inclusion with radius 5 mm was simulated in a slice along
the transverse plane of the phantom by reducing the magnitudes of scatterers in that
region. The focus of the ultrasound beam was set at 30 mm depth from the phantom
surface. The transducer center frequency was selected as 10 MHz and the sampling
frequency as 40 MHz. The total number of scan lines (M) was set to 128 to match with
that of available commercial ultrasound scanners. To simulate noisy data zero-mean
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 29
additive white Gaussian noise was added to the data so as to obtain an SNR of 30 dB.
The performance of the proposed bMCFLMS algorithm on the simulation data is
shown in Fig. 2.5. From visual comparison of the images provided in Figs. 2.5(a)-
(c) we see that our estimated TRF (Fig. 2.5(c)) has finer texture compared to the
standard RF image (Fig. 2.5(b)) and the sharpness of the inclusion boundary has been
restored in the deconvolved TRF image. Fig. 2.5(d) shows that the estimated RF
spectrum, derived from the estimated TRF along the green marked line of the image
and estimated PSF, matches closely with that of the true RF. Again, since we have
the true TRF for the simulation data, we can evaluate NPM for our estimated TRF
which is shown in Fig. 2.5(e). Fig. 2.5(e) proves the effectiveness of our proposed
correlation-based constraint to prevent misconvergence of the proposed algorithm as
the NPM remains stable at �22.1 dB.
Comparative results of different algorithms on simulation phantom data at SNR
= 30 dB are presented in Table 2.3. The quantitative performance measures used for
comparison are NPM (see (2.49)) and Gd (see (2.51)). It is obvious from Table 2.3 that
our proposed bMCFLMS algorithm gives better image quality in terms of resolution
gain (5 and 10 dB level) and NPM than that of the other methods. In order to show
the improvement after using the missing data in our algorithm, results on the final
deconvolved image from the bMCFLMS are presented.
2.4.2 In-Vivo Results
The performance of the proposed algorithms was tested on two in-vivo data. One of
these data is of fibrocystic breast with the largest cyst size of 4.75 mm⇥4.26 mm and
the other one is of a carotid artery. These data were collected from the patients who
appeared for medical examination at the Medical Centre of Bangladesh University of
Engineering and Technology (BUET), Dhaka, Bangladesh. This study was approved
by the Institutional Review Board (IRB) and prior patient consent was taken.
Ultrasound image quality is measured in terms of texture, tissue structure boundary
and size in the reconstructed image [17], [18], [25]. To demonstrate that these features
improve, we present in Fig. 2.6 the performance of our proposed bMCFLMS algorithm
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 30
on the fibrocystic breast data. Fig. 2.6(a) shows the wideband TRF spectrum where
the TRF line is chosen along the red line of Fig. 2.6(e), and Fig. 2.6(b) shows the
narrowband, smooth PSF spectrum. The estimated RF spectrum evaluted from the
multiplication of TRF spectrum and PSF spectrum, matches closely with true RF
spectrum which is shown in Fig. 2.6(c). This ensures the deconvolution efficacy of the
proposed algorithm. From visual perspective, Fig. 2.6(e) shows a significant increase
in texture quality compared to Fig. 2.6(d). For better visualization zoomed-in view
of the two images are provided in Figs. 2.6(f)-(g). It is evident from Figs. 2.6(d)-(e)
that the deconvolved image offers finer texture compared to the blurred RF image.
Again, from the zoomed-in view it is apprent that the partially obscured cyst in the
standard RF image is more prominently visible in the deconvolved image with better
boudary, shape and size. In addition to that, some tiny structures are also visible in the
deconvolved image which are marked by a ellipse in Figs. 2.6(d)-(e). From the visual
inspection of the marked region we can infer that the tiny structures have emerged as
a result of deconvolution that were obscured in the standard RF image.
As in the case of in-vivo data in Fig. 2.6, the overall image quality of the carotid
artery depicted in Fig. 2.7 has also improved. From the zoomed-in view of the carotid
artery in Figs. 2.7(f)-(g), we see that the deconvolved image has finer texture and
comparatively sharp arterial boundary than the RF image. Again, Fig. 2.7(c) shows the
close match between our estimated RF spectrum and the original RF spectrum which
proves the estimation accuracy of the estimated TRF using our proposed algorithm,
and of the PSF estimated using the MINT algorithm.
As shown in Table 2.4, our proposed bMCFLMS algorithm gives significantly better
resolution gain at 10 dB level compared to its time-domain counterpart, CR-based
method and cepstrum method for the in-vivo data. The latter methods also show
data dependency as they give resolution gain improvement in different range for the
breast cyst and the carotid artery data. As discussed in the introduction section, the
effectiveness of cepstrum method depends on the separability of the TRF and the PSF
spectrum. Again, the CR-based method requires thresholding to find the null space
bases of the correlation matrix. However, l1-bMCLMS and our proposed bMCFLMS
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 31
Data Length N in the Axial Direction
Tim
e P
er
Itera
tion (
logari
thm
ic s
cale
)
a
100 200 800700600500400300 900 1000
100
101
102
l1-bMCLMS
md-bMCFLMS
Figure 2.8: Time required per iteration by the time- and the frequency-domain algo-
rithms for the simulation phantom data of size Na ⇥ 128 where Na is made variable.
algorithm do not impose any stringent requirements on the data. Therefore, they
show similar resolution improvement for both type of data which justifies their less
dependence on parameter tuning. The parameters for correlation constraint once set
remain effective for all types of data presented in the result section. Again to investigate
the speckle noise level before and after deconvolution signal-to-noise ratio (sSNR) was
calculated in Table 2.4 as [31]. sSNR is defined as the ratio of µ and �, where µ and �
are the mean and standard deviation of the absolute value of the RF data, respectively.
Insignificant changes in the value of sSNR justify the claim that deconvolution can
significantly improve the resolution though has negligible effect on speckle noise as
reported in [31].
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 32
Table 2.1: Symbols and Description
Symbols Description
Lb Axial length of an RF data block
xbi (n) RF data of the b-th block of the i-th A-line
hbi TRF of the b-th block of the i-th A-line
Cxpi
Convolution matrix of length 2Lb � 1 for the
p-th block of the i-th A-line RF data
epqij Cross-relation error between the p-th block of
the i-th and q-th block of the j-th A-line
epqij Truncated epqij in the frequency domain
m Iteration index
rbJb(m) Gradient of the b-th block cost function with
respect to the estimated b-th block TRF at the
m-th iteration
rJb(m) Gradient of the b-th block cost function with
respect to the estimated total TRF at the m-
th iteration
µb(m) Variable step-size for the update of the b-th
block TRF at the m-th iteration
Jbcorr(m) Cost function for the correlation constraint of
the b-th block at the m-th iteration
Jbt(m) Total cost function for the b-th block including
correlation constraint
(m) Coupling factor for the correlation constraint
at the m-th iteration
⇠, ⇢ and � Empirical constants for the correlation cou-
pling factor (m)
s(n) Estimated PSF using the MINT algorithm
⌫i Scaling factor for the estimated i-th A-line
missing RF data
Jb0(m) Total cost function for the b-th block including
estimated missing data and correlation con-
straint at the m-th iteration
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 33
Table 2.2: Constrained bMCFLMS Algorithm
Step 1 . Set total block number B and appropriate value for ⇠, ⇢ and �
. Initialize the i-th channel TRF, hi = [1 01⇥(L�1)]T for i = 1, 2, · · · ,M
Step 2 . Set current block number b = 1
Step 3 . Set iteration index m = 1
Step 4 . Calculate the error functions for b-th block using (3.6), (3.6), (2.18) and
(2.21)
. Calculate correlation of estimated TRF with RF data using (2.43)
Step 5 . Calculate @Jbt
@hb⇤k
according to (2.46)
. Calculate @Jbt
@hq⇤k
according to (2.47)
. Calculate step-size for b-th block and m-th iteration µb(m) using J
bt instead
of J b in (2.32), (3.7) and (2.33)
Step 6 . Update h using (2.28)
. Normalize h according to (2.37)
Step 7 . If m is less than required iterations, set m = m+ 1 and go to step 4
. Else set b = b+ 1, m = 1 and go to step 3
CHAPTER 2. DECONVOLUTION OF ULTRASOUND IMAGES 34
Table 2.3: Performance of different algorithms on simulation phantom data with addi-
tive noise
Gd(RF)
Data Method NPM (dB) 5 dB 10 dB
CR-based Method �15.13 3.82 4.23
Cepstrum �6.11 1.77 1.78
Simulation data l1-bMCLMS �19.99 4.11 5.14
Proposed bMCFLMS �21.20 4.42 6.03
Table 2.4: Performance of different algorithms on in-vivo data
Gd(RF) sSNR
Data Method 5 dB 10 dB Original Deconv.
Fibrocystic Breast CR-based Method 2.45 1.32 0.7835 0.6788
Cepstrum 1.13 0.72 0.7835 0.7722
l1-bMCLMS 3.45 4.84 0.7835 0.6405
Proposed bMCFLMS 4.00 6.17 0.7835 0.8366
Carotid Artery CR-based Method 3.44 5.03 0.7451 0.7223
Cepstrum 2.44 2.16 0.7451 0.6743
l1-bMCLMS 3.72 5.89 0.7451 0.7564
Proposed bMCFLMS 4.01 7.05 0.7451 0.7129
Chapter 3
B-mode Image Generation Framework
In this chapter, we propose a multiframe-based adaptive despeckling (MADS) algo-
rithm that treats the speckle noise in its multiplicative form and utilizes the speckle
patterns estimated using the multiframe-based adaptive ultrasonic speckle noise esti-
mation (MSNE) algorithm proposed in this thesis. The MSNE algorithm is based on
formulating the true image as single input and the envelope of deconvolved consecutive
US image frames with multiplicative speckle noise pattern in each frame as multiple
outputs. The despeckling algorithm, on the contrary, treats the envelope of deconvolved
consecutive US image frames as multiple inputs and the true image as single output.
According to the mathematical model representing the US imaging system, deconvolu-
tion is necessary prior to despeckling for resolution enhancement of the raw RF data.
Hence, a 2-D deconvolution approach as an extension of our previously published 1-D
deconvolution algorithm, i.e., bMCFLMS has been also described in this chapter. To
prevent misconvergence of the MSNE algorithm in the presence of additive noise and
estimation error resulting from the deconvolution step, a zero-lag correlation contraint
derived from the deconvolved image and the estimated speckle pattern is attached with
the original cost function. As the overall despeckling approach has been derived by
completely following the signal generation models, it is likely to maximally preserve
the diagnostically important details and tissue texture present in the image. Finally, a
complete framework including deconvolution, despeckling, and post-processing for ul-
35
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 36
trasound B-mode image generation with superior quality in terms of resolution, edges,
small details and texture is also established in this chapter for greater interest of the
researchers.
3.1 Despeckling of the Envelope of Deconvolved Data:
Literature Review
A number of algorithms have been reported in the literature for despeckling US images
which attempts to remove speckle noise either in their mulplicative form or converting
the noise in the additive form. Among the first group of algorithms, the spatial averag-
ing based approaches exploit the repetitive nature of the US image, and among them,
linear filtering methods, such as Gaussian filter and mean filter are effective in reducing
speckle noise [2]. However, they tend to oversmooth the texture and blur edges present
in the image. To overcome this problem, nonlinear approaches based on local [33], [34]
and non-local statistics [35], [36] of the image have been proposed. These algorithms
are mainly weighted filters, in which the weights depend on the similarity between the
intensity values of the patches surrounding the pixels [37]. The main difference between
local and non-local means methods is that the non-local means method employs the
most similar pixels in the image to denoise the current pixel regardless of their Eu-
clidean distance. Although these approaches tend to preserve textures and edges, their
performance is dependent on tuning parameters, such as filter and patch size. Again,
the nonlinear approaches based on the diffusion equation [38] not only preserves edges
but also enhances edges by inhibiting diffusion across edges and allowing diffusion on
either side of the edge. However, selection of the parameter-values is a major issue
in this method, as a value of parameter that is smaller than the optimum one leads
to unsatisfactory noise suppression whereas a higher parameter-value results in poor
structure preservation [39].
The second group of algorithms converts the multiplicative speckle noise into an
additive one by using logarithmic transformation. Now, as shown in [1], the additive
noise can be handled using any traditional denoising scheme, and the performance
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 37
of this scheme determines the overall efficacy of despeckling. Finally, the denoised
image is exponentially transformed back to give the despeckled image. The overall
process is termed as homomorphic filtering approach [1], and among the denoising
schemes used in this process, the transfrom domain or multi-resolutional based ap-
praoches [40–43] are of higher efficacy. In [44], an advanced ultrasound despeckling
algorithm is proposed based on the intra-scale correlation between the wavelet coeffi-
cients. Among the multi-resolutional approaches, as shown in [4], despeckling based
on non-subsampled contourlet gives superior performance. However, in wavelet-based
despeckling, a threshold is a critical parameter that is to be determined based on the a
priori knowledge of the distribution of the speckle pattern. In addition, the threshold-
based filtering of wavelet coefficients implies texture smoothening, and it gives rise to
artifact such as Gibbs phenomenon near the edges [45]. All of the approaches discussed
so far relies on ad-hoc filtering or smoothening technique without addressing mathe-
matically the speckle noise generation model. Therefore, it cannot be guaranteed that
these algorithms only operate on the speckle noise without significantly distorting the
true image.
3.2 Method
In this chapter, our main concern is to derive a novel despeckling algorithm that is
likely to preserve maximum features present in the image. However, according to (1.1)
and (1.2), deconvolution is a necessary pre-processing step to despeckling. Therefore,
for the completeness of a high-resolution B-mode image generation, we also consider
it important to include a 2-D extension of our previously proposed 1-D deconvolution
(bMCFLMS) algorithm.
3.2.1 Deconvolution of RF Echo Data
In this section, we attempt to estimate the TRF, h(m,n), in (1.1) with increased
resolution by removing the effect of the PSF, s(m,n), from the raw RF data, x(m,n).
However, to estimate h(m,n), similar to the approach as described in [31], we consider
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 38
s (m)
h1(m)
h2(m)
hN
(m)
x1(m)
x2(m)
xN
(m)
v1(m)
v2(m)
vN
(m)
PSF
TRFBackscattered
RF Data
Additive
Noise
a
a
a
a
‘
‘
‘
Figure 3.1: A SIMO model for backscattered RF signal.
the 2-D distortion kernal, i.e., PSF s(m,n) in (1.1) decomposible into two 1-D distortion
kernals (PSFs): one along the axial direction and the other along the lateral direction.
Following this assumption, (1.1) can be modified as
x(m,n) = sa(m) ⇤a sl(n) ⇤l h(m,n) + v(m,n) (3.1)
where sl(n) and sa(m) are the lateral and axial PSFs, respectively, and ‘⇤l’ and ‘⇤a’represent convolution along the lateral and the axial directions, respectively. A novel
technique for removing the effect of axial PSF sa(m) from the measured RF image
x(m,n) was reported in the previous chapter using a single input multiple output
(SIMO) model as shown in Fig. 3.1, where sa(m) convolves with the i-th A-line of the
axial TRF denoted as ha
i(m) and with additive noise vi(m) gives the i-th A-line RF
data xi(m):
xi(m) = sa(m) ⇤a ha
i(m) + vi(m) (3.2)
where
ha
i(m) = sl(n) ⇤l h(m,n) (3.3)
In matrix form, (3.2) can be written as
X = Saha + v (3.4)
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 39
where Sa is the convolution matrix formed using the axial PSF sa(m) and
X =hX1 X2 · · · XN 0
i,
ha =
hha
1 ha
2 · · · ha
N 0
i.
Here, N 0 is the total number of A-lines, and Xi and ha
iare the i-th A-line with M
0
samples taken from xi(m) and ha
i(m), respectively where m = 1, 2, · · · ,M 0. To account
for the non-stationarity of the axial PSF, the RF data were divided into B blocks with
equal length Lb, and a block-based cost function Jb for the b-th block was formulated
(for details see [46]) to estimate the axial TRF block-by-block in the frequency-domain
as
Jb =
n�1X
i=1
nX
j=i+1
ebH
ijeb
ij(3.5)
where, ‘H’ denotes the Hermitian operation, any variable with ‘ ’ represents the vari-
able in the frequency-domain, and eb
ijis the Fourier transform of eb
ijdefined as
eb
ij=
b�1X
p=1
A1ep(b�p)ij
+bX
p=1
A2ep(b�p+1)ij
Here,
ep(b�p)ij
= Cxpiha(b�p)j
�Cxpjha(b�p)i
, p = 1, 2, · · · , b� 1,
ep(b�p+1)ij
= Cxpiha(b�p+1)j
�Cxpjha(b�p+1)i
, p = 1, 2, · · · , b,
and A1, A2 are the truncation matrices truncating the last (Lb � 1) and the first Lb
samples of the error function, respectively. Cxpi
is the convolution matrix formed using
the RF data, x(m,n) along the i-th A-line and the p-th block. Now, the b-th block
axial TRF hab was estimated as
hab
= arghab min J b, subject to ||h
a
|| = 1 (3.6)
where ‘|| · ||’ denotes the l2-norm and
ha
=hha1T
ha2T
· · · haBTiT
(3.7)
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 40
with ‘T ’ denoting matrix transpose operation and
hab
=hhab
1 hab
2 · · · hab
N 0
i(3.8)
Here, hab
idenotes the estimated axial TRF along the i-th A-line and the b-th block.
In sample-domain, the estimated axial TRF along the i-th A-line can be written as
ha
i(m). Now, the estimated axial TRF hi(m) along the i-th A-line at discrete time m
can be modeled as
ha
i(m) = sl(n) ⇤l h(m,n) + v
0(m,n) (3.9)
where v0(m,n) is the noise resulting from the estimation error of the axial TRF. There-
fore, from (3.2) and (3.9), it is apparent that an attempt, similar to the approach
adopted in the axial direction, can be made in the lateral direction to undo the effect
of the lateral distortion kernal (PSF) from the estimated axial TRF ha
i(m) to estimate
the 2-D deconvolved TRF h(m,n). In this approach, the PSF is considered to be later-
ally stationary as described in [25], and therefore, no blocking is required in the lateral
direction. The method is summarized in Table 3.1. The estimated TRF after lateral
deconvolution is given by
h =hh1 h2 · · · h
M 0
iT(3.10)
where, hiis the estimated lateral TRF along the i-th sample line, i.e., samples along
the i-th row of h. In sample-domain, the m-th row and n-th column sample of the
estimated TRF, h can be denoted as h(m,n).
3.2.2 Proposed Despeckling Algorithm
The envelope of the estimated TRF h(m,n) in the previous subsection is corrupted
with speckle noise as given by (1.2). It is apparent from (1.2) that knowledge of the
speckle pattern u(m,n) in an image frame can help despeckling that frame. However, in
the absence of additive noise, direct division of he(m,n) by u(m,n) may amplify noise
and/or give rise to division by zero problem. Hence, in this subsection, we attempt to
formulate an energy constrained iterative approach to find an equalization multiplying
factor to despeckle the frame. In what follows, we attempt to formulate a novel SIMO
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 41
Table 3.1: bMCFLMS algorithm for 2-D deconvolution of ultrasound RF image
Step 1 . Select the bMCFLMS method reported in [46]
Step 2 . Set X = raw radio-frequency (RF) data, where X is the data to be decon-
volved
. Set the data length, L = axial length, and the number of channels, M =
lateral length of the raw RF data. Execute the bMCFLMS algorithm with
block number, B = 2 along the axial direction.
Step 3 . Set X = axially deconvolved data in step 2
. Set the data length, L = lateral length, and the number of channels, M =
axial length of the raw radio-frequency (RF) data. Execute the bMCFLMS
algorithm with block number, B = 1 along the lateral direction.
model for the deconvolved image frames and thereby, estimate the speckle noise in
the respective deconvolved frames using an adaptive filtering technique. Then a novel
MISO model will be proposed to despeckle the decconvolved image frames using the
estimated speckle pattern of the respective frames.
w (m,n)
Additive
Noise
Multiplicative
Speckle Noise
u
u
u
1
2
p
(m,n)
(m,n)
(m,n)
r(m,n)
h (m,n)
h (m,n)
h (m,n)
e1
e2
ep
Decon
volv
ed
data
1
2
p
w (m,n)
w
(m,n)
True Image
Figure 3.2: A new SIMO model for deconvolved 2-D RF data.
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 42
In an ultrasound imaging system, images are generally acquired at a frame rate
ranging from 10�60 frames per second (FPS) with speckle patterns generated randomly
in each image frame from the interference of the US pulse at the receiving transducer.
In our work, we attempt to use p consecutive deconvolved frames to formulate a single
input multiple output (SIMO) model as shown in Fig. 3.2. Here, the true ultrasound
image r(m,n) that is considered stationary throughout the p frames, multiplies with
the speckle noise ui(m,n) of the i-th frame, and with an additive noise wi(m,n) gives
the envelope, hei(m,n), of the estimated deconvolved image hi(m,n) of the i-th frame:
hei(m,n) = r(m,n)ui(m,n) + wi(m,n) (3.11)
In matrix form, (3.11) can written as
Hei = R · ⇤Ui +Wi (3.12)
where ‘·⇤’ denotes elementwise multiplication. Here, hei(m,n), r(m,n), ui(m,n) and
wi(m,n) represents the m-th row and n-th column elements of Hei, R, Ui and Wi,
respectively. Therefore, here the challenge is to estimate the speckle noise ui(m,n)
from each frame in the presence of additive noise wi(m,n) and then remove the speckle
noise from the deconvolved envelope image hei(m,n). The assumptions behind the
SIMO model formulation and the identifiability condition [47] for the speckle noise
pattern ui(m,n) of the i-th frame are
1. The true image r(m,n) is stationary throughout the p consecutive frames.
2. The speckle patterns of each frame do not share common zeroes with the rest
p� 1 consecutive frames.
These assumptions are realistic because for an ultrasound video recording with 30 FPS,
consecutive 5�10 frames take around 0.17�0.33 second during which the hand motion
can be ignored. Then we can consider the true ultrasound image r(m,n) as stationary
throughout these frames. Again, formation of speckle noise in consecutive frames is
a completely random phenomena, and hence, they are unlikely to contain common
zeros. However, the probablity of sharing common zeros between the speckle patterns
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 43
of the consecutive frames can be further reduced by increasing the number of fames
into consideration; doing this will improve the identification accuracy of the patterns
provided that the stationarity assumption remains valid as shown in the result section
later. In what follows, we derive a multiframe-based adaptive speckle noise estimation
algorithm using the proposed SIMO model.
Speckle Noise Estimation
In the absence of additive noise, the following error function eij(m,n) can be used to
estimate the speckle noise:
eij(m,n) = hei(m,n)uj(m,n)� hej(m,n)ui(m,n) (3.13)
where ui(m,n) is the estimated speckle noise of the i-th frame. Notice that for additive
noiseless case if we can estimate the speckle pattern accurately, the error function
defined in (3.13) becomes zero. Using this fact, we can build the following cost function
to iteratively estimate the speckle noise:
J =p�1X
i=1
pX
j=i+1
||Eij · ⇤Eij||2F (3.14)
where
Eij = Hei · ⇤Uj � Hej · ⇤Ui (3.15)
and ‘|| · ||F ’ indicates the Frobenius norm. An estimate of the speckle noise U can be
obtained by minimizing the cost function J as
U = argU min J, subject to ||U||F = 1 (3.16)
where
U =hU1 U2 · · · Up
i(3.17)
Taking the gradient of J in (3.14), we get
rkJ =@J
@Uk
= 2pX
i=1
Hei. ⇤ Eik (3.18)
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 44
The update equation for the MSNE algorithm at the q-th iteration is
U(q + 1) = U(q)� µ(q)rJ(q)��U=U(q)
(3.19)
where,
rJ(q) =@J(q)
@U
=hr1J(q) r2J(q) · · ·r pJ(q)
i(3.20)
and µ(q) is the variable step-size (VSS) which is such that the misalignment of U(q+1)
with the true noise pattern U is minimum at every iteration, given the current estimate
U(q):
Jµ(q) = (||U� ↵U(q + 1)||F )2|U(q)
=����U · ⇤U� 2↵U · ⇤U(q) + 2↵µ(q)U ·⇤r J(q)
+ ↵2U(q) · ⇤U(q)� 2↵2
µ(q)U ·⇤r J(q)
+ ↵2µ2(q)rJ(q) ·⇤r J(q)
����S|U(q) (3.21)
where ↵ is a scaling constant inherent in any blind channel identification approach
based on the cross-relation, and we define an operator ‘|| · ||S’ which evaluates the
sum of the matrix elements. Minimizing (3.21), i.e., setting the gradient of Jµ(q) with
respect to µ(q) to zero, we get µ(q) as
µ(q) =
����U(q) ·⇤r J(q)����S����rJ(q) ·⇤r J(q)����
S
(3.22)
Equation (3.22) can be considered as a variant of VSS derived in [12].
So far additive noise has been ignored in the derivation of the proposed MSNE
algorithm. However, it has a similar effect on the convergence of the MSNE algorithm
as described in [46]. To solve the problem, we need to impose a constraint on (3.14)
so as to prevent the deviation of the estimated speckle pattern from the true speckle
pattern. To this end, consider the following model of speckle corrupted image for the
k-th frame as described in [35]:
hek(m,n) = r(m,n) + r⇣(m,n) Vk(m,n) (3.23)
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 45
where Vk(m,n) ⇠ N (0, �2). Ignoring the additive noise in (1.2) and comparing with
(3.23), our proposed algorithm is basically estimating the true speckle pattern as
uk(m,n) = 1 + r(⇣�1)(m,n) Vk(m,n) (3.24)
Then the speckle pattern uk(m,n), estimated using the proposed method, can be ex-
presses as
uk(m,n) = 1 + r(⇣�1)(m,n) Vk(m,n) (3.25)
Since r(m,n) is not a variable here, to make the estimated speckle pattern uk(m,n)
in (3.25) close to the true speckle pattern uk(m,n) in (3.24), we need to maximize
the zero-lag correlation between Vk(m,n) and Vk(m,n). However, to attach this as
a constraint on (3.14), estimates of the true image r(m,n) and the parameter ⇣ are
necessary. Alternatively, consider the zero-lag correlation between the deconvolved
image hek(m,n) and the estimated speckle pattern uk(m,n):
Jcorr =����Hek · ⇤Uk
����S
=����(Hek +Wk) · ⇤Uk
����S
=����Hek · ⇤Uk
����S
(3.26)
where the zero-lag correlation between the additive noise Wk and speckle noise Uk is
considered zero. Using (3.23) and (3.25) in (3.26), we get
Jcorr =����R+R
⇣ ·⇤V k +R⇣ · ⇤Vk + Vk · ⇤Vk
����S
=����Hek +R
⇣ · ⇤Vk + Vk · ⇤Vk
����S
= c+����Vk · ⇤Vk
����S
(3.27)
where at the point of misconvergence when the estimated speckle pattern is close to
the true speckle pattern, the zero-lag correlation between r⇣(m,n) and Vk(m,n) can be
considered zero, and c is a constant defined as c =����Hek
����S. From (3.27), it is apparent
that maximizing Jcorr or equivalently minimizing �Jcorr is analogous to maximizing the
zero-lag correlation between Vk(m,n) and Vk(m,n). To prevent the misconvergence of
the proposed algorithm in the noisy case, we propose to use the zero-lag correlation
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 46
constraint in (3.27) with the MSNE cost function in (3.14). Then modifying (3.14) for
noisy case, we obtain
Jc(q) = J(q)� �1Jcorr(q) (3.28)
where �1 is the Lagrange multiplier or also known as the coupling factor. Taking
gradient of (3.28) with respect to Uk, we get
rkJc(q) = rkJ(q)� �1Hek (3.29)
Replacing rkJ(q) in (3.18) by rkJc(q), we can estimate the speckle noise pattern in
each of the p frames.
Estimation of the True Ultrasound Image
So far, a novel algorithm for estimating the speckle pattern has been explained with
a view to estimating the true ultrasound image r(m,n) in (1.2) using U from (3.16).
Now, we describe a novel MISO model as shown in Fig. 3.3, where the speckle noise
cancellation (SNC) factors for each of the i-th frame gi(m,n) multiplies with hei(m,n),
to obtain an estimate of the true US image ri(m,n) for that frame. In the absense of
additive noise in (3.11) and estimation error in the estimated speckle pattern Ui, the
estimated ultrasound image of the i-th frame, Ri, can be obtained in matrix form as
g (m,n)
g (m,n)
g (m,n)
h (m,n)
r(m,n)
e1
1
2
p
Envelope ofDeconvolved
RF data
Speckle Noise Cancellation
EstimatedImage
h (m,n)e2
h (m,n)ep
Figure 3.3: The block diagram for the estimation procedure of true ultrasound image.
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 47
Ri = Hei · ⇤Gi (3.30)
= Ri · ⇤Ui · ⇤Gi (3.31)
where Gi is the elementwise multiplying SNC factor to equalize the speckle pattern
for the i-th frame. Here for an accurate estimation of Ri, the elements of the matrix
Ui. ⇤Gi in (3.31) should be equal to 1. Therefore, we can formulate the following cost
function to estimate the SNC factors:
Jeq = ||U · ⇤G�D||2F
(3.32)
where
G =hG1 G2 · · · Gp
i(3.33)
and D is a matrix with all entries equal to 1. Taking gradient of (3.32) with respect
to G, we get
rJeq =@Jeq
@G= 2(U. ⇤G�D). ⇤U (3.34)
The update equation for estimating G at the q0-th iteration is given by
G(q0 + 1) = G(q0)� µ(q0)rJeq(q0) (3.35)
where µ(q0) is the VSS for the q0-th iteration, and can be obtained following (3.22) as
µ(q0) =
����G(q0) ·⇤r Jeq(q0)����S���� rJeq(q0) ·⇤r Jeq(q0)����S
(3.36)
Up to now, we have ignored the effect of estimation error in U. If we assume that Es
be the estimation error in U, then we can write
U = U+ Es (3.37)
Now, replacing the true speckle pattern U with the estimated speckle pattern U, the
cost function in (3.32) becomes
Jeq = ||U. ⇤G�D||2F
(3.38)
= ||(U+ Es). ⇤G�D||2F
= ||U. ⇤G�D||2F+ E (3.39)
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 48
where E represents the terms including the estimation error Es. Therefore, the gradient
in (3.34) will have two components– one from the desired part of the cost function and
the other from the estimation error part, i.e.,
rJeq(q0) = rJ
desired
eq(q0) +rJ
error
eq(q0) (3.40)
From (3.38), we observe that U and G have inverse relation so as to make their elemen-
twise product equal to 1. Therefore, in (3.37), the term associated with the estimation
error in U gives rise to an equalization factor G in which the small estimation error
is magnified. To make G less sensitive to such phenomenon, we impose an energy
regularization constraint on G in (3.32):
J0eq= ||U · ⇤G�D||2
F+ �2||G||2
F(3.41)
where �2 is the Lagrange multiplier. Now, the gradient in (3.34) becomes
rJ0eq= 2(U · ⇤G�D) · ⇤U+ 2�2G (3.42)
In addition to the energy regularization constraint, we propose the following gradi-
ent averaging technique to average out or at least reduce the detrimental effect of
rJerror
eq(q0):
rJ00eq(q0) = ↵rJ
0eq(q0) + (1� ↵)J 0
eq(q0 � 1) (3.43)
where ↵ is the weighting factor given on the current gradient rJeq(q0). Now, using the
average gradient rJ00eq(q0) in (3.35) and (3.36) in place of rJeq(q0), we can estimate
the equalization factor G which can then be used to get an estimate of the estimated
true image of the i-th frame, Ri, using (3.30). Finally, averaging the estimates for
i = 1, 2, · · · , p for SNR improvement, the true image is reconstructed as
R =pX
i=1
Ri (3.44)
The additive noise in (3.11) has been ignored so far. However, considering the
additive noise in (3.30), we get
Ri = (R · ⇤Ui +Wi) · ⇤Gi
= R0i+W
0i
(3.45)
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 49
where W0iis the modified additive noise in the i-th frame. From (3.44) and (3.45), we
can write
R = R0 +W
0 (3.46)
Due to superior performance of non-subsampled shearlet transform (NSST) to capture
the geometric and mathematical properties of an image such as scales, directionality,
elongated shapes and oscillations as described in [48], we attempt to denoise the im-
age R with a hard-thresholding on the NSST co-efficients following a similar method
as described in [49]. Here, W0 is somewhat minimized due to averaging. However,
according to [1], [8], [50], the effect of speckle noise is more pronounced compared to
the additive noise and hence, the SNR for the additive noise can be considered high in
the despeckled image. In this approach, the estimated true image is decomposed into
4 levels with each having 3, 3, 4 and 4 directions. The coarse scales of the NSST co-
efficients are not thresholded, but the finest scale is hard-thresholded using a tunable
low threshold value.
3.3 Post-processing
To match with the characteristics of the display monitor and control the overall bright-
ness of the image, further post-processing like gamma correction [51] is necessary. The
gamma correction [52] of the estimated true image R is done using
I(m,n) =
r(m,n)
max(R)
!�
(3.47)
Finally, to control the image contrast, gray level transformation [53] of the image
I(m,n) is done as
G(m,n) =
8>>>>>>><
>>>>>>>:
0, if G(m,n) < Wlow
(I(m,n)/max(I)�Wlow)(Whigh�Wlow) ,
if Wlow G(m,n) < Whigh
1, otherwise
(3.48)
CHAPTER 3. B-MODE IMAGE GENERATION FRAMEWORK 50
where Wlow and Whigh are tunable parameters (intended for tuning the contrast of the
image) such that
Wlow < Whigh < 1
The complete framework of the proposed ultrasound image enhancement and recon-
struction process is depicted in Fig. 3.4.
Raw Image
Data
bMCFLMS
along Axial
Direction
bMCFLMS
along Lateral
Direction
Deconvolution
Despeckling
N
ois
e
Est
imati
on
Speckle Noise
Removal
Gamma Correction
&
Gray Level
Transformation
Image
Post-processings
B-mode
Image
True
US Image
Estimation
MS
NE
Figure 3.4: Block diagram of the proposed ultrasound image reconstruction method
from the raw RF data.
Chapter 4
Results
In this section, the efficacy of our proposed framework for high-resolution B-mode im-
age generation is evaluated on both the simulation and in-vivo data. The contents of
the thesis cover: deconvolution, despecking via MSNE, and post-processing for a com-
plete B-mode image generation. However, as the main contribution is the despeckling
algorithm, the simulation study is designed to show the effectiveness of our proposed
method for speckle noise estimation and despeckling only. The convergence of the
algorithm in the presence of additive noise is also shown to justify the use of the pro-
posed constraint. On the other hand, the in-vivo images suffer from low resolution and
speckle noise arising from the physical phenomena related with the US image acquisi-
tion system. Therefore, according to the signal generation models, the in-vivo study
includes the 2-D deconvolution to enhance the resolution as the first phase for all the
methods involved for comparing despeckling performance. Two types of deconvolu-
tion methods, namely– bMCFLMS and cepstrum [54] are investigated. Finaly, post
processing stage is included to make a complete investigation of the high-resolution
B-mode image generation pipeline in a single thesis.
The quality of the despeckled image is compared with those of SRAD (speckle
reducing anisotropic diffusion filter) [38] and OBNLM (optimized Bayesian non-local
means-based filtering) [35] methods. The performance matrices, used in this case, are
SNR (signal-to-noise ratio), PSNR (peak signal-to-noise ratio), SSIM (structural sim-
51
CHAPTER 4. RESULTS 52
ilarity index measure) [55], and NIQE (natural image quality evaluator) [56]. Among
these indices SNR, PSNR, SSIM require reference image and hence, cannot be used
for in-vivo data. In the absence of a reference image, NIQE and visual evaluation are
the only ways to evaluate the performance of the proposed algorithm. On the other
hand, in case of simulation data with the original noiseless image at hand, we attempt
to build the intuition behind different aspects, i.e., the number of image frames to
be chosen, runtime, efficacy in preserving small details, of the proposed despeckling
algorithm.
(f )(e)(d)
(a) (b) (c)
Figure 4.1: Effect of the number of frames in despeckling the modified Shepp-Logan
phantom image using the proposed algorithm: (a) clean phantom, (b) noisy phantom
(� = 0.4), (c)-(f) despeckled using 5,10, 15 and 20 frames, respectively.
4.0.1 Simulation Data
Simuation data were generated using the ‘Modified Shepp-Logan’ phantom available
in MATLAB with size 256 ⇥ 256. In the simulation study, we have investigated the
despeckling efficacy of the proposed framework and have not considered the PSF effect
on the image. Hence, the image was corrupted with speckle noise only as described
in [35] following (3.23) where V(m,n) ⇠ N (0, �2) and as in [57], ⇣ = 0.5 was used.
CHAPTER 4. RESULTS 53
(a) (b)
(d) (e) (f )
(g)
(c)
(i)
NP
M(d
B)
-22
-20
-18
-16
-14
20 40 60 800Iteration
NP
M(d
B)
-15
-14
-13
-11
-12
20 40 60 800
(h) Iteration
Figure 4.2: Despeckling of Shepp-Logan phantom image corrupted by synthetic speckle
noise. (a) clean phantom, (b) noisy phantom (� = 0.4), (c) true speckle noise in the
5th frame; despeckled image using (d) SRAD, (e) OBNLM, (f) proposed MSNE; (g)
extracted noise from the 5th frame using MSNE; (h) NPM measure between the true
and estimated noise using MSNE without constraint; (i) NPM measure between the
true and estimated noise using MSNE with constraint.
The level of noise was varied by setting � = {0.2; 0.4; 0.8}. At a particular noise level,
the speckle pattern was varied using different ⌫(m,n) patterns with the same distribu-
tion for different frames. Among the frames used for despeckling using our proposed
method, the last frame was despeckled using SRAD and OBNLM for comparison with
the proposed algorithm. Here, the implementation platform used were: CPU: IntelR�
CoreTM i7-8700K, RAM: 32 GB, software: MATLAB R�, The MathWorks, Natick, MA.
Comparing (1.2) and (3.23) in additive noiseless case, we can write the speckle pattern
CHAPTER 4. RESULTS 54
of the k-th frame as
uk(m,n) = 1 + r�0.5(m,n)⌫k(m,n) (4.1)
Therefore, (4.1) can be used to calculate the true speckle pattern for the simulation data
that can be used to quantify the performance of the proposed despeckling algorithm.
As claimed in Section 3.2.2, increasing the number of image frames in the proposed
speckle pattern estimation algorithm has an impact on the accuracy of estimation. In
what follows, we attempt to establish a suitable frame number that optimally meets
all the assumptions made in Section 3.2.2, consumes less runtime, and produces a
visually pleasant despeckled image. The performance index used in this case is NPM
(normalized projection misalignment) defined as
NPM(q) = 20log10✓k⇢(q)kkUk
◆dB (4.2)
⇢(q) = U� UTU(q)
UT (q)U(q)U(q) (4.3)
A lower value of NPM indicates better estimation of U. From Table 4.1, it is apparent
that increasing the level of noise deteriorates the estimation accuracy of the speckle
pattern. However, it can be improved by around 6 dB for the three different noise lev-
els as mentioned above by increasing the number of image frames from 5 to 20 in the
proposed algorithm. As described in [47], for an accurate estimation using the blind
multichannel algorithm, the channels should not have common zeros. As we introduce
more image frames in the estimation process, the probability of having common zeros
decreases. This in turn improves the speckle estimation accuracy. However, increasing
the number of image frames imply more computational complexity leading to higher
runtime, and at the same time, it causes the violation of the quasi-stationarity assump-
tion for the true image as described earlier in Section 3.2.2. The despeckled images
using different number of frames for noise level � = 0.4 are shown in Fig. 4.1. It is
apparent from this figure that consideration of frame number greater than 10 results
in visually imperceptible change in the despeckled images. Therefore, in Fig. 4.2,
we have used 10 image frames to compare our simulation phantom results with other
algorithms.
CHAPTER 4. RESULTS 55
Table 4.1: Simulation results on the estimation accuracy in terms of NPM (dB) of
speckle pattern using the proposed method for different noise levels
NPM (dB)
Number of frames Noise level Runtime (sec)
� =0.2 � =0.4 � =0.8
5 �31.65 �25.69 �19.96 0.61
10 �34.69 �28.72 �22.96 2.01
15 �36.41 �30.45 �24.68 3.95
20 �37.64 �31.64 �25.83 6.63
Width (mm)
De
pth
(m
m)
0 10 20 30 40
0
10
20
30
40
De
pth
(m
m)
Width (mm)
0 10 20 30 40
0
10
20
30
40
Depth
(m
m)
Width (mm)
0 10 20 30 40
0
10
20
30
40
(a) (c)(b)
(d) (e) (f )
Figure 4.3: Deconvolution of ultrasound images using adaptive bMCFLMS algorithm.
(a) Raw RF image, (b) 1-D deconvolved image, (c) 2-D deconvolved image, (d)-(f)
zoomed-in views of image segments of (a)-(c), respectively.
The performnce of the proposed algorithm in comparison to others is illustrated
in Fig. 4.2. From Fig. 4.2(d), we can deduce that SRAD distorts the texture in the
homogeneous region and blurs the small details in the phantom. Again, as evident
from Fig. 4.2(e), although OBNLM is superior in performance compared to SRAD
in preserving texture and edges, it fails to remove speckle noise completely when the
CHAPTER 4. RESULTS 56
Table 4.2: Performance measures computed for the simulation study with different
noise level (�) using diffetrent despeckling approaches
Methods Noise level SNR PSNR SSIM NIQE
(�) (dB) (dB)
0.2 20.33 24.98 0.9996 7.55
OBNLM 0.4 14.24 19.35 0.9985 11.03
0.8 10.45 15.94 0.9969 13.17
0.2 12.97 18.28 0.9980 7.45
SRAD 0.4 9.32 15.04 0.9954 10.06
0.8 7.89 13.62 0.9935 12.97
0.2 28.88 33.22 0.9999 6.88
Proposed 0.4 24.30 28.79 0.9998 6.21
0.8 21.63 26.23 0.9997 6.10
noise level is high, e.g., � = 0.4. On the other hand, our proposed algorithm (see
Fig. 4.2(f)) shows significant visual improvement in terms of maintaing original tex-
ture, edges and small details compared to the SRAD and OBNLM approaches, and
the despeckled image is visually close to the clean phantom image. Again, the quanti-
tative metrics as presented in Table 4.2 also demonstrate that our proposed algorithm
performs significantly better in terms of quantitative indices SNR, PSNR, SSIM, and
NIQE at different noise levels compared to SRAD and OBNLM. To show the effect of
the zero-lag correlation constraint on the convergence profile of the proposed MSNE
algorithm, we depict in Fig. 4.2(h)-(i) the NPM curve with 20 dB SNR. As can be
seen in 4.2(h), the algorithm misconverges near �16 dB, whereas in 4.2(i), there is no
sign of misconvergence and the constrained MSNE algorithm smoothly convergences to
around NMP= �22.96 dB implying that the zero-lag correlation constraint is effective
in preventing misconvergence.
CHAPTER 4. RESULTS 57
Table 4.3: Axial and lateral correlation energy for raw RF, 1-D and 2-D deconvolved
data using the b-MCFLMS algorithm.
Data Correlation energy
Axial Lateral
RF 0.0379 0.0501
1-D deconvolved 0.0287 0.0497
2-D deconvolved 0.0279 0.0329
4.0.2 In-Vivo Data
Performance of the proposed complete framework for US B-mode image generation
comprising of deconvolution, despeckling and post-processing, respectively, is evalu-
ated on the in-vivo data, collected using a commercial SonixTOUCH Research (Ultra-
sonix Medical Corporation, Richmond BC, Canada) scanner integrated with a linear
array transducer, L14-5/38, operating at 10 MHz with sampling frequency of 40 MHz.
These data were collected from the patients who appeared for medical examination at
the Medical Centre of Bangladesh University of Engineering and Technology (BUET),
Dhaka, Bangladesh. This study was approved by the Institutional Review Board (IRB),
and prior patient consent was taken.
The performance evaluation of the deconvolution step is done subjectively as elab-
orate performance evaluation of the 1-D bMCFLMS algorithm is already done in our
published work [46]. However, the performance of the complete framework is eval-
uated using two approaches. First, we keep the deconvolution step fixed for all the
despeckling algorithms to be compared with the proposed MADS algorithm and eval-
uate their comparative performance both from visual and quantitative perspectives.
Second, we compare subjectively the image generated using our complete framework
with the data acquiring machine B-mode image. All the in-vivo images shown here
were log compressed and dynamic range was set to 35 dB as described in [46] for display
purpose.
CHAPTER 4. RESULTS 58
Width (mm)
Depth
(m
m)
0 10 20 30 40
0
10
20
30
40
Width (mm)
0 10 20 30 40
0
10
20
30
40
Width (mm)
0 10 20 30 40
0
10
20
30
40
Width (mm)
Depth
(m
m)
0 10 20 30 40
0
10
20
30
40
Width (mm)
0 10 20 30 40
0
10
20
30
40
Width (mm)
0 10 20 30 40
0
10
20
30
40
(a) (b) (c)
(d) (e) (f )
Figure 4.4: Despeckling of breast ultrasound image- 1. (a) Deconvolved image, images
obtaind using (b) SRAD, (c) OBNLM, (d) proposed algorithm, (e) machine B-mode
image, (f) estimated speckle pattern of the 5-th frame.
2-D Deconvolution Performance Evaluation on In-Vivo Data
The successive stages of deconvolution offer images with improved and finer texture
as shown in Figs. 4.3(a)-(c) and the zoomed-in views of their marked portion in Figs.
4.3(d)-(f), respectively. The speckle pattern in Fig. 4.3(d) is blurry and highly auto-
correlated in the spatial domain as convolution of point speckle with the US PSF results
in the spreading of the point spatially and thereby, reduces the resolution of the image.
This large spatial coverage of speckle leads to considerable correlation between speckle
noise not only in the same frame but also in the consecutive frames leading to a greater
number of common zeros. To justify our claim, we select an axial and a lateral line
from two consecutive frames along the marked lines in Figs. 4.3(a)-(c) and calculate
the energy of the normalized correlation among two axial lines as well as lateral lines
of two consecutive frames as presented in Table 4.3. Higher value of correlation energy
CHAPTER 4. RESULTS 59
Width (mm)
De
pth
(m
m)
0 10 20 30 40
0
10
20
30
40
Width (mm)
0 10 20 30 40
0
10
20
30
40
Width (mm)
0 10 20 30 40
0
10
20
30
40
Width (mm)
De
pth
(m
m)
0 10 20 30 40
0
10
20
30
40
Width (mm)
0 10 20 30 40
0
10
20
30
40
Width (mm)
0 10 20 30 40
0
10
20
30
40
(a) (b) (c)
(d) (e) (f )
Figure 4.5: Despeckling of breast ultrasound image- 2. (a) Deconvolved image, images
obtaind using (b) SRAD, (c) OBNLM, (d) proposed algorithm, (e) machine B-mode
image, (f) estimated speckle pattern of the 5-th frame.
indicates higher correlation among the two frames along the axial or the lateral direc-
tion. From Table 4.3, it is evident that the raw RF data of a particular frame is more
correlated with the next frame in both the axial and the lateral directions compared to
those of the 1-D and the 2-D deconvolved data. However, the 1-D deconvolved data has
higher lateral correlation compared to that of the 2-D deconvolved data with nearly
the same axial correlation (see Table 4.3). And in Fig. 4.3(e), the speckle pattern
becomes fiber-like with greater spatial coverage along the lateral direction than the ax-
ial direction leading to higher correlation with the next frame in the lateral direction.
Finally, in Fig. 4.3(f), the lateral correlation is sufficiently removed, and the speckle
pattern becomes randomly distributed and uncorrelated. As the speckles in the final
2-D deconvolved image occupies lesser space, this has additional advantage of reducing
common zeros between the speckle patterns in consecutive frames along with increasing
CHAPTER 4. RESULTS 60
resolution of the image. Hence, the speckle pattern estimation can be done efficiently
with 5 � 10 number of frames without violating the quasi-stationarity assumption of
the true image.
Despeckling Performance on in-vivo Data
Table 4.4: NIQE measure for images despeckled with different algorithms
OBNLM SRAD Proposed
Image- 1 7.38 7.69 5.16
Image- 2 7.61 8.49 5.32
We have used two in-vivo breast images with one containing a solid mass (RF
image-1) and the other containing a cyst (RF image-2) to validate the performance
of our proposed algorithm. Similar type of post-processing with � = 0.97 in (3.47)
and Wlow = 1e � 2,Whigh = 0.98 in (3.48) were set for all the images despeckled
with different algorithms for illustration purpose. The performance index used here
is NIQE that relies on the deviation from the statistical regularities of distortionless
images to rate an image as defined in [37], [56]. The lower the value of the NIQE
metric, the better the quality of the despeckled image. However, the NIQE index was
measured on the despeckled image without post-processing. As shown in Table 4.4,
our proposed algorithm gives the lowest NIQE score and hence, the best quality image
compared to that of SRAD and OBNLM. Figs. 4.4 and 4.5 are provided for subjective
evaluation of the proposed algorithm. As shown in Figs. 4.4(b) and 4.5(b), SRAD
algorithm succeeds in preserving edges although it blurs the texture and degrades
the contrast of the image. On the other hand, according to Figs. 4.4(c) and 4.5(c)
OBNLM shows superior performance compared to SRAD in preserving undistorted
texture and contrast. However, it fails to remove the speckle noise completely from
the image. To visually compare the performance of our proposed framework to that of
the commercial US image acquiring machine used in this experiment, Figs. 4.4(d)-(e)
and 4.5(d)-(e) are portrayed. From these figures, observe that tissue texture is more
CHAPTER 4. RESULTS 61
(a) (b)
Figure 4.6: Estimated B-mode image with the deconvolution step as (a) bMCFLMS,
and (b) cepstrum.
prominent in the images provided by our proposed framework compared to those in
the machine B-mode images. To facilitate the observations, significant structures of
the images are marked with arrow and circle in Figs. 4.4(d)-(e) and 4.5(d)-(e) which
show the machine B-mode images have blurred and distorted the tissue structures.
Again, the cyst boundary in Fig. 4.5(d) is sharper and well-defined comapred to that
of Fig. 4.5(e). The estimated speckle noise patterns of the 5-th image frame as shown
in Figs. 4.4(f) and 4.5(f) contain tissue structures that justify the relevance of true
image dependent modeling [35] of speckle pattern as shown in (3.23).
In spite of offering an elegant solution to the speckle removal problem, the proposed
framework has a flaw in its complete pipeline as the deconvolution step is not realtime
implementable requiring 76 minutes in total for a single image of 128 A-lines with each
line having 1040 samples. However, as an alternate approach, the deconvolution step
can be replaced by a time-efficient cepstrum-based deconvolution as described in [54]
with a little cost paid in image quality as evident from higher NIQE score of 5.59 (see
Fig. 4.6 (b)) compared to that of 5.32 (see Fig. 4.6(a)) using the bMCFLMS algorithm,
and this brings down the total runtime to 6.3 seconds. A graphical processing unit
(GPU) based deconvolution technique to be investigatd in future may bring down the
overall B-mode image generation framework into real-time.
Chapter 5
Discussion and Conclusion
5.1 Discussion
In this study, we present a complete framework of signal processing approaches com-
prising of deconvolution, despeckling, gamma correction, and gray level transformation
to produce a high-resolution B-mode image with superior edge and texture from the
raw RF image. The parameters for SRAD and OBNLM algorithms were tuned for
the lowest NIQE score of the despeckled image. While deriving SRAD and OBNLM
algorithms in [38] and [35], respectively, deconvolution of raw RF image to enhance
resolution was not addressed. Hence, introducing deconvolution prior to SRAD and
OBNLM may have resulted in their poor performance. However, although their speckle
removal efficacy may be good without deconvolution, the image resolution will be poor.
In the proposed framework, the performance of the despeckling algorithm (MADS) is
dependent on the number of consecutive image frames to be considered in the MSNE
algorithm. As mentioned earlier, there is a tradeoff between the number of frames that
can be used without violating the quasi-stationarity assumption of the true US image
and the speckle noise estimation accuracy. In our experiment, we observed that five
consecutive image frames are good enough for a visually pleasant B-mode image gener-
ation. Again, the Lagrange multipliers– �1 in the constraint preventing misconvergence
of the MSNE algorithm (see (3.28)) and �2 in the energy constraint of the iterative
62
CHAPTER 5. DISCUSSION AND CONCLUSION 63
despeckling algorithm (see (3.41)) remain effective once set at an optimum level for
a particular US imaging set-up. To make the framework independent of the display
monitor, gamma correction as a post-processing step has been introduced. Again, to
offer the user a tunable contrast adjustment, two parameters Wlow and Whigh have been
used in the gray level transformation step.
In addition to providing a guideline for high-resolution B-mode image generation,
the thesis introduces a method to extract the speckle pattern inherent in a US image.
Despite of being a random process, speckle noise is not devoid of information. Since the
statistics of the speckle depends on the microstructure of the tissue parenchyma, it can
be useful for differentiating between different tissue compositions or types [58], [59].
5.2 Conclusion
This thesis has dealt with a complete framework for high-resolution ultrasound image
reconstruction from raw RF data. The proposed method relies on SIMO models for
both deconvolution and speckle noise estimation, and MISO model for despeckling. The
proposed framework completely follows the signal generation model and sequentially
addresses the issues with US imaging such as low resolution, speckle noise, and additive
noise. In the first step, to enhance the resolution, a 2-D deconvolution technique has
been introduced as an extension of our previously proposed 1-D bMCFLMS algorithm
which is necessary prior to despeckling according to the mathematical model of US
imaging. In the next step, a novel multiframe-based adaptive speckle noise estimation
(MSNE) algorithm estimates the speckle pattern without any a priori information on
the statistics of the image or the noise pattern. Using the estimated speckle pattern, an
energy constrained iterative algorithm estimates the true US image following a MISO
model. As the despeckling procedure is completely based on signal generation model
and does not involve any kind of ad-hoc filtering operation as reported in the literature,
it has resulted in a high quality tissue texture and edges in the image. Finally, gamma
correction and gray level transformation have been done as post-processing to produce
a complete B-mode image. The efficacy of the proposed algorithm has been tested
CHAPTER 5. DISCUSSION AND CONCLUSION 64
both quantitatively and qualitatively on simulation and in-vivo data. The results
have demonstrated the superiority of our proposed despeckling algorithm compared to
SRAD and OBNLM methods. Again, the proposed framework offers B-mode image
with superior texture and image details compared to those provided by a commercial
ultrasound scanner.
As our proposed framework preserves original image features such as texture, details
and edges, it may have a far reaching impact on medical imaging for diagnostic purpose.
At the same time, the proposed despeckling algorithm may be efficacious in dealing with
the speckle noise problem in other imaging such as synthetic aperture radar (SAR) [60]
and optical coherence tomography (OCT) [61].
List of Publications
Journal
1. Jayanta Dey, and M. K. Hasan, “Ultrasonic tissue reflectivity function esti-
mation using correlation constrained multichannel flms algorithm with missing
RF data”, Biomed. Phys. Eng. Express, vol. 4, no. 4, pp. 045024, 2018.
2. Jayanta Dey, Sharmin R. Ara, and M. K. Hasan, “Multiframe-based Adaptive
Despeckling Algorithm for Ultrasound B-mode Imaging with Superior Edge and
Texture,” submitted to IEEE Transactions on Ultrasonics, Ferroelectrics, and
Frequency Control, 2019.
65
Bibliography
[1] O. V. Michailovich and A. Tannenbaum, “Despeckling of medical ultrasound im-
ages,” IEEE Trans. Ultrason. Ferroelectr. Freq. Control, vol. 53, no. 1, pp. 64–78,
2006.
[2] H. Choi and J. Jeong, “Speckle noise reduction in ultrasound images using srad
and guided filter,” in Advanced Image Technology (IWAIT), 2018 International
Workshop on. IEEE, 2018, pp. 1–4.
[3] F. P. X. De Fontes, G. A. Barroso, P. Coupé, and P. Hellier, “Real time ultrasound
image denoising,” Journal of real-time image processing, vol. 6, no. 1, pp. 15–22,
2011.
[4] T. Joel and R. Sivakumar, “An extensive review on despeckling of medical ul-
trasound images using various transformation techniques,” Applied Acoustics, vol.
138, pp. 18–27, 2018.
[5] F. Xue, J. Liu, and X. Ai, “Parametric psf estimation based on predicted-SURE
with l1-penalized sparse deconvolution,” Signal, Image and Video Processing, pp.
1–8, 2018.
[6] Z. Al-Ameen, “Faster deblurring for digital images using an ameliorated
richardson-lucy algorithm,” IEIE Transactions on Smart Processing & Computing,
vol. 7, no. 4, pp. 289–295, 2018.
66
BIBLIOGRAPHY 67
[7] X. Zong, A. F. Laine, and E. A. Geiser, “Speckle reduction and contrast enhance-
ment of echocardiograms via multiscale nonlinear processing,” IEEE Trans. Med.
Imaging, vol. 17, no. 4, pp. 532–540, 1998.
[8] A. Achim, A. Bezerianos, and P. Tsakalides, “Novel bayesian multiscale method
for speckle removal in medical ultrasound images,” IEEE Trans. Med. Imaging,
vol. 20, no. 8, pp. 772–783, 2001.
[9] F. Beaufays, “Transform-domain adaptive filters: an analytical approach,” IEEE
Transactions on Signal processing, vol. 43, no. 2, pp. 422–431, 1995.
[10] R. Ahmad, A. W. Khong, M. K. Hasan, and P. A. Naylor, “An extended nor-
malized multichannel FLMS algorithm for blind channel identification,” in Signal
Processing Conference, 2006 14th European, 2006, pp. 1–5.
[11] M. Haque and M. Hasan, “Variable step-size multichannel frequency-domain LMS
algorithm for blind identification of finite impulse response systems,” IET Signal
Processing, vol. 1, no. 4, pp. 182–189, 2007.
[12] N. D. Gaubitch, M. K. Hasan, and P. A. Naylor, “Generalized optimal step-size for
blind multichannel lms system identification,” IEEE Signal Process Lett., vol. 13,
no. 10, pp. 624–627, 2006.
[13] M. K. Hasan, M. Shifat-E-Rabbi, and S. Y. Lee, “Blind deconvolution of ultra-
sound images using l1-norm-constrained block-based damped variable step-size
multichannel LMS algorithm,” IEEE Transactions on Ultrasonics, Ferroelectrics,
and Frequency Control, vol. 63, no. 8, pp. 1116–1130, 2016.
[14] M. A. Haque and M. K. Hasan, “Noise robust multichannel frequency-domain
LMS algorithms for blind channel identification,” IEEE Signal Processing Letters,
vol. 15, pp. 305–308, 2008.
[15] L. Liao, X.-L. Li, A. W. Khong, and X. Liu, “Analysis of the noise robustness prob-
lem and a new blind channel identification algorithm,” in Digital Signal Processing
(DSP), 2015 IEEE International Conference on. IEEE, 2015, pp. 838–842.
BIBLIOGRAPHY 68
[16] T. Taxt, “Restoration of medical ultrasound images using two-dimensional homo-
morphic deconvolution,” IEEE Transactions on Ultrasonics, Ferroelectrics, and
Frequency Control, vol. 42, no. 4, pp. 543–554, 1995.
[17] O. Michailovich and D. Adam, “Phase unwrapping for 2-d blind deconvolution of
ultrasound images,” IEEE Transactions on Medical Imaging, vol. 23, no. 1, pp.
7–25, 2004.
[18] O. Michailovich and A. Tannenbaum, “Blind deconvolution of medical ultrasound
images: A parametric inverse filtering approach,” IEEE Transactions on Image
Processing, vol. 16, no. 12, pp. 3005–3019, 2007.
[19] D. Adam and O. Michailovich, “Blind deconvolution of ultrasound sequences using
nonparametric local polynomial estimates of the pulse,” IEEE Transactions on
Biomedical Engineering, vol. 49, no. 2, pp. 118–131, 2002.
[20] O. Michailovich and D. Adam, “Shift-invariant, dwt-based" projection" method
for estimation of ultrasound pulse power spectrum,” IEEE Transactions on Ul-
trasonics, Ferroelectrics, and Frequency Control, vol. 49, no. 8, pp. 1060–1072,
2002.
[21] C. Yu, C. Zhang, and L. Xie, “A blind deconvolution approach to ultrasound
imaging,” IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Con-
trol, vol. 59, no. 2, 2012.
[22] G. Ayers and J. C. Dainty, “Iterative blind deconvolution method and its applica-
tions,” Optics letters, vol. 13, no. 7, pp. 547–549, 1988.
[23] R. Tekalp and J. Biemond, “Maximum likelihood image and blur identification: a
unifying approach,” Optical Engineering, vol. 29, no. 5, pp. 422–435, 1990.
[24] S. J. Reeves and R. M. Mersereau, “Blur identification by the method of generalized
cross-validation,” IEEE Transactions on Image Processing, vol. 1, no. 3, pp. 301–
311, 1992.
BIBLIOGRAPHY 69
[25] O. V. Michailovich and D. Adam, “A novel approach to the 2-d blind deconvolution
problem in medical ultrasound,” IEEE Trans. Med. Imaging, vol. 24, no. 1, pp.
86–104, 2005.
[26] M. K. Hasan, “Damped variable step size multichannel wiener LMS algorithm for
blind channel identification with noise,” Proc. Communication Systems, Networks
and Digital Signal Processing, pp. 374–377, 2006.
[27] M. A. Haque, M. S. A. Bashar, P. A. Naylor, K. Hirose, and M. K. Hasan, “En-
ergy constrained frequency-domain normalized LMS algorithm for blind channel
identification,” Signal, Image and Video Processing, vol. 1, no. 3, pp. 203–213,
2007.
[28] M. K. Hasan, J. Benesty, P. A. Naylor, and D. B. Ward, “Improving robustness of
blind adaptive multichannel identification algorithms using constraints,” in Signal
Processing Conference, 2005 13th European, 2005, pp. 1–4.
[29] M. K. Hasan and P. A. Naylor, “Analyzing effect of noise on LMS-type approaches
to blind estimation of SIMO channels: robustness issue,” in Signal Processing
Conference, 2006 14th European, 2006, pp. 1–4.
[30] G. Turin, “An introduction to matched filters,” IRE transactions on Information
theory, vol. 6, no. 3, pp. 311–329, 1960.
[31] U. R. Abeyratne, A. P. Petropulu, and J. M. Reid, “Higher order spectra based
deconvolution of ultrasound images,” IEEE Trans. Ultrason. Ferroelectr. Freq.
Control, vol. 42, no. 6, pp. 1064–1075, 1995.
[32] J. A. Jensen, “Simulation of advanced ultrasound systems using field ii,” in Biomed-
ical Imaging: Nano to Macro, 2004. IEEE International Symposium on. IEEE,
2004, pp. 636–639.
[33] V. S. Frost, J. A. Stiles, K. S. Shanmugan, and J. C. Holtzman, “A model for radar
images and its application to adaptive digital filtering of multiplicative noise,”
BIBLIOGRAPHY 70
IEEE Transactions on Pattern Analysis & Machine Intelligence, no. 2, pp. 157–
166, 1982.
[34] D. T. Kuan, A. A. Sawchuk, T. C. Strand, and P. Chavel, “Adaptive noise smooth-
ing filter for images with signal-dependent noise,” IEEE Transactions on Pattern
Analysis & Machine Intelligence, no. 2, pp. 165–177, 1985.
[35] P. Coupé, P. Hellier, C. Kervrann, and C. Barillot, “Nonlocal means-based speckle
filtering for ultrasound images,” IEEE Trans. Image Process., vol. 18, no. 10, pp.
2221–2229, 2009.
[36] A. Buades, B. Coll, and J.-M. Morel, “A review of image denoising algorithms,
with a new one,” Multiscale Modeling & Simulation, vol. 4, no. 2, pp. 490–530,
2005.
[37] H. R. Shahdoosti and Z. Rahemi, “A maximum likelihood filter using non-local
information for despeckling of ultrasound images,” Machine Vision and Applica-
tions, vol. 29, no. 4, pp. 689–702, 2018.
[38] Y. Yu and S. T. Acton, “Speckle reducing anisotropic diffusion,” IEEE Trans.
Image Process., vol. 11, no. 11, pp. 1260–1270, 2002.
[39] S. K. Jain and R. K. Ray, “Non-linear diffusion models for despeckling of images:
achievements and future challenges,” IETE Technical Review, pp. 1–17, 2019.
[40] S. Gupta, R. Chauhan, and S. Saxena, “Locally adaptive wavelet domain bayesian
processor for denoising medical ultrasound images using speckle modelling based
on rayleigh distribution,” IEE Proceedings-Vision, Image and Signal Processing,
vol. 152, no. 1, pp. 129–135, 2005.
[41] N. Gupta, M. Swamy, and E. Plotkin, “Despeckling of medical ultrasound images
using data and rate adaptive lossy compression,” IEEE Trans. Med. Imaging,
vol. 24, no. 6, pp. 743–754, 2005.
BIBLIOGRAPHY 71
[42] Y. Yue, M. M. Croitoru, A. Bidani, J. B. Zwischenberger, and J. W. Clark, “Non-
linear multiscale wavelet diffusion for speckle suppression and edge enhancement
in ultrasound images,” IEEE Trans. Med. Imaging, vol. 25, no. 3, pp. 297–311,
2006.
[43] D. Koundal, S. Gupta, and S. Singh, “Speckle reduction method for thyroid ultra-
sound images in neutrosophic domain,” IET Image Processing, vol. 10, no. 2, pp.
167–175, 2016.
[44] F. Luisier, T. Blu, and M. Unser, “A new sure approach to image denoising:
Interscale orthonormal wavelet thresholding,” IEEE Trans. Image Process., vol. 16,
no. 3, pp. 593–606, 2007.
[45] M. I. H. Bhuiyan, M. O. Ahmad, and M. Swamy, “Spatially adaptive wavelet-
based method using the cauchy prior for denoising the sar images,” IEEE Trans.
Circuits Syst. Video Technol., vol. 17, no. 4, pp. 500–507, 2007.
[46] J. Dey and M. K. Hasan, “Ultrasonic tissue reflectivity function estimation using
correlation constrained multichannel flms algorithm with missing rf data.” Biomed.
Phys. Eng. Express, 2018.
[47] G. Xu, H. Liu, L. Tong, and T. Kailath, “A least-squares approach to blind channel
identification,” IEEE Trans. Signal Process., vol. 43, no. 12, pp. 2982–2993, 1995.
[48] H. R. Shahdoosti and O. Khayat, “Image denoising using sparse representation
classification and non-subsampled shearlet transform,” Signal, Image and Video
Processing, vol. 10, no. 6, pp. 1081–1087, 2016.
[49] D. Gupta, R. S. Anand, and B. Tyagi, “Speckle filtering of ultrasound images
using a modified non-linear diffusion model in non-subsampled shearlet domain,”
IET Image Proc., vol. 9, no. 2, pp. 107–117, 2014.
[50] A. Garg and V. Khandelwal, “Despeckling of medical ultrasound images using fast
bilateral filter and neighshrinksure filter in wavelet,” Advances in Signal Processing
and Communication: Select Proceedings of ICSC 2018, vol. 526, p. 271, 2018.
BIBLIOGRAPHY 72
[51] P.-M. Lee and H.-Y. Chen, “Adjustable gamma correction circuit for tft lcd,”
in Circuits and Systems, 2005. ISCAS 2005. IEEE International Symposium on.
IEEE, 2005, pp. 780–783.
[52] X. Huang, Z. Jia, J. Zhou, J. Yang, and N. Kasabov, “Speckle reduction of re-
constructions of digital holograms using gamma-correction and filtering,” IEEE
Access, vol. 6, pp. 5227–5235, 2018.
[53] P. Suetens, Fundamentals of medical imaging. Cambridge university press, 2002.
[54] T. Taxt and J. Strand, “Two-dimensional noise-robust blind deconvolution of ul-
trasound images,” IEEE Trans. Ultrason. Ferroelectr. Freq. Control, vol. 48, no. 4,
pp. 861–866, 2001.
[55] H. R. Shahdoosti, “Two-stage image denoising considering interscale and intrascale
dependencies,” J. Electron. Imaging, vol. 26, no. 6, p. 063029, 2017.
[56] A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a" completely blind"
image quality analyzer.” IEEE Signal Process. Lett., vol. 20, no. 3, pp. 209–212,
2013.
[57] T. Loupas, W. McDicken, and P. L. Allan, “An adaptive weighted median filter
for speckle suppression in medical ultrasonic images,” IEEE Trans. Circuits Syst.,
vol. 36, no. 1, pp. 129–135, 1989.
[58] R. F. Wagner, “Statistics of speckle in ultrasound b-scans,” IEEE Trans. Sonics
& Ultrason., vol. 30, no. 3, pp. 156–163, 1983.
[59] C. Sehgal, “Quantitative relationship between tissue composition and scattering
of ultrasound,” The Journal of the Acoustical Society of America, vol. 94, no. 4,
pp. 1944–1952, 1993.
[60] C. Wang, L. Xu, D. A. Clausi, and A. Wong, “A bayesian joint decorrelation and
despeckling of sar imagery,” IEEE Geoscience and Remote Sensing Letters, 2019.
BIBLIOGRAPHY 73
[61] A. Paul, D. P. Mukherjee, and S. T. Acton, “Speckle removal using diffusion
potential for optical coherence tomography images,” IEEE journal of biomedical
and health informatics, vol. 23, no. 1, pp. 264–272, 2019.