Sudipan Saha Mtech Ultrasound CAD

Ultrasound Image Processing for CAD
and Other Applications
M.Tech. Dissertation
submitted in partial fulfillment of the requirements

for the degree of
Master of Technology
by
Sudipan Saha
under the guidance of
Prof. S. N. Merchant
Department of Electrical Engineering

Indian Institute of Technology, Bombay
Mumbai - 400076.
June 2014
Abstract
Ultrasound imaging is a popular medical imaging modality which is routinely used in
many clinical applications. Computer assistance can make the diagnosis comparatively
easier and efficient. But devising computer aided diagnosis tools for ultrasound images are
comparatively complicated than other medical imaging modalities because of less visual
cues and presence of speckle noise. Many algorithms have been proposed in literature
regarding removal of speckle noise. In this work we have studied some of the existing
methods of speckle noise removal, which can be termed as preprocessing. More emphasis
has been laid on the post-processing tasks i.e. on studying and devising algorithms related
to the computer aided diagnosis. We have grouped computer aided diagnosis algorithms
into two groups. The first group consists of those algorithms which are based on the
explicit methods and do not depend on machine learning. In this work, explicit algorithms
have been proposed for artery detection, cyst localization and gallstone detection. The
second group consists of those algorithms which rely on machine learning. Some of the
existing popular machine learning related methods has been thoroughly studied. New
algorithms have also been proposed. The algorithms have been tested on a prostate
cancer database. Ultrasound imaging devices are often misused for illegal abortion. In
this work, the scope of using image processing techniques to stop the misuse of ultrasound
imaging devices have been explored.
i
Contents
Abstract i
List of Figures v
List of Tables vi
Abbreviations vii
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Organization of the report . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Fundamentals of Ultrasound Imaging 3

2.1 Frequency range of ultrasound wave . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Converse piezoelectric effect . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Different ultrasound imaging modes . . . . . . . . . . . . . . . . . . . . . . 3
2.4 Ultrasound image generation . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.5 Advantages of ultrasound imaging . . . . . . . . . . . . . . . . . . . . . . . 6
2.6 Drawbacks of ultrasound imaging . . . . . . . . . . . . . . . . . . . . . . . 6
2.7 Future of ultrasound imaging . . . . . . . . . . . . . . . . . . . . . . . . . 7
3 Ultrasound Image Preprocessing 8

3.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Various speckle removal algorithms . . . . . . . . . . . . . . . . . . . . . . 9
3.3 Estimation of statistical parameters . . . . . . . . . . . . . . . . . . . . . . 10
3.3.1 SNR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3.2 MSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3.3 RMSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
ii
3.3.4 PSNR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3.5 Correlation coefficient . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3.6 Speckle index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4 Explicitly Programmed CAD 14

4.1 Carotid artery localization . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.1.1 Roundness metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.1.2 Proposed algorithm for artery localization . . . . . . . . . . . . . . 16
4.1.3 Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 Cyst localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.3 Gallbladder stone detection . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5 Machine Learning Based CAD 22

5.1 Features for classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.1.1 First order intensity histogram features . . . . . . . . . . . . . . . . 23
5.1.2 Gray-level co-occurrence matrix . . . . . . . . . . . . . . . . . . . . 24
5.1.3 Fractal features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.1.4 Other features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.2 Dimensionality reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.3 Classification methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.3.1 Using kNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.3.2 Using SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.3.3 Classification using k-Means clustering method . . . . . . . . . . . 29
5.3.4 Combining SVM and k-Means . . . . . . . . . . . . . . . . . . . . . 30
5.3.5 One class SVM and SVDD . . . . . . . . . . . . . . . . . . . . . . . 31
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.5 Adapting to domain change . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6 Fetal Ultrasound Image Processing 37

6.1 Analyzing reasons for preference of male child . . . . . . . . . . . . . . . . 37
6.2 Gender imbalance in India and other countries . . . . . . . . . . . . . . . . 39
6.3 Consequences of gender imbalance . . . . . . . . . . . . . . . . . . . . . . . 39
6.4 Misuse of ultrasound technology . . . . . . . . . . . . . . . . . . . . . . . . 39
6.5 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.5.1 Amniotic fluid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.5.2 Curvy contour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
iii
6.5.3 Shape based features . . . . . . . . . . . . . . . . . . . . . . . . . . 41
7 Conclusion and Future Work 43

7.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
A GUI to help radiogist to provide manually segmented images 46
iv
List of Figures
2.1 Block diagram of ultrasound imaging device . . . . . . . . . . . . . . . . . 6
3.1 Ultrasound phantom image and results after applying various speckle re-
ducing techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.1 Two different instances of artery detection. Detected artery shown by red
circle. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Detected cyst shown in red . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 Illustration of gallstone detection . . . . . . . . . . . . . . . . . . . . . . . 21
6.1 Male genitalia composed of two elliptical regions . . . . . . . . . . . . . . . 42
A.1 Screen-shot of the GUI developed to collect ground-truth data . . . . . . . 46
v
List of Tables
3.1 Estimated statistical parameters for various speckle reduction techniques

for ”Shepp-Logan” image . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
5.1 Stotal variation with k when using kNN algorithm . . . . . . . . . . . . . . 34

5.2 Stotal variation with γ when C is fixed at 0.1 . . . . . . . . . . . . . . . . . 35
5.3 Stotal variation with C when γ is fixed at 0.5 . . . . . . . . . . . . . . . . . 35
6.1 Gender ratio (0-6 yrs) change over time . . . . . . . . . . . . . . . . . . . . 38

6.2 Gender ratio of few states of India . . . . . . . . . . . . . . . . . . . . . . . 38
vi
Abbreviations
ASM Angular Second Moment.
CAD Computer Aided Diagnosis.

CSVM Clustered Support Vector Machine.
GLCM Grey-Level Co-occurrence Matrix.
HMAF Homogeneous Mask Area Filtering.
kNN K Nearest Neighbor.
MSE Mean Square Error.
PCA Principal Component Analysis.

PSNR Peak Signal to Noise Ratio.
RBF Radial Basis Function.

RMSE Root Mean Square Error.
ROI Region of Interest.
SFTA Segmentation-based Fractal Texture Analysis.

SNR Signal to Noise Ratio.
SVDD Support Vector Data Description.
SVM Support Vector Machine.
vii
Chapter 1
Introduction
1.1 Motivation
Ultrasound imaging is one of the most popular medical imaging techniques. The main
reasons behind its popularity are its non-invasive nature, suitability for the real time
imaging. It is also comparatively less expensive than other medical imaging modalities.
It is routinely used in many clinical applications and different areas of biomedical research.
Though it is already a mature technology, it is not yet stagnant, as ultrasound technology
is evolving every day. It is striving towards more automation and decrease in size of the
ultrasound device, thus increasing portability. Ultrasound imaging can be more effec-
tively utilized if coupled with suitable image processing modules to perform automated
diagnosis. But image processing is comparatively more intricate for ultrasound images
than other medical images, due to presence of multiplicative noise and absence of visual
cues. So algorithms specialized for the ultrasound domain needs to be devised. Currently
lot of attention has been paid by the research community across the globe to develop ul-
trasound image processing based computer aided diagnosis (CAD) techniques. Also, with
the advent of ultrasound imaging, it has been misused for the gender detection of the fetus
and illegal abortion. Besides automating the diagnosis techniques, it is also essential to
develop techniques to stop misuse of ultrasound imaging for such illegal purposes. Com-
paratively less attention has been paid on this aspect by the research community. In this
thesis, we have dealt with the B-mode (basic two dimensional intensity mode) ultrasound
images. Basic preprocessing steps required for ultrasound image processing like speckle
noise removal has been studied. Major aspects and algorithms related to CAD has been
studied and also new algorithms has been proposed. We have also studied and proposed
algorithms to stop of ultrasound imaging for the illegal abortion.
1
1.2 Contribution
1. Various existing speckle noise reduction techniques have been studied and few of
those have been implemented.
2. The computer aided diagnosis techniques has been exhaustively studied from both
explicit (non-machine learning) perspective and machine learning based perspec-
tive. Novel rule based algorithms has been devised and implemented for the carotid
artery detection. Various machine learning techniques have been studied for our
application. Some of those have been implemented. Also new algorithms has been
proposed.
3. Various possible techniques regarding stopping misuse of ultrasound devises for the
fetal abortion have been studied.
4. Other perspectives has also been looked at, e.g. social impact analysis of the misuse
of ultrasound device for illegal abortion, feasibility analysis of the ultrasound image
processing algorithms on the android platform.
1.3 Organization of the report

The thesis outlay evolves its structure from initiating an overview of the ultrasound imag-
ing in Chapter 2. Chapter 3 discusses about the presence of multiplicative noise (speckle
noise) in the ultrasound image and some of the existing speckle removal techniques. Few
techniques has been implemented and compared in terms of the various statistical param-
eters. Chapter 4 discusses how explicitly programmed algorithms can be applied for the
CAD applications. Few explicit algorithms has also been designed e.g. automated detec-
tion of cyst, automated detection of the carotid artery. The objective of this chapter is to
demonstrate how effective the explicit techniques can be in devising CAD applications,
while demonstrating few illustrations of those. The algorithm proposed for detection of
the artery is novel. Chapter 5 discusses how machine learning algorithms can be applied
for the CAD applications. Few algorithms like k Nearest Neighbor (kNN), Support Vec-
tor Machine (SVM) has been implemented and tested on a prostate dataset. The social
impact of the misuse of ultrasound imaging for the illegal abortion, and various possible
ways to stop it has been explored in Chapter 6. Though very few literatures are available
in this context, those available have been explored. Conclusions and future directional
advancements in the field have been outlined in the Chapter 7.
2
Chapter 2
Fundamentals of Ultrasound Imaging
Ultrasound imaging is a popular medical imaging modality which is routinely used in

many clinical applications and different areas of biomedical research. Though it is already
a mature technology, it is not stagnant, as ultrasound technology is evolving everyday,
aiming towards more automation and decrease in size.
2.1 Frequency range of ultrasound wave

Ultrasound waves are characterized by the frequency range above 20 KHz, which is also
the upper limit of human hearing. In medical ultrasound applications, frequencies ranging
from 500 KHz to 30 MHz is normally used [1].
2.2 Converse piezoelectric effect

Ultrasound waves are generated by the piezoelectric crystals. When mechanical stress is
applied to piezoelectric crystals, electric current is generated. The reverse is also true,
when an electric current is applied to the piezoelectric crystal, mechanical stress and
deformation is induced in it [2]. The ultrasound waves are generated by this converse
piezoelectric effect.
2.3 Different ultrasound imaging modes

Different imaging modes in current ultrasound are as follows [3, 4]:
1. A (Amplitude) mode imaging which displays amplitude of sampled signal for a single
sound wave as function of time.
3
2. B (Brightness) mode imaging which is performed by sweeping the transmitted ul-
trasound wave over plane to generate a 2D image.
3. M (Motion) mode in which pulses are emitted in quick succession and each time
either an A-mode or a B-mode image is captured. It is used to analyze moving
organs.
4. CW (Continuous Wave) Doppler which can be used for estimating blood flow in
veins.
5. PW (Pulse Wave) Doppler which can be used to estimate both blood flow and
velocity location.
6. Color Doppler, where the Pulse Wave Doppler is used to create color image which
is super-imposed on B-mode image. A color code is used to indicate the direction
of the flow. Doppler can differentiate between the RBCs moving away and moving
towards the probe [2]. RBCs moving away from the probe return ultrasound wave
at lower frequency and displayed as blue. but color can vary depending on the angle
between probe and skin. So color cannot be used as a very reliable indicator.
2.4 Ultrasound image generation

Electric current is applied to the piezoelectric crystals in the ultrasound probe, which leads
to generation of ultrasound wave. The ultrasound probe plays a dual role of transmitter
and receiver. It generates ultrasound wave for certain fraction of the time and listen
for return of ultrasound waves remaining of the time [2]. The energy of the returning
wave depends on the reflection occurring at the interface of the different types of tissue.
The greatest reflection occurs from tissues with great difference in acoustic impedance
[2]. Using the piezoelectric effect, the crystals in probe converts mechanical energy of
the returning ultrasound wave into electric current, which is further processed by the
machine to produce a grayscale image. The grayscale image ranges from complete black
to complete white and is more whitish dependent on the energy of the returning echo.
The image regions can be divided into these three broad classes [2]:
1. Hyper-echoic areas which are whitish, and have great amount of energy in returning
echoes. As there is a large impedance difference between the acoustic impedance of
the bone and soft tissue, so the bone and soft tissue interface produces hyper-echoic
image.
4
2. Hypo-echoic areas which are between white and black. Amount of energy in return-
ing echo is medium. As difference in acoustic impedance between various types of
soft tissues like blood, muscle, fat etc. are very less, so those results in hypo-echoic
image.
3. Anechoic areas which are seen as black. As blood is relatively homogeneous in its
acoustic impedance, so ultrasound wave can flow through it with ease. So large
blood vessels are anechoic in appearance.
Ultrasound probes or transducers can be classified based on their frequency range [2]. High
frequency probes have better axial and lateral resolution, but greater tissue attenuation.
Attenuation is basically caused by conversion to the heat. Due to attenuation, imaging
of the deeper structure is not possible with the high frequency probes. So high frequency
probes are better suited for capturing small structures. On the other hand, low frequency
probes have greater tissue penetration, but poor resolution. So low frequency probes are
suitable for capturing larger size object.
An ultrasound system overview is shown in Fig. 2.1. The first stage is beam forming.
Transducers are aligned to focus the ultrasound waves at different focal lengths along a
given scan-line. The transducers which are repeatedly switched on and then off, listens to
the reflected wave during the off time. When passing through the human body, as portion
of the ultrasound waves gets absorbed by the body, waves traveling deeper become weaker.
To compensate this distance varying signal attenuation, a distance (time) varying gain is
applied to the reflected signal, which is called called time gain compensation [5]. After
beam-forming the signal is then passed through the signal processing blocks. These include
filtering, envelope extraction, log compression. Band pass filtering is done to reduce noise
outside region of interest. Then envelope is extracted, which is further log compressed to
achieve the desired dynamic range required for display [3]. The processed signal is then
fed to the scan conversion module which generates the image [5]
5
Figure 2.1: Block diagram of ultrasound imaging device
2.5 Advantages of ultrasound imaging

As given in [6, 7, 4] and [7], ultrasound has following advantages compared to other
medical imaging techniques:
1. Ultrasound is considered safe than other medical imaging techniques.
2. It is relatively inexpensive in comparison to other medical imaging techniques.
3. It is painless and easily tolerated by the patients.
4. It produces image in real time.
5. Ultrasound scanning gives clear image of soft tissues which do not show up well in
X-ray images.
6. It is portable.
2.6 Drawbacks of ultrasound imaging

The major disadvantages of ultrasound imaging as pointed out by Shung [6], Kaur[4] are:
1. Gasy and bony structures are not easy to image, needs use of specialized procedures.
6
2. Interpretation of the image is operator skill-dependent.
3. Good image cannot be obtained from obese patients.
4. More presence of noise components.
5. Images are of low resolution.
2.7 Future of ultrasound imaging

There are many scopes of improvement yet in the ultrasound imaging field. Development
of portable scanners, miniature pocket size scanners and high frequency scanners are yet
in progress [6]. Use of ASICs in design has contributed to drastic reduction in physical
size [6]. General Electronics has released Ipod size scanner which weighs only 390 gm
[6]. More compact size scanner is expected to come in near future. But major improve-
ment is expected in the field of automated CAD with development of more sophisticated
ultrasound image processing algorithms.
7
Chapter 3
Ultrasound Image Preprocessing
3.1 Background
Noise is a common problem in any imaging system. Noise is generally unwanted data
which decreases contrast in image or causes blurring of edges. While most of the times
noise is due to the environment, many times it can be due to the physical nature of the
system. Noise are broadly of two types: additive and multiplicative. Whereas additive
noise mainly depends on the system and can be removed by modeling the system; multi-
plicative noise is dependent on image and thus not easy to model.
In the ultrasound image, the most dominant noise is speckle noise which is multiplicative
in nature. It appears as a granular pattern because of the image formation under coherent
waves. Speckle occurs due to diffusion scattering. It happens when sound wave interferes
with small objects of size comparable to sound wavelength.
Speckle noise removal is mainly important for the following reasons:
1. For enhancing visualization, i.e. to improve the human interpretation of the images.
2. Speckle removed image is easy to segment.
But following factors needs to be taken into account when designing despeckling algo-
rithms:
1. Speckle noise, though regarded as noise, may present useful diagnostic information
for radiologist.
2. Speckle removal algorithm should not blur the edges.
8
3.2 Various speckle removal algorithms
The most basic filter used for denoising purposes is mean filter. It has effect of blurring
the edges [8]. Mean filter is generally used for additive Gaussian noise and not suitable
for speckle noise which obeys multiplicative model with non-Gaussian nature. So mean
filter is not suitable for speckle noise. Some of the modifications of the mean filtering has
been proposed in literature [9]. Lee filter takes a convex combination of the pixel intensity
and the average intensity within a given window. Here given an input image I(x, y) the
output image Y (x, y) is computed as:
Y (x, y) = I(x, y)W (x, y) + I 0 (x, y)(1 − W (x, y)) (3.1)
where I 0 (x, y) is the mean intensity within a certain window region and W (x, y) is an
adaptive coefficient.
Other simple filtering methods has been proposed [10, 11]. In [10], 2-D FFT of the
ultrasound image is taken and then local peaks are detected in the frequency domain.
After removing the local peaks, inverse 2-D FFT is taken. But this method don’t deal
with multiplicative nature of speckle noise. In [11] multiplicative model of speckle noise is
converted into additive model by taking logarithm of the original speckled image. Then
DWT of the image is taken, image is thresholded and inverse DWT is performed on the
image. Finally exponential is computed to obtain the denoised image.
The most basic non-linear filter is median filtering. It produces less blurred image in
comparison to the mean filtering. It is very simple to implement, but still quite robust
in performance [8]. Various variants of the median filtering has also been proposed in
literature [12, 13, 14]. In [12] an adaptive weighted median filter has been proposed which
uses the ratio of mean to variance as a index of how uniform in value the pixels in an
area are. If the mean to variance ratio is small, it is assumed that the central pixel is
in vicinity of any image discontinuity. So additional weight is given to the central pixel.
If the ratio is found to be large, then it is assumed that the region consists of nearly
uniform pixel values and distant pixels are given greater weight for estimating the median
value. In [13] at every point of the original ultrasound image, a set of median values are
computed, by taking the median value of the pixel intensities that lie along the stick.
Then the final speckle removed image is formed by selecting the largest median value at
each point. In [14] rough set theory has been applied along with median filtering. In this
method, a 3 × 3 window is considered taking each pixel as central pixel. If the central
pixel is maximum or minimum value in the given window, then this pixel is considered as
noise pixel. Median filtering is applied only for those pixels. The major drawback of using
9
any median filtering based method is the high computational cost. Sorting of N pixels
require time complexity of O(N logN ) [15]. Homogeneous mask area filtering (HMAF)
can also be used for speckle reduction. In homogeneous mask area filtering, the most
homogeneous neighborhood around a pixel is found at location (i, j) using 3 × 3 subset
windows within a neighborhood of 5 × 5 window [16]. Speckle index for each of the 3×3
window is computed by taking ratio of the variance and the mean. 9 such speckle index
are obtained. Minimum of those values are obtained and mean value corresponding to
that 3 × 3 subset window is assigned to the pixel under consideration.
Farzana et al. has proposed an adaptive bilateral filtering based method [17]. The bilateral
filter consists of a domain filter and a range filter. Bilateral filter is non-linear, noise-
reducing filter with edge-preserving capability.
Squeeze box filter has also been used for despeckling [18, 19]. Tay et al. proposed the
method which considers local extrema of the B mode image f (i, j) as outliers. But
the method is iterative and hence not suitable for real time implementation / should be
restricted to few iterations, which will affect the performance. Yu and Acton [20] proposed
an anisotropic diffusion based technique which is also iterative in nature. Results of various
speckle reduction techniques on a noisy ultrasound phantom has been shown in Fig. 3.1.
3.3 Estimation of statistical parameters

Various statistical parameters like Signal to Noise Ratio (SNR)[5], Peak Signal to Noise
Ratio (PSNR)[10], Mean Square Error (MSE) [5], Root Mean Square Error (RMSE)[5],
Correlation coefficient, Speckle index [15] are used to evaluate the performance of the
speckle noise reduction techniques.
3.3.1 SNR
SNR directly computes ratio of the variance of the signal and the noise. Higher SNR
indicates the despeckling algorithm is better.
σf2
SNR = 10 log10 (3.2)
σe2
where σf2 and σe2 is the variance of the original noise free image and the error image
(difference between original noise free image and the despeckled image) respectively.
10
3.3.2 MSE
MSE is used as an index to estimate difference between the original image and the denoised
image. When f and g indicates the original noiseless image and the denoised image
respectively, then,
M N
1 XX
MSE = (f (i, j) − g(i, j))2 (3.3)
M N i=1 j=1
where M ,N represents nos. of rows and nos. of columns of the image, i.e. M N is the size
of the image.
3.3.3 RMSE
RMSE is calculated by taking square root of the MSE.
√
RMSE = MSE (3.4)
Lower RMSE indicates despeckling algorithm is better.
3.3.4 PSNR
PSNR can be regarded as the ratio between power of a signal and the power of corrupting
noise. PSNR in decibels is expressed as:
PSNR = 20 log10 (255/RMSE) (3.5)
Higher PSNR indicates despeckling algorithm is better.
3.3.5 Correlation coefficient

When f and g indicates the original noiseless image and the denoised image respectively,
then correlation coefficient (COC) is defined as:
(f − f¯)(g − ḡ)
P
COC = pP (3.6)
(f − f¯)2 (g − ḡ)2
P
3.3.6 Speckle index

Lower value of speckle index means the despeckling algorithm is better. It is defined as
follows:
M N
1 X X σ(i, j)
SI = (3.7)
M N i=1 j=1 µ(i, j)
11
where σ(i, j) and µ(i.j) represents the standard deviation and mean corresponding to a
certain window domain. M ,N represents nos. of rows and nos. of columns of the image,
i.e. M N is the size of the image.
(a) Input noisy ultra- (b) Mean filtered out- (c) Median filtered out-
sound phantom put put
(d) Modified median fil- (e) HMAF [16] output (f) Squeeze box [18] fil-
tered [14] output tered output
(g) Adaptive bilateral

[17] filtered output
Figure 3.1: Ultrasound phantom image and results after applying various speckle reducing
techniques
The MSE, RMSE, SNR, PSNR values for different speckle reduction techniques tested
on the ”Shepp-Logan” phantom image has been listed in Table 3.1.
12
Method MSE RMSE SNR PSNR
Mean filtering 0.011 0.11 5.50 67.49
HMAF [16] 0.009 0.09 6.62 68.56
Median filtering 0.006 0.08 8.36 70.25
Modified median filtering [14] 0.004 0.06 10.12 72.00
Squeeze box filtering[18] 0.002 0.04 13.28 75.26
Adaptive bilateral[17] filtering 0.002 0.04 13.28 75.26
Table 3.1: Estimated statistical parameters for various speckle reduction techniques for
”Shepp-Logan” image
13
Chapter 4
Explicitly Programmed CAD
Computer aided diagnosis (CAD) are procedures which uses computing devices to assist
in the analysis / interpretation of the medical images. Ultrasound image processing based
computer diagnosis can be approached in two ways:
1. Explicitly programmed CAD (no use of machine learning)
2. Machine learning based CAD
In the explicit approach, some particular rules are known to the algorithm developer or
some rules are derived by experimenting with the available images / data. The algorithm
developer uses those rule set to formulate an explicit algorithm. In this method the
algorithm developer depends on his/her knowledge about the domain to formulate the
rules. Then those rules are used for designing the required CAD application. Some of
the examples where such approach has been partially / fully used are [21, 22, 23, 24, 25].
In [21], a computerized breast lesion segmentation technique has been proposed. In [22],
an algorithm for tumor detection in breast ultrasound images has been proposed. The
breast images are normalized followed by fuzzification. Then the image is enhanced and an
optimum threshold is estimated which is used to binarize the image. Watershed transform
is used on the binarized image to obtain the final tumor boundary. In [23], Hafizah et al.
have experimented with four different groups of the kidney ultrasound images, normal,
bacterial infection, cystic and the kidney stone. They have estimated different intensity
histogram features and different Gray level co-occurrence matrix (GLCM) features for the
different classes. Based on those features authors have tried to conclude which features are
the best suited to differentiate between different categories. While their work is mainly
to derive a set of rules, their work can be used further for machine learning / ANN
based systems also. In [24], Raja et al. has considered three different classes of kidney;
14
normal, cystic and renal disease. Authors have obtained range of values for five different
parameters for three different cases of kidney. In [25], Usha and Sandya have proposed
an algorithm for measurement of ovarian shape and size.
We have implemented algorithms for following tasks based on explicit approach:
• Carotid artery localization
• Cyst localization
• Gallbladder stone detection
4.1 Carotid artery localization

The artery perpendicular cut captured by sonograph appears to be nearly circular [26].
Automatic localization of the artery is important as it can be used for further processing
to predict cardio-vascular events. An algorithm has been proposed which would work on
the single frame to locate the artery. The algorithm has been tested on the database
which has been used in [26, 27].
Every explicit algorithm uses some well defined properties to formulate the algorithm.
Following properties has been used to devise algorithm for carotid artery localization:
• The artery perpendicular cut is nearly circular.
• The artery generally appears darker in ultrasound images.
4.1.1 Roundness metric

Roundness of any object has been evaluated by a roundness metric. Given an object, the
roundness metric is evaluated in the following way:
i. The target object’s filled area is calculated. Filled area does not include any other
object which is completely enclosed by our target object, but not a part of the target
object. Let filled area be a.
ii. The target object’s perimeter is calculated. Let it be p.
iii. We define a metric m1 as:

4∗π∗a
m1 = (4.1)
p2
Above metric for roundness evaluation was proposed by Cox [28].
15
iv. Major axis length of the ellipse which has the same normalized second central mo-
ments as the given object is calculated. Let major axis length be l1 . Similarly the
minor axis length is evaluated. Let the minor axis length be l2 .
v. We define a metric m2 as:

l1
m2 = (4.2)
l2
vi. The metric m1 will attain value equal to one for a circle and lesser value for non-
circular objects. The metric m2 will attain value equal to one for a non-circular
object and greater value for non-circular objects. So we would like to minimize m1
and maximize m2 . We formulated the metric mround as:
m1
mround = (4.3)
m2
The proposed roundness metric can be minimized to assess the roundness of any given
object.
4.1.2 Proposed algorithm for artery localization

Using the proposed roundness metric and property that artery looks darker in the ultra-
sound images, following algorithm has been devised for carotid artery localization:
i. Let the input image Iartery1 . Histogram equalization operation is applied on Iartery1
to obtain Iartery2 .
ii. Then the image is binarized and then inverted to obtain Iartery3 . So, darker regions in
the original image correspond to the white pixels in the binarized image and whiter
regions in the original image correspond to the black pixels in the binarized image.
iii. Morphological closing operation is applied on the image Iartery3 to obtain Iartery4 . It
is performed successively with two different structuring elements. It is first performed
with a circular disk as structural element and then with a line as structuring element
with varying angle (0◦ to 90◦ ).
iv. All the connected white components are detected in Iartery4 . A connected component
is a group of spatially connected adjacent pixels which has same intensity value (either
all white or all black) in the binarized image. Let the connected components are Ci
where i varies from 1 to Ncount .
16
v. A value tmaxlength is calculated by calculating the 0.3 times the value of maximum of
the width and height of the image. This value is used as a threshold. The connected
components having major axis value greater than this threshold value are removed
from further consideration. This concept has grown from the fact that region of
interest (here artery) cannot occupy a very large area of the ultrasound image. If
the target object was indeed occupying such a major chunk of the ultrasound image,
then we would have not been interested in localizing it. Let the remaining connected
components are Ci where i varies from 1 to Mcount and Mcount ≤ Ncount .
vi. Corresponding to all the connected components Ci a counter gi is initiated to zero.

A threshold area tarea is defined. Following steps are repeated for tarea =2000, 1600,
1200, 800.
(a) Only those connected components which has pixel value greater than tarea are
kept into consideration.
(b) Roundness metric mround is evaluated for all connected components in consid-
eration. Let index of the connected component having best roundness metric
value (i.e. minimum) is ibest .
(c) The counter corresponding to the ibest is is increased by one, i.e.
gibest = gibest + 1 (4.4)
vii. The connected component with higher value of g is declared as the region correspond-
ing to the detected carotid artery. In case of tie between two regions the region with
more numbers of pixel is declared as the desired region.
viii. Centroid of the identified connected component is calculated to estimate the center
of the artery.
ix. The average value of major axis length and minor axis length of the ellipse having
same normalized second central moments as the identified connected component is
calculated to estimate the radius of the artery area.
4.1.3 Result
The proposed algorithm has been tested on images from the database which has been used
in [26, 27]. The algorithm has been tested total amount of 283 images from that database
17
(a)
(b)
Figure 4.1: Two different instances of artery detection. Detected artery shown by red
circle.
18
both with and without applying any noise removal step. While no noise removal step
was applied, the artery was correctly identified in 188 images, thus yielding an accuracy
of 66.43%. The algorithm has been tested with two types of noise removal step, median
filtering and adaptive bilateral filtering. While applying median filtering, artery was
detected in only 180 images and while applying adaptive bilateral filtering 187 images
yielded expected output. Thus, speckle removal did not improve the performance and
actually degraded the performance in case of median filtering. This might be due to the
fact that roundness of the artery might be negatively affected / reduced while applying
the noise removal step.
4.2 Cyst localization

A cyst is generally a closed sac which may contain fluids. The properties which we have
used to formulate the algorithm are:
• Cyst generally appears much dark, i.e. pixel values corresponding to cyst are very
less.
• Cyst is elliptic or almost round.
Following algorithm has been devised for cyst localization:
i. Let the input image be Icyst1 .
ii. A threshold value is calculated by estimating the 20 percentile value of the all the pixel
values present in the image. The image is binarized with this calculated threshold
value to obtain Icyst2 .
iii. The binarized image is inverted to obtain Icyst3 . So, darker regions in the original
image correspond to the white pixels in the binarized image and whiter regions in the
original image correspond to the black pixels in the binarized image.
iv. Morphological closing and opening operations are successively applied on the bina-
rized image to obtain Icyst4 .
v. Connected component detection is done on Icyst4 to find white connected components.

Very small white connected components are removed from further consideration.
vi. Among the remaining white connected regions, roundness metric mround is calculated
for all the regions.
19
Figure 4.2: Detected cyst shown in red
vii. The region with the maximum value of the roundness metric is declared to be the
region corresponding to the cyst.
viii. Centroid of the identified connected component is calculated to estimate the center
of the cyst area.
ix. The major axis length and minor axis length of the ellipse having same normalized
second central moments as the identified connected component is used along with
identified center to define an ellipse. This is defined to be the region corresponding
to the cyst.
As any publicly available database for cyst is not available to the best of our knowledge,
so the algorithm was tested on few images collected from internet.
4.3 Gallbladder stone detection

Gallbladder stone is an abnormal growth within the gallbladder. Following properties of
the gallbladder stone has been used to formulate the algorithm:
• Gallbladder stone generally looks whitish in the ultrasound images.
• It lies within the gall bladder which generally looks darker in the ultrasound images.
Using the above properties, Following algorithm has been devised for gallbladder stone
detection:
i. Let the input image be Igall1 .
20
(a) Input image (b) Output image (pixels probable to be gallstone
shown in white)
Figure 4.3: Illustration of gallstone detection
ii. Median filtering is applied on Igall1 to remove local variation. The resulting image is
Igall2 .
iii. Canny edge detection is applied on Igall2 to obtain edge detected image Igall3 .
iv. Morphological closing operation is applied on Igall3 to obtain Igall4 . The structuring
element used is line with varying angle (0◦ to 90◦ ).
v. White connected regions with very small filled area are removed from the image to
obtain Igall5 .
vi. The connected components are detected from the inverted version of the image Igall5 ,
i.e. dark connected components are detected. Let the dark connected component
detected image be Igall6 .
vii. Now for any white pixel in the image Igall5 , we traverse from the same pixel co-
ordinate in Igall6 towards left to reach the first dark connected component, i.e. first
pixel with value one in Igall6 . Similarly, we traverse towards right to find the first right
dark connected component, i.e first pixel with value one in Igall6 . If these two pixels
belongs to the same dark connected component, then the pixel in Igall5 is assumed
probable to be a part of a gallstone.
As any publicly available database for gallstone is not available to the best of our
knowledge, so the algorithm was tested on few images collected from the internet.
21
Chapter 5
Machine Learning Based CAD
Machine learning can be used in CAD whenever we cannot directly write an explicit
program to solve a given problem. The relation between input image and the output
inference is sometimes intricate. Mere human observation is not sufficient to find out such
relations. Moreover, even if some complicated relations are discovered through rigorous
observation, those relation may not generalize well. The relationship between input and
output may hold differently in different time / environment. So, rather than writing
explicit different programs for all such different circumstances, it is preferable to design
general purpose systems which would adapt to the circumstances.
Even the machine learning methods can be classified into various groups like super-
vised learning, unsupervised learning, semi-supervised learning. If a substantially large
amount of dataset (sets of input images and inferred result from it) is available to us,
then supervised learning is most suitable for use in CAD applications. Supervised learn-
ing algorithm can learn from the given example data. It learns a mapping from the input
to the output while a supervisor provides the correct values [29]. In the unsupervised
learning, only input data is given and the regularity / statistical similarity in it is found
by the machine learning system.
While designing a machine learning based system, some of the important factors to
be taken into account are:
1. Choosing suitable features
2. Dimensionality reduction
3. Choosing suitable classifier
22
5.1 Features for classification
To develop a ultrasound image processing based CAD system, selection of suitable texture
feature is required. Perception of texture is dependent on the spatial arrangement of
intensity values. Texture is characterized by the spatial distribution of intensity values in
a neighbourhood and so it is an area attribute.
Some of the popularly used texture features are:
5.1.1 First order intensity histogram features

First order gray level statistical features[30] can be used for ultrasound image classifi-
cation. Most commonly used first order gray level statistical features are mean (m1 ),
variance (µ2 ), skewness (µ3 ) kurtosis (µ4 ). If random variable I represents the gray level
of the image region, then, the first order histogram can be defined as p(I), where [31]:
number of pixels with gray level I

p(I) = (5.1)
total number of pixels in the given region
Now mean can be defined as:
Ng −1
X
m1 = Ip(I) (5.2)
I=0
where Ng is the nos. of possible gray levels. Similarly, variance, skewness and kurtosis
can be defined as [31]:
Ng −1
X
µ2 = (I − m1 )2 p(I)
I=0
Ng −1
X
µ3 = (I − m1 )3 p(I)
I=0
Ng −1
X
µ4 = (I − m1 )4 p(I)
I=0
All these first order statistical features has different physical significance. The variance
can be interpreted as the measure of the deviation of the intensity level from the mean
intensity level. The skewness can be interpreted as the measure of histogram asymmetry
around the mean [31].
To capture the first order statistics, the semivariogram approach can also be used [32].
23
5.1.2 Gray-level co-occurrence matrix
Though the features generated from the first order gray level statistical features provides
information related to the gray level distribution of the image but they do not provide
any information regarding the relative position of various gray levels within the image
[31]. To exploit the information embedded in the gray level configuration, various second
order statistical features are defined. Gray-level co-occurrence matrix (GLCM) is most
popular among those.
GLCM can be defined a matrix of relative frequencies Pθ,d (I1 , I2 ) which describes how
frequently two pixels with gray-levels I1 , I2 appear in the window separated by distance
d in direction θ. Thus co-occurrence matrix is function of two parameters, the distance
between two pixels d and the relative orientation θ. Generally, θ = 0◦ , 45◦ , 90◦ , 135◦ is
considered. Various features can be defined based on the GLCM matrix.
Haralick et al. has described a set of textual features based on the GLCM matrix.
Some important features among these are Angular Second Moment (ASM) or energy,
contrast, correlation, homogeneity and entropy. These features are defined as below [31]:
X
ASM = P (I1 , I2 )2
I1 ,I2
X
Contrast = |I1 − I2 |2 log P (I1 , I2 )
I1 ,I2
X P (I1 , I2 )
Homogeneity =
I1 ,I2
1 + |I1 − I2 |2
Entropy = −P (I1 , I2 ) log P (I1 , I2 )
ASM measures the smoothness of the image. It is the sum of the squared elements in
the GLCM. More smooth the image, more will be the value of the ASM. This is also called
energy. It measures textual uniformity. ASM has high value when gray level distribution
has constant or periodic form. Contrast measures local variations and takes high value
when there is high local variation. Homogeneity measures high values for low-contrast
images. Entropy is an index of randomness and it takes low values for smooth images. It
measures complexity of the image. Entropy is inversely correlated to energy. GLCM based
features has been used in many works related to ultrasound image processing [33, 30, 23].
5.1.3 Fractal features

Fractal is a concept more popular among computer graphics society which can also be
used for image texture analysis. A fractal surface is considered self- similar, i.e., the
24
local surface at a finer scale resembles the global surface at a coarser scale [34]. Fractal
dimension of an object is greater than the Euclidean dimension. Concept of the fractal
dimension can be regarded as an indicator of the surface roughness. It has been used in
breast ultrasound image classification [35]. Various features can be computed from the
concept of fractal. One such feature set is Segmentation-based Fractal Texture Analysis,
or SFTA) proposed by Costa et al. in [36]. To extract SFTA features, input image is
decomposed into a set of binary images by two-threshold binary decomposition algorithm.
From the set of binary images, fractal dimension of the resulting regions are computed to
describe the texture pattern.
5.1.4 Other features

Gray level run length matrix has also been used for ultrasound image processing [37].
Power spectral features can also be used [30]. To obtain power spectral features, 2-D
FFT algorithm is applied to the image which is followed by computation of the magnitude
square of the spectral components. It results in the power spectrum P (u, v). The power
spectrum is used to compute the various power spectral features.
5.2 Dimensionality reduction

Dimensionality reduction is often required while designing ultrasound image processing
based CAD systems. It is an optional step which may be required in case of large dimen-
sionality of the feature set. Often many of the textual features are correlated and hence
redundant. Also processing with many features is more expensive. Also, as per Hughes
phenomenon [38], the requirement for the large training sample increases with the dimen-
sionality of the input feature set. Thus increasing numbers of features may degrade the
performance of the classifier. E.g. if 14 Haralick features are taken at four different angles
and four different distance, then nos. of features would be 224. So to reduce dimension
in such case dimensionality reduction algorithms can be used. The methods for reducing
dimensionality can be divided into two types : feature selection and feature extraction
[29]. In feature selection, we select a subset of features which give most of the information
and we discard the rest. In feature extraction, we find a new set of features which is a
combination of the earlier set of feature but fewer in dimension. Principal component
analysis (PCA) is one of the popular feature extraction method. In PCA, the criteria to
be maximized is variance [29]. In [39], in one of the experiments, the gray maps of varying
25
window size were extracted (from 8 pixels × 8 pixels to 25 pixels × 25 pixels). Then the
resulting vector of 64–625 dimensions was projected using PCA into spaces ranging from
10 to 50 dimensions.
5.3 Classification methods

After suitable selection of features, machine learning based techniques can be used for
classification. In [30], artificial neural network based technique has been used. A multi-
layer back-propagation network has been trained based on the selected features. In [39],
prostate has been categorized into various categories like cancerous, BHP, prostaitis using
K-Nearest Neighbor (kNN) classification. The same method can be used for kidney. In
[40], ovary has been classified into normal, cystic and polycystic using SVM based tech-
nique. In [41], breast tumor diagnosis system has been designed using SVM. As pointed
out by authors in [41], advantage of using SVM is that it is more faster and stable than
neural network based systems.
5.3.1 Using kNN

The kNN algorithm is a simple yet powerful algorithm. It is a non-parametric algorithm
because it does not make any assumption regarding the underlying class distribution. It
has been used in many works related to ultrasound image processing [39, 42]. It has a
minimal training phase. But testing phase is expensive in terms of time. It assigns the
input to the class having most number of examples among the k neighbors of the input,
where all the neighbors are generally given equal weightage. Suppose a training dataset
{xi , yi }N
i=1 is given where xi corresponds to the input data and yi corresponds to the
known output label. Now when a unknown test data is presented to the system, k nearest
neighbor of it in the given training dataset is found. In case of ties, it can be broken either
arbitrarily or based on a weighted vote [29]. Neighbors are selected based on a suitable
metric, e.g. euclidean distance. When k = 1 it is called nearest neighbor algorithm. any
variants of kNN like adaptive kNN [43] have also been proposed in literature.
5.3.2 Using SVM

Support Vector Machine [14] is from the class of maximum margin classifiers. It has been
used in many works related to ultrasound image classification:
26
• Breast ultrasound image classification [44, 45].
• Thyroid ultrasound image classification [46].
• Prostate ultrasound image classification [47, 48, 49].
It is a non-parametric technique as it does not make assumption about underlying data

distribution. It performs classification by finding a decision surface between two classes
that has the maximum distance to the closest points in the training set. These closest
points are known as support vectors. SVM can be used for non-linearly separable data
also, hence its usefulness. The main trick behind success of SVM to handle non-linear
data is to map the features of the input space into higher dimensional feature space. This
is achieved by using non-linear transformation called kernel function. In addition to this:
1. SVM has flexibility to deal with high-dimensional data, which is important in our
case. The upper bound on the generalization error does not depend on dimension
of the feature space in case of SVM. If four first order features mentioned and 14
Haralick features are taken at four different angles and four different distances, then
the data would have dimension of 228. Hence our data would have high dimension
and here SVM plays its role.
2. SVM shows high accuracy compared to most of the supervised learning algorithms.
3. It belongs to category of the kernel methods. Kernel based algorithms can be

formulated as convex optimization problem and so there is optimum solution which
can be solved analytically. This is one of the main reasons of using SVM [29].
Let there be a training data with two classes, and labels for the classes be +1 (assume,
corresponding to cancerous) and -1(assume, corresponding to non-cancerous). Let the
data be {xi , yi }N
i=1 . Inner product is defined as:
X
wT x = w i xi (5.3)
i
Let there be a discriminant function of the form:
f (x) = wT x + b (5.4)
where b is called the bias.

A linear classifier is based on the above discriminant function and the boundary region
27
between the positive and the negative regions is called the decision boundary of the
classifier, which is defined by a hyperplane given by the following equation:
x : f (x) = wT x + b = 0 (5.5)
As the decision boundary is linear, so it is called a linear classifier. In many applications,

non-linear classifier yields better result than the linear classifier, but linear classifier has
many advantages to use. One simple approach to develop non linear classifier out of the
linear classifier is to map the data from the input space X to a feature space F using
some non-linear mapping φ : X → F so that in the new space F the new discriminant
function will be given by [50]:
f (x) = wT φ(x) + b (5.6)
A maximum margin classifier can be defined as a function which maximizes the geometric
margin [50]. Maximizing the geometric margin is to minimize ||w||2 . So we get need to
minimize ( 21 )||w||2 subject to:
yi (wT xi + b) ≥ 1 ∀i = 1, 2, ..., n (5.7)
This equation classifies each example correctly which is possible for completely sepa-
rable data. For data which is not completely separable / there are some misclassified data
even in the training data, often greater margin can be achieved by allowing the classifier
to mis-classify few points. Thus to allow errors, we introduce a new term ξi and we modify
(5.7) as:
yi (wT xi + b) ≥ 1 − ξi ∀i = 1, 2, ..., n (5.8)
Now the parameter to minimize will be modified and a new term will be introduced to
penalize the misclassification and the margin error. So now we will minimize w.r.t. w
and b, ( 12 )||w||2 + C ni=1 ξi , where (5.8) is the constraint.
P
The above formulation is called the soft margin SVM. The above equation is quadratic in
convex and can be solved using any convex optimization technique like Lagrange multiplier
[51], Sequential Minimal Optimization (SMO). Using the method of Lagrangian multiplier,
we can obtain the dual formulation of the above equation which is expressed in terms of
the variable αi [50]:
n n n
X 1 XX
maximize α αi − yi yj αi αj xTi xj (5.9)
Pn i=1
2 i=1 j=1
subj. to i=1 yi αi =0
o≤αi ≤c
28
The example xi for which αi > 0 are the points on the margin and are called the support
vectors. This dual formulation depends on data only through dot product. The dot
product can be replaced with a non-linear kernel function and in this way can handle
non-linear separation case. The basic kernels used are [52]:
1. Linear k(xi , xj ) = xTi xj
2. Polynomial k(xi , xj ) = (γxTi xj + r)d , γ > 0
3. Radial basis function (RBF) k(xi , xj ) = exp(−γ||xi − xj ||2 ), γ > 0
4. Histogram intersection kernel

where γ, r, d are kernel parameters. Mostly used kernel is RBF due to its robustness and
it has been found to outperform other kernels [53]. It is suitable for non-linearly separable
data. The neighborhood parameter of RBF which gives relative weight of other samples
to a given sample can be selected by trial and error method such as grid-search [29]. While
using grid-search for RBF kernel, grid-search needs to be done in two dimension, to select
C and γ. SVM has been implemented using LibSVM [54].
5.3.3 Classification using k-Means clustering method

k-Means algorithm is a un-supervised clustering algorithm [29]. Given a dataset it groups
the data into few cohesive clusters. We are given a dataset {xi }N
i=1 but no label is generally
given. The numbers of clusters k is a parameter of the algorithm, which denotes the
numbers of clusters in the data. In the k-Means algorithm, the k cluster centroids are
initialized randomly at beginning. Then each training examples are assigned to the closer
cluster centroid and the cluster centroids are recomputed according to the mean of the
points assigned to it. This second step is repeated until convergence.
The k-Means algorithm is susceptible to the local minima. Though most of the times
it works fine, but it can yield solution which is not satisfactory. It is a clustering method,
not used for supervised classification. We have devised an algorithm where k-Means
algorithm is used for classification. The proposed algorithm is as follows:
i. All the features of the training data are separately normalized to values between 0
and 1. Suppose there are N data in the training set and a certain feature value for
this N data is {fi }N
i=1 . Let, minima of the feature is M inf and the maxima of the
feature is M axf . Then the feature value for all the N data point is recomputed as:
fi − M inf
fi = (5.10)
M axf − M inf
29
ii. Feature values of the test data are also normalized with the same values as the training
data.
iii. PCA is applied on the training data. Thus training data is transformed to new feature
set. Only those feature set which accounts for 90% of the variance in the training
data is retained. Other features are deleted. Test data is also transformed using the
same vectors used to transform the training data. Same set of features as training
data are also retained in the test data.
iv. At the end of the each training data, we append its label (converted to numeric value),
i.e. +1 for one class and -1 for the other. At the end of the each test data, we append
a label 0. Here it is to be noted that 0 is the mean of +1 and -1. Now this numeric
values will also be treated as a separate feature for the data. While for the training
data, the values indicate clear separation, value 0 indicates uncertainty of the test
data.
v. Now training data and test data are combined and k-Means algorithm is applied on
it with k = 2.
vi. In one of the clusters more training data with label equal to 1 will be present. Test
data present in that cluster are assigned label +1. In the other cluster more training
data with label equal to -1 will be present. Test data present in that cluster are
assigned label -1.
Instead of using PCA, kernel PCA can also be used in this method, which is a kernel-
ized version of the PCA [29].
5.3.4 Combining SVM and k-Means

In [55] Gu and Han have proposed a Clustered Support Vector Machine (CSVM) which is
a divide and conquer method. In this approach, data is first divided into several clusters
by k-means clustering. Then a linear SVM is applied on each cluster. The author has
proposed it as an alternative for the use of non-linear classifier.
Here we are proposing to use the same divide and conquer method. But however we
argue that the same approach can be used for non-linear classifier also. There might
be data like positive data and negative data in concentric circles, where they are indeed
linearly non-separable. Due to the fact that sometimes some class of the data are combined
to form a single class, it may happen that in the feature space such linearly non-separable
30
positive data and negative data lies in two or more different zones of the feature space,
where each zones are not related with each other. So we are proposing to use RBF kernel
SVM with the k-Means algorithm. The algorithm we have used is described as following:
Training Phase
i. Training data is clustered into k numbers of clusters by using k-Means algorithm.

Here, k is to be empirically chosen and should be chosen such that it is equal to or
greater than numbers of the clusters present in the data.
ii. All such clusters are separately trained with SVM with RBF kernel.
Test Phase
i. Distance of the test data is measured from the centroid of all the k clusters. The
cluster from which data is nearest is assumed to be cluster where the test data belongs.
ii. The test data is tested with the corresponding SVM of that cluster and the output
label is predicted.
5.3.5 One class SVM and SVDD

When there is a training data of a specific class and we would like to determine whether
new test data is member of that specific class (the class determined by our training
data), then one class SVM can be used [56]. Support vector data description (SVDD)
was proposed by Tax and Duin [57]. It can also be used for similar uses like one class
SVM. It is a method to find boundary around a given data set and thus used for outliers
detection. It has been implemented in LibSVM [58]. The method can be used when we
want to detect data of one type only and would like to classify all other data as outliers.
5.4 Results
Publicly available database for ultrasound images are very scarce. We have used the
prostate database which has been used in [39]. It is a prostate cancer database. In the
database, along with the images, two points for each image has been provided which
defines a region of interest (ROI). In [39], the pixels in that ROI are analyzed.
31
In our work, we have tried to differentiate prostatitis and tumor in advanced stage.
Henceforth we would be referring advanced stage tumor as cancer. Prostatitis is a medical
complication involving inflammation of the prostate.
We have used sensitivity and specificity to evaluate the performance of the system.
Suppose we have two class, i.e. C1 and C2 and our prime objective is to identify C1 , then
sensitivity (S1 ) as:
Numbers of occurrence correctly detected as C1
S1 = (5.11)
Numbers of actual occurrence of C1
We define specificity (S2 ) as:
Numbers of occurrence correctly detected as C2
S2 = (5.12)
Numbers of actual occurrence of C2
Sensitivity is also known as true positive rate. In our case if we define C1 as cancer and
C2 as prostatitis, then, sensitivity is redefined as:
Numbers of occurrence correctly detected as cancer
S1 = (5.13)
Numbers of actual occurrence of cancer
Specificity is redefined as:
Numbers of occurrence correctly detected as prostatitis
S2 = (5.14)
Numbers of actual occurrence of prostatitis
We define another parameter, Stotal as
Stotal = S1 + S2 (5.15)
For an ideal classification system, sum of sensitivity and specificity (Stotal ) would be equal
to two as both of those are expected to be one for an ideal classification system. But
ideal classification system does not only depend on the classifier, but also on the classes
to be classified. Classes should be completely separable to yield sum of sensitivity and
specificity as two. But in most practical system, classes are not completely separable.
Algorithms has been tested with window size (wsize ) equal to 7, 15, 31, 121. Four
GLCM features contrast, correlation, energy, homogeneity were calculated at d = 1, 2, 3, 4
and at varying angles θ = 0◦ , 45◦ , 90◦ , 135◦ . Thus we get 64 features. We defined a
straight line joining the two points given in the database which defines the ROI of an
image. For wsize = 7, 15, 31, points were sampled along this straight line and centering
on those points the window was defined and the features were calculated. Then median
and standard deviation of all those features were calculated to get a feature vector of size
128 for an image. 45 images per class were used for training and rest of the images were
32
used for testing. The algorithms failed to perform in satisfactory manner with window
size 7. When using window size 15, applying SVM with RBF kernel and C = 0.1, γ = 0.5
sensitivity of 0.65 and specificity of 0.53 was attained. kNN failed to perform satisfactorily
with window size 15 and with k = 3 sensitivity of 0.30 and specificity of 0.74 was attained.
When using window size 31, applying SVM with RBF kernel and C = 0.1, γ = 0.5 sen-
sitivity of 0.75 and specificity of 0.47 was attained. kNN did not yield satisfactory result
with this window size also.
When using wsize =121, only one window was defined per image, centering on the
middle point of the straight line joining two given points in the database. A feature
vector of length 64 was defined for this window. When using kNN, with k = 3, sensitivity
of 0.60 and specificity of 0.47 was attained. With k = 5 sensitivity of 0.7 and specificity
of 0.52 was attained. Variation of Stotal with k has been shown in table 5.1. When
using SVM with linear kernel with C = 0.1 sensitivity of 0.65 and specificity of 0.66 was
attained. With C = 0.2 sensitivity of 0.72 and specificity of 0.5 was attained. When
using the RBF kernel, with C = 0.1, γ = 0.5 sensitivity of 0.70 and specificity of 0.65
was attained. With C = 0.1, γ = 0.7 sensitivity of 0.70 and specificity of 0.73 was
attained. When using the k-Means based method described in section 5.3.3 sensitivity of
0.70 and specificity of 0.58 was attained. Using the method combining SVM and k-Means
as described in section 5.3.4, for C = 0.1, γ = 0.5 sensitivity of 0.90 and specificity of
0.45 was attained. Variation of sum of sensitivity and specificity (Stotal ) with γ when C
is fixed at 0.1 has been shown in table 5.2. Variation of Stotal with C when γ is fixed
at 0.5 has been shown in table 5.3. The same set of experiments were performed with
applying PCA and transforming feature vector to length 15. But that did not induce any
significant change in the performance.
The SFTA features [36] were also calculated and appended with the calculated GLCM
features. But the performance did not show any improvement. When using the GLCM
features and the SFTA features together, when using SVM with RBF kernel, with C = 0.1,
γ = 0.5 sensitivity of 0.75 and specificity of 0.56 was attained. When using the linear
kernel with C = 0.1 sensitivity of 0.70 and specificity of 0.55 was attained. When using
kNN with k = 3 sensitivity of 0.6 and specificity of 0.47 was attained. When using the
k-Means based method described in section 5.3.3 sensitivity of 0.65 and specificity of 0.60
was attained. Thus we conclude that adding SFTA features did not yield any additional
benefit in this case. In fact the performance degrade a bit which might be due to Hughes
phenomenon [38]. Similarly combining the features from window size 15, 31 with features
from window size 121 also degraded the performance. Another experiment was conducted
33
k Sum of sensitivity
and specificity (Stotal )
1 1.11
3 1.07
5 1.22
7 1.23
9 1.25
11 1.25
13 1.26
15 1.26
17 1.21
19 1.21
21 1.22
23 1.23
25 1.19
Table 5.1: Stotal variation with k when using kNN algorithm
by obtaining results separately from features of window size 15, 31, 121 and performing
voting based on those results to obtain the final result. Voting is a simple method to
combine multiple classifiers [29]. No improvement was observed and the performance
actually degraded from the performance when window size is 121. With C = 0.1, γ = 0.5
sensitivity of 0.70 and specificity of 0.54 was attained.
Though we failed to attain very high sum of sensitivity and specificity, it is primarily
due to the poor image quality in the database and high overlap between the classes. An
experiment using SVDD was conducted to verify the overlap between the classes. SVDD
was trained using RBF kernel and training data corresponding to cancer. While tested
on the cancer test data, it accurately identified 65% of the data as belonging to cancer
data. But, when it was tested on the prostatitis data it identified 57% of the prostatitis
data also as belonging to cancer data. It indicates a high overlap between the data. It
is to be noted the authors in [39], the work where this database was originally used, also
failed to attain satisfactory results. The authors has also emphasized that they failed to
achieve desirable results mainly due to poor quality images in the database.
34
γ Sum of sensitivity
0.1 1.30
0.2 1.25
0.3 1.24
0.4 1.31
0.5 1.35
0.6 1.40
0.7 1.43
0.8 1.41
0.9 1.38
1.0 1.37
1.3 1.27
1.5 1.26
Table 5.2: Stotal variation with γ when C is fixed at 0.1
C Sum of sensitivity
0.1 1.35
0.2 1.32
0.3 1.24
0.4 1.28
0.5 1.27
0.6 1.27
0.7 1.26
0.8 1.25
0.9 1.25
1.0 1.25
1.3 1.14
1.5 1.13
Table 5.3: Stotal variation with C when γ is fixed at 0.5
35
5.5 Adapting to domain change
The ultrasound images taken by one device at certain settings may differ largely from the
images taken by other device at certain other settings. Generally, the literatures related
to ultrasound image processing also reports the device from which the images has been
collected. Generally, the training and test images are taken from the same device. But, if
we aspire to design a CAD system, which would work for images taken from any device,
then we would require to design in such way that system would work even if training
and test images have been taken from different devices. Fortunately, solution for such
problem has already been explored in literature, though not in context of ultrasound
image processing. This problem can be considered to be a domain adaptation problem,
which has been discussed in [59].
In supervised learning, labeled training samples T = (xi , yi ) are drawn from domain
D and unlabeled test data are also drawn from the same domain D. So it is possible
to estimate a distribution P̂ (x, y) from training data which will correctly model the true
distribution P (x, y) which governs D.
In unsupervised learning, labeled training samples not available. Only unlabeled data
available. We have to infer an approximation P̂ (x, y) for the true unknown P (x, y) by
exploiting the statistics of the given data.
Transfer learning refers to problem of retaining knowledge from one task / domain to
another task / domain.
Single task transfer learning is of two types:
1. Sample selection bias: Unlabeled test data are drawn from same domain D of the
training data. But very small amount of the training data leads to poor estimation
of the P̂ (x, y) which hence do not approximate well true distribution P (x, y).
2. Domain adaptation: Unlabeled test data are drawn from a target domain Dt
different from the source domain Ds of the training samples.
Let, source domain data set be Xs = {xsl }N

l=1 i.e. composed of N source domain data
Let, target domain data set be Xt = {xtu }M
u=1 i.e. composed of M target domain data
Let xs and xt be d-dimensional vectors related to Ds and Dt respectively.
In domain adaptation problem, our goal is to develop a domain adaptation classifier
g(x|Xs , Xt , ) which would would accurately classify target domain samples Xt from Dt by
exploiting labeled samples Xs from Ds . The model adopted for classification is described
by parameter .
36
Chapter 6
Fetal Ultrasound Image Processing
Ideally gender ratio of male and female should have been 1:1. But it has never been true.
Leaving few exception, in all society across the globe, male child are preferred. This unfair
preference has been aided recently by growth in technology. With advent of ultrasound
imaging devices, gender of the fetus can be easily determined. Ultrasound imaging device
can be considered as a bliss for mankind and has catalyzed the recent advancement of the
health care system. But unfortunately this technology has been misused causing worser
gender imbalance. Gender-selective abortion is a common practice mainly in the third
world nations. While the source of the problem is technology, the problem needs to be
solved technologically. One of the possible solution is to segment the genitalia from the
fetal ultrasound image so that it can be blocked from being displayed in the ultrasound
imaging device [60]. This will contribute to curbing of female foeticide. As preventing
display of the genitalia of the fetus will stop knowing the gender of fetus, gender based
abortion is expected to reduce.
6.1 Analyzing reasons for preference of male child

Almost in all parts of the globe, male child has been preferred historically to female child.
In India, as evident from table 6.1 [61, 62], sex ratio has deteriorated over time. Some
reasons behind this are:
1. Almost in all societies there is trend of marrying off girls and sending to the hus-
band’s house. So there is a stereotype that male will look after parents.
2. There is a belief that male child will carry forward family lineage. They will be the
bread-earner for the family.
37
Year 0-6 yrs girls per 1000 boys
1961 976
1971 964
1981 962
1991 945
2001 927
Table 6.1: Gender ratio (0-6 yrs) change over time
State Sex ratio Literacy rate

Uttar Pradesh 912 67.68
Bihar 918 61.80
Rajasthan 928 66.11
Jammu & Kashmir 889 67.16
Madhya Pradesh 931 69.32
Kerala 1084 94
Himachal Pradesh 972 82.80
Tamil Nadu 996 80.09
Delhi 868 86.21
Table 6.2: Gender ratio of few states of India
3. In countries like India, one of the main reason is dowry system [63]. The amount
to be paid while marrying a girl is very high and makes many family pauper. So
people prefer male child.
4. In the table 6.2 [64], gender ratio and literacy rates of few states of India has been
shown which shows strong correlation between sex ratio and literacy rate. From
this we can conclude that poor education is also one of the reason behind skewed
sex ratio. Only in some exceptional states like Delhi, despite higher literacy rate,
sex ratio is skewed.
5. Polygamy is also a cause behind male preference. In societies where polygamy exists,
people with better economic conditions are more biased towards male child. This
has been explained by Robert Trivers and Dan Willard (Trivers-Willard hypothesis,
[65]).
38
6. In many cases, parents are neutral in case of 1st child, but preference for son in-
creases with subsequent child.
6.2 Gender imbalance in India and other countries

A highly skewed gender ratio has been observed in some of the Asian countries like India,
South Korea, Georgia, Azerbaijen, China, and Armenia [66]. All of these nations are
densely populated and forms a large chunk of the world population.
A drastic reduction in female infant has been reported in Indian census. The national
gender ratio between 0 to 6 yrs has been reported to be 927 females to 1000 males
[62]. From table 6.2, we can see that gender ratio is more skewed in comparatively less
developed states.
In as many as 36 nations, gender determination has been banned [67]. In India also, Pre-
natal Diagnostic Techniques (Regulation and Prevention of Misuse) Act, 1994, has been
implemented to stop female foeticide. This law bans anybody from using any technology
to detect gender of unborn fetus.
6.3 Consequences of gender imbalance

1. Skewed gender ratio is indirectly promoting lesser female participation in the demo-
cratic decision making bodies. Lesser participation in decision making bodies is
causing male dominance in societies. Thus it is iteratively increasing female perse-
cution.
2. Increase in number of inhuman activities like rape etc.
3. Trend of trade of female (women trafficking) increases. Women are sold to areas
where foeticide has caused skewed sex ratio. Such incidence has been reported in
Haryana. Skewed sex ratio of Haryana has caused trafficking of girls from other
states to Haryana [68].
6.4 Misuse of ultrasound technology

It has been detected that majority of the pre-natal gender detection occurs via ultrasound
devices [61]. Though there was always a preference towards male child, but it has now been
catalyzed largely by advancement of technology. Mere desire for male child could not have
39
translated into female foeticide unless it was aided by advanced ultrasound technology.
It is clearly evident from table 6.1 that with advent of the ultrasound technology, gender
ratio has deteriorated in faster pace. A steep decrease is observed from 1981-2001. Though
ultrasound imaging technique was invented earlier, it ushered in the same period. In the
early 1980s, computer software and ultrasound imaging technology combined and age of
modern ultrasound stared [69]. Ultrasound devices are becoming cheaper, portable and
accuracy has increased manifold increasing its misuse. Even companies like GE, who are
major saler of Ultrasound device, has agreed that the misuse exists and has taken steps
to reduce misuse of ultrasound technology and promoting ethical ultrasound [70]. It has
been reported by Burns [71] that in India even local entrepreneurs are running portable
clinics for sex determination in running vans.
Due to easy availability of technology, sex ratio is getting worser in the cities and towns
[61]. Thus we see that misuse of ultrasound technology is one of the major reason behind
skewed gender ratio. While problem has been scaled up by technology, we cannot rely
simply on ethics of doctor or wait for government to implement strict rules. There is a
urge to solve this problem using technology itself. More sophisticated techniques needs
to be devised using which misuse of ultrasound device can be curbed.
6.5 Methods
Not many literatures exist related to this topic and also there is no publicly available fetal
ultrasound image database best to our knowledge. A method has been proposed by Tang
and Chen [72] where authors has proposed a two stage algorithm. In the first stage (called
as ”rough” stage) pixels of interest are identified by feature based detection. In the final
(called as ”fine” stage) a supervised learning framework has been used. A method has been
proposed in [60], a sliding window approach is used where for the training images, features
for each window is calculated and those are marked as whether containing genitalia or
not. Then the training data is used to train a SVM based system. The trained system is
then used for the identifying the window containing genitalia.
Generalizing, when we have a large dataset, a supervised learning framework can be
used, which can be summarized in the following steps:
1. Manually segmenting the images of the training dataset into genital and non-genital
regions.
2. Calculating textual features corresponding to both the regions and using those es-
40
timated textual features values to train a supervised learning system. Any suitable
supervised learning algorithm can be used. Samples corresponding to the genitalia
region can be treated as positive samples and samples corresponding to the non-
genitalia regions can be treated as negative samples.
3. Using the trained system for further classification.
But currently no such database is publicly available. Supervised learning algorithms are
not very practical when such a dataset is not available.
Suitable features can be identified and an explicit algorithm can be implemented for
the genitalia detection. For feature based identification, characteristics / features need to
be identified which distinguishes the genitalia region from the non-genitalia region. Some
of the features which have been identified to be suitable [72] are discussed in following
subsections.
6.5.1 Amniotic fluid

Amniotic fluid is present around the genitalia. Amniotic fluid is the protective fluid which
is in the amniotic sac of the pregnant female. It is very dark, so it is supposed to produce
high contrast with the genitalia which is comparatively much brighter. So by assessing
contrast in windowed regions, candidate regions for genitalia can be found.
6.5.2 Curvy contour

Contour of the fetal genitalia is curvy. So large numbers of corner points are expected
to be present in the genitalia in comparison to the other regions. So by implementing
corner point detection algorithm and then counting numbers of corner points in windowed
regions, genitalia might be located.
6.5.3 Shape based features

Various shape based features can be used. E.g. male genitalia are approximately com-
posed of two elliptical regions. Similar features can be found out for the female genitalia
and used. More such shape based features can be found out by experimenting with min-
imal size dataset.
A major aspect in designing such a system would be the choice of window area. If we
take a small window, then the features mentioned would not be captured appropriately.
41
(a) Male genitalia, taken from [72] (b) Our intended output
Figure 6.1: Male genitalia composed of two elliptical regions
E.g. small window won’t capture the contrast between the amniotic fluid and the genitalia
properly. So, the window should be appropriately chosen to to capture these properties.
42
Chapter 7
Conclusion and Future Work
7.1 Conclusion
Ultrasound imaging is currently very popular medical imaging modality. More automation
can be engineered in the field of ultrasound imaging to facilitate better health services.
We tried to explore the field in depth and breadth.
Speckle noise removal is very important for ultrasound image processing. Some of
the existing techniques of speckle noise removal has been studied. Few of those has
been implemented and performance has been compared in terms of various statistical
parameters. Adaptive bilateral filtering and squeeze box filtering outperformed other
methods.
The computer aided diagnosis has been studied from both machine learning based
approach and non-machine learning based explicit approach. Explicit algorithm for cyst
detection, artery detection, gall bladder stone detection has been proposed and imple-
mented. Artery detection has been done both on the unmodified input image and also on
the speckle removed input image. Our algorithm for artery detection performed better
on the unmodified image. It may be inferred that roundness property of the artery gets
distorted during speckle removal.
The aspects involved in designing a machine learning based CAD has been discussed
in details. A GUI has been developed which can help radiologists to provide researchers
ultrasound images with ground truth data. Methods like SVM, kNN has been studied,
tested on available data and results has been compared. A novel classification method
based on k-Means clustering has been proposed and implemented. A method combining
SVM with kNN has also been studied and implemented. SVM with RBF kernel performed
better than other methods. The k-Means based method also yielded satisfactory result
43
both in terms of sensitivity and specificity. In our case, larger window size for capturing
features yielded better result. Appending SFTA features with GLCM features did not
improve performance. Also combining the features from different window size degraded
the performance. Reducing feature size by applying PCA also did not improve perfor-
mance in terms of sensitivity and specificity. The experiment with SVDD also proved the
overlap between classes in the data.
Adapting to changing domain is an important aspect to incorporate adaptability in
the CAD systems. A SVM based method for domain adaptation has been studied.
Misuse of ultrasound technology for fetal gender determination and its ill impact on
gender ratio has been studied. Analysis has been done about how growth of ultrasound
imaging has increased illegal abortion. Possible ways to detect the fetal genitalia auto-
matically so that it can be stopped from being displayed has been discussed. Supervised
learning algorithm cannot be used unless there is a substantially large database available.
Hence explicit algorithm needs to be devised for fetal genitalia detection. On successful
implementation of such technique, misuse of ultrasound imaging could be stopped.
Also some other aspects like possibility of implementation of ultrasound image pro-
cessing modules in Android Smart devices has been studied. Opportunities of image
processing in Android device has been explored and simple video processing Apps. has
been developed.
7.2 Future Work

The scope for improvement is large and can be summarized as following:
1. As currently there is a scarcity of free ultrasound image database, it is a major

drawback for research related to ultrasound image processing. Such database needs
to be created. It will benefit the research community to large extent.
2. Both the machine learning and non-machine learning (explicit) techniques for CAD
needs to be studied more subtly. A demarcation needs to be decided regarding
where we can use which type of method. Suitability of different machine learning
methods for different tasks need to be studied in details.
3. Existing texture features need to be studied in more details while aiming at inno-
vating more suitable texture feature for ultrasound images. Ultrasound images are
much different than most of the other images. So it is required to investigate in
more depth for finding suitable feature set for the ultrasound images.
44
4. Deep learning [73] can also be explored for better representation of the ultrasound
images. It needs to be investigated in details.
5. The other dimensionality reduction methods in addition to PCA needs to be stud-

ied. A demarcation needs to be decided regarding when dimensionality reduction is
required.
6. Domain adaptation needs to be studied in details. A suitable database needs to

be devised from at-least two different devices with different conditions to test the
domain adaptation algorithms.
7. Real time implementation issues of the CAD devices need to be studied.
8. Techniques for stopping misuse of ultrasound device for gender determination and
illegal abortion need to be studied in more details. More suitable features need to
be studied for the fetal genitalia detection. Successfully integrating such algorithms
with the ultrasound device is required. The system needs to be designed in such a
way so that it is not easily vulnerable to hackers. So the security aspects also need
to be studied.
9. Implementation of ultrasound image processing algorithms in Android devices /

smart phones needs to be studied in more details. Currently, feasibility of such
system has been studied by implementing small image processing modules. Actual
ultrasound image processing modules needs to be made. In such a system in addition
to the image processing aspect, other aspects like communication would also require
attention.
45
Appendix A
GUI to help radiogist to provide

manually segmented images
To validate the devised algorithms, ultrasound images along with ground-truth data are
required. A Matlab based Graphical User Interface has been developed which can be used
by radiologist / expert doctor in providing us ground-truth data. The expert can identify
the contour corresponding to the organ, input the name of organ and input any special
comment (e.g. whether infected with cancer). All those data get saved to a database. A
screen-shot of the GUI has been shown in figure A.1.
Figure A.1: Screen-shot of the GUI developed to collect ground-truth data
46
References
[1] S. K. Narayanan and R. S. D. Wahidabanu, “A view on despeckling in ultrasound

imaging,” Int. J. Signal Proc., Image Proc. and Patter Recog., vol. 2, pp. 85–98,
September 2009.
[2] F. L. Arbona, B. Khabiri, and J. A. Norton, Ultrasound guided regional anesthesia.

Cambridge University Press, 2010.
[3] M. Ali, D. Magee, and U. Dasgupta, Signal processing overview of ultrasound systems
for medical imaging. Texas Instruments, NOVEMBER 2008.
[4] K. Kaur, “Digital image processing in ultrasound images,” Int. J. Recent Innovation
Trends Comp. Comm., vol. 1, pp. 388–393, March 2013.
[5] N. K. Ragesh, A. R. Anil, and R. Ragesh, “Digital image denoising in medical ultra-
sound images: A survey,” in Proc. ICGST, pp. 67–73, 2011.
[6] K. K. Shung, “Diagnostic ultrasound: past, present, and future,” J. Med. and Bio.
Engg., vol. 31, no. 6, pp. 371–375, 2011.
[7] V. Chan and A. Perlas, Basics of Ultrasound Imaging. Springer, 2011.
[8] S. K. Narayanan and R. Wahidabanu, “A view on despeckling in ultrasound imaging,”

2009.
[9] J.-S. Lee, “Speckle analysis and smoothing of synthetic aperture radar images,” Com-
puter graphics and image processing, vol. 17, no. 1, pp. 24–32, 1981.
[10] K. Kaur, B. Singh, and M. Kaur, “Speckle noise reduction using 2-d fft in ultrasound
images.,” International Journal of Advances in Engineering & Technology, vol. 4,
no. 2, 2012.
47
[11] A. Vishwa and M. S. Sharma, “Speckle noise reduction in ultrasound images by
wavelet thresholding,” International Journal of Advanced Research in Computer Sci-
ence and Software Engineering, vol. 4, no. 6, 2012.
[12] T. Loupas, W. McDicken, and P. Allan, “An adaptive weighted median filter for
speckle suppression in medical ultrasonic images,” Circuits and Systems, IEEE
Transactions on, vol. 36, no. 1, pp. 129–135, 1989.
[13] N. Ragesh, A. Anil, and R. Rajesh, “Digital image denoising in medical ultrasound
images: a survey,” in Icgst Aiml-11 Conference, Dubai, UAE, pp. 67–73, 2011.
[14] C. Bo, G. Zexun, Y. Yang, and L. Xiaosong, “Application of the rough set to im-
age median denoising,” in Software Engineering, Artificial Intelligence, Networking,
and Parallel/Distributed Computing, 2007. SNPD 2007. Eighth ACIS International
Conference on, vol. 1, pp. 75–78, IEEE, 2007.
[15] S. Wu, Q. Zhu, and Y. Xie, “Evaluation of various speckle reduction filters on medical
ultrasound images,” in Engineering in Medicine and Biology Society (EMBC), 2013
35th Annual International Conference of the IEEE, pp. 1148–1151, IEEE, 2013.
[16] M. Nagao and T. Matsuyama, “Edge preserving smoothing,” Computer graphics and
image processing, vol. 9, no. 4, pp. 394–407, 1979.
[17] E. Farzana, M. Tanzid, K. Mohsin, M. Bhuiyan, and S. Hossain, “Adaptive bilateral

filtering for despeckling of medical ultrasound images,” in TENCON 2010-2010 IEEE
Region 10 Conference, pp. 1728–1733, IEEE, 2010.
[18] P. C. Tay, S. T. Acton, and J. A. Hossack, “A stochastic approach to ultrasound

despeckling,” in Biomedical Imaging: Nano to Macro, 2006. 3rd IEEE International
Symposium on, pp. 221–224, IEEE, 2006.
[19] D. Tamilkudimagal and K. Kalpana, “Squeeze box filter for contrast enhancement in
ultrasound despeckling,” in Emerging Trends in Electrical and Computer Technology
(ICETECT), 2011 International Conference on, pp. 524–530, IEEE, 2011.
[20] Y. Yu and S. T. Acton, “Speckle reducing anisotropic diffusion,” Image Processing,

IEEE Transactions on, vol. 11, no. 11, pp. 1260–1270, 2002.
[21] W. Gomez, L. Leija, W. Pereira, and A. Infantosi, “Morphological operators on the

segmentation of breast ultrasound images,” in Health Care Exchanges, 2009. PAHCE
2009. Pan American, pp. 67–71, IEEE, 2009.
48
[22] L. Zhang, Y. Ren, C. Huang, and F. Liu, “A novel automatic tumor detection
for breast cancer ultrasound images,” in Fuzzy Systems and Knowledge Discovery
(FSKD), 2011 Eighth International Conference on, vol. 1, pp. 401–404, IEEE, 2011.
[23] W. M. Hafizah, E. Supriyanto, and J. Yunus, “Feature extraction of kidney ultra-

sound images based on intensity histogram and gray level co-occurrence matrix,” in
Modelling Symposium (AMS), 2012 Sixth Asia, pp. 115–120, IEEE, 2012.
[24] K. B. Raja, M. R. Reddy, S. Swaranamani, S. Suresh, M. Madheswaran, and K. Thya-

garajah, “Study on ultrasound kidney images for implementing content based image
retrieval system using regional gray-level distribution,” in Proc. International Con-
ference on advances in infrastructures for electronic business, education, science,
medicine, and mobile technologies on the internet, pp. 6–12, 2003.
[25] B. S. Usha and S. Sandya, “Measurement of ovarian size and shape parameters,” in
Proc. Annual IEEE India Conference (INDICON), pp. 1–6, 2013.
[26] K. Riha and R. Benes, “Circle detection in pulsative medical video sequence,” in
Signal Processing (ICSP), 2010 IEEE 10th International Conference on, pp. 674–
677, IEEE, 2010.
[27] K. Řı́ha, J. Mašek, R. Burget, R. Beneš, and E. Závodná, “Novel method for lo-
calization of common carotid artery transverse section in ultrasound images using
modified viola-jones detector,” Ultrasound in medicine & biology, vol. 39, no. 10,
pp. 1887–1902, 2013.
[28] E. Cox, “A method of assigning numerical and percentage values to the degree of
roundness of sand grains,” Journal of Paleontology, vol. 1, no. 3, pp. 179–183, 1927.
[29] E. Alpaydin, Introduction to machine learning. MIT press, 2004.
[30] K. Bommanna Raja, M. Madheswaran, and K. Thyagarajah, “Analysis of ultrasound

kidney images using content descriptive multiple features for disorder identification
and ann based classification,” in Computing: Theory and Applications, 2007. IC-
CTA’07. International Conference on, pp. 382–388, IEEE, 2007.
[31] N. Aggarwal and R. Agrawal, “First and second order statistics features for clas-
sification of magnetic resonance brain images.,” Journal of Signal & Information
Processing, vol. 3, no. 2, 2012.
49
[32] J. R. Carr and F. P. De Miranda, “The semivariogram in comparison to the co-
occurrence matrix for classification of image texture,” Geoscience and Remote Sens-
ing, IEEE Transactions on, vol. 36, no. 6, pp. 1945–1952, 1998.
[33] W. Gómez, W. Pereira, and A. F. C. Infantosi, “Analysis of co-occurrence texture

statistics as a function of gray-level quantization for classifying breast ultrasound,”
Medical Imaging, IEEE Transactions on, vol. 31, no. 10, pp. 1889–1899, 2012.
[34] N. Lu, Fractal imaging. Academic Press, 1997.
[35] D.-R. Chen, R.-F. Chang, C.-J. Chen, M.-F. Ho, S.-J. Kuo, S.-T. Chen, S.-J. Hung,
and W. K. Moon, “Classification of breast ultrasound images using fractal feature,”
Clinical imaging, vol. 29, no. 4, pp. 235–245, 2005.
[36] A. F. Costa, G. Humpire-Mamani, and A. J. M. Traina, “An efficient algorithm for

fractal analysis of textures,” in Graphics, Patterns and Images (SIBGRAPI), 2012
25th SIBGRAPI Conference on, pp. 39–46, IEEE, 2012.
[37] A. S. M. Sohail, P. Bhattacharya, S. P. Mudur, and S. Krishnamurthy, “Local relative

glrlm-based texture feature extraction for classifying ultrasound medical images,” in
Electrical and Computer Engineering (CCECE), 2011 24th Canadian Conference on,
pp. 001092–001095, IEEE, 2011.
[38] J. A. Richards and X. Jia, Remote sensing digital image analysis, vol. 3. Springer,
1999.
[39] R. Llobet, J. C. Pérez-Cortés, A. H. Toselli, and A. Juan, “Computer-aided detection

of prostate cancer,” International Journal of Medical Informatics, vol. 76, no. 7,
pp. 547–556, 2007.
[40] J. R. Tegnoor, “Automated ovarian classification in digital ultrasound images using

svm,” vol. 1, no. 6 (August-2012), 2012.
[41] Y.-L. Huang, K.-L. Wang, and D.-R. Chen, “Diagnosis of breast tumors with ultra-
sonic texture analysis using support vector machines,” Neural Computing & Appli-
cations, vol. 15, no. 2, pp. 164–169, 2006.
[42] Á. Gómez-Ferrer and S. Arlandis, “Computer-aided analysis of transrectal ultrasound

images of the prostate,” Actas Urológicas Españolas (English Edition), vol. 35, no. 7,
pp. 404–413, 2011.
50
[43] S. Sun and R. Huang, “An adaptive k-nearest neighbor algorithm,” in Fuzzy Systems
and Knowledge Discovery (FSKD), 2010 Seventh International Conference on, vol. 1,
pp. 91–94, IEEE, 2010.
[44] R.-F. Chang, W.-J. Wu, W. K. Moon, Y.-H. Chou, and D.-R. Chen, “Support vector
machines for diagnosis of breast tumors on us images,” Academic radiology, vol. 10,
no. 2, pp. 189–197, 2003.
[45] R.-F. Chang, W.-J. Wu, W. K. Moon, and D.-R. Chen, “Automatic ultrasound
segmentation and morphology based diagnosis of solid breast tumors,” Breast Cancer
Research and Treatment, vol. 89, no. 2, pp. 179–185, 2005.
[46] D. Selvathi and V. Sharnitha, “Thyroid classification and segmentation in ultrasound

images using machine learning algorithms,” in Signal Processing, Communication,
Computing and Networking Technologies (ICSCCN), 2011 International Conference
on, pp. 836–841, IEEE, 2011.
[47] S. M. Han, H. J. Lee, and J. Y. Choi, “Computer-aided prostate cancer detection

using texture features and clinical features in ultrasound image,” Journal of digital
imaging, vol. 21, no. 1, pp. 121–133, 2008.
[48] S. Mohamed, M. Salama, M. Kamel, E. El-Saadany, K. Rizkalla, and J. Chin,

“Prostate cancer multi-feature analysis using trans-rectal ultrasound images,”
Physics in Medicine and Biology, vol. 50, no. 15, p. N175, 2005.
[49] S. M. Han, H. J. Lee, and J.-Y. Choi, “Prostate cancer detection using texture and
clinical features in ultrasound image,” in Information Acquisition, 2007. ICIA’07.
International Conference on, pp. 547–552, IEEE, 2007.
[50] A. J. Smola and B. Schölkopf, “A tutorial on support vector regression,” Statistics

and computing, vol. 14, no. 3, pp. 199–222, 2004.
[51] A. D. Belegundu and T. R. Chandrupatla, Optimization concepts and applications

in engineering. Cambridge University Press, 2011.
[52] C.-W. Hsu, C.-C. Chang, C.-J. Lin, et al., “A practical guide to support vector
classification,” 2003.
[53] A. S. M. Sohail, M. M. Rahman, P. Bhattacharya, S. Krishnamurthy, and S. P.

Mudur, “Retrieval and classification of ultrasound images of ovarian cysts combining
51
texture features and histogram moments,” in Biomedical Imaging: From Nano to
Macro, 2010 IEEE International Symposium on, pp. 288–291, IEEE, 2010.
[54] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector machines,” ACM
Transactions on Intelligent Systems and Technology, vol. 2, pp. 27:1–27:27, 2011.
Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm.
[55] Q. Gu and J. Han, “Clustered support vector machines,” in Proceedings of the Six-
teenth International Conference on Artificial Intelligence and Statistics, pp. 307–315,
2013.
[56] B. Schölkopf, J. C. Platt, J. Shawe-Taylor, A. J. Smola, and R. C. Williamson,

“Estimating the support of a high-dimensional distribution,” Neural computation,
vol. 13, no. 7, pp. 1443–1471, 2001.
[57] D. M. Tax and R. P. Duin, “Support vector data description,” Machine learning,
vol. 54, no. 1, pp. 45–66, 2004.
[58] W.-C. Chang, C.-P. Lee, and C.-J. Lin, “A revisit to support vector data description
(SVDD),” Last accessed : June 10, 2014.
[59] L. Bruzzone and M. Marconcini, “Domain adaptation problems: A dasvm classifi-

cation technique and a circular validation strategy,” Pattern Analysis and Machine
Intelligence, IEEE Transactions on, vol. 32, no. 5, pp. 770–787, 2010.
[60] H. Godil, S. Davey, and R. Shekhar, “A novel algorithm for feature detection and
hiding from ultrasound images,” in Proceedings of the International Conference on
Bioinformatics, Computational Biology and Biomedical Informatics, p. 681, ACM,
2013.
[61] B. R. Sharma, N. Gupta, and N. Relhan, “Misuse of prenatal diagnostic technology

for sex-selected abortions and its consequences in india,” J. Royal Inst. Pub. Health,
vol. 121, pp. 854–860, 2007.
[62] K. C. Jena, “Female foeticide in india: A serious challenge for the society,” Orissa
review, 2008.
[63] R. Patel, “The practice of sex selective abortion in india: may you be the mother of
a hundred sons,” Carolina Paper Series. Chapel Hill: Center for Global Initiatives,
University of North Carolina, vol. 20, 1996.
52
[64] “State census 2011.” http://www.census2011.co.in/states.php. Last accessed :
May 25, 2014.
[65] “Trivers-Willard hypothesis.” http://en.wikipedia.org/wiki/Trivers-Willard_

hypothesis. Last accessed : May 25, 2014.
[66] C. Z. Guilmoto, “Sex-ratio imbalance in Asia: trends, consequences and policy

responses.” http://www.unfpa.org/gender/docs/studies/summaries/regional_
analysis.pdf. Last accessed : May 25, 2014.
[67] M. Darnovsky, “Countries with laws or policies on sex selection.” http:

//geneticsandsociety.org/downloads/200904_sex_selection_memo.pdf, April
2009. Last accessed : May 25, 2014.
[68] T. Thakur, “Skewed sex ratio behind trafficking.” http://archive.asianage.com/

india/skewed-sex-ratio-behind-trafficking-134, July 2013. Last accessed :
June 20, 2014.
[69] “Sound medicine : understanding ultrasound and its benefit.” http://www.sdms.

org/public/soundmedicine.asp. Last accessed : May 25, 2014.
[70] “Promoting ethical ultrasound use in India.” http://files.gecompany.com/gecom/

citizenship/pdfs/ge_ethical_ultrasound_use_india_casestudy.pdf. Last ac-
cessed : May 25, 2014.
[71] J. F. Burns, “India fights abortion of female fetuses.” http://www.nytimes.

com/1994/08/27/world/india-fights-abortion-of-female-fetuses.html, Au-
gust 1994. Last accessed : May 25, 2014.
[72] S. Tang and S. Chen, “A fast automatic recognition and location algorithm for fetal
genital organs in ultrasound images,” Journal of Zhejiang University SCIENCE B,
vol. 10, no. 9, pp. 648–658, 2009.
[73] I. Arel, D. C. Rose, and T. P. Karnowski, “Deep machine learning-a new frontier in
artificial intelligence research,” Computational Intelligence Magazine, IEEE, vol. 5,
no. 4, pp. 13–18, 2010.
[74] “The ultrasound scanner that plugs into a smartphone and could revolu-
tionise medical care in third world countries.” http://www.dailymail.co.uk/
53
sciencetech/article-2363964/The-ultrasound-scanner-plugs-SMARTPHONE-
revolutionise-medical-care-world-countries.html, July 15 2013.
[75] “Ultrasound is now on smartphones.” https://www.engineeringforchange.org/

news/2011/06/02/ultrasound_is_now_on_smart_phones.html, June 2 2011. Last
accessed : May 25, 2014.
[76] “Android phones sold more than iphones in April-June quarter, says
IDC.” http://businesstoday.intoday.in/story/android-phones-sold-more-
than-iphones-in-april-june-idc/1/197617.html, August 8 2013.
[77] G. Bradski, “The opencv library,” Doctor Dobbs Journal, vol. 25, no. 11, pp. 120–126,
2000.
[78] “Android NDK.” http://developer.android.com/tools/sdk/ndk/index.html.

Last accessed : May 25, 2014.
[79] “ADT Plugin.” http://developer.android.com/tools/sdk/eclipse-adt.html.

Last accessed : May 25, 2014.
[80] “Eclipse CDT.” http://www.eclipse.org/cdt/. Last accessed : May 25, 2014.
54

Sudipan Saha Mtech Ultrasound CAD

Uploaded by

Copyright:

Available Formats

Sudipan Saha Mtech Ultrasound CAD

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sudipan Saha Mtech Ultrasound CAD

Uploaded by

Copyright:

Available Formats

Ultrasound Image Processing for CAD

and Other Applications

submitted in partial fulfillment of the requirements

Department of Electrical Engineering

2 Fundamentals of Ultrasound Imaging 3

3 Ultrasound Image Preprocessing 8

4 Explicitly Programmed CAD 14

5 Machine Learning Based CAD 22

6 Fetal Ultrasound Image Processing 37

7 Conclusion and Future Work 43

A GUI to help radiogist to provide manually segmented images 46

2.1 Block diagram of ultrasound imaging device . . . . . . . . . . . . . . . . . 6

6.1 Male genitalia composed of two elliptical regions . . . . . . . . . . . . . . . 42

A.1 Screen-shot of the GUI developed to collect ground-truth data . . . . . . . 46

3.1 Estimated statistical parameters for various speckle reduction techniques

5.1 Stotal variation with k when using kNN algorithm . . . . . . . . . . . . . . 34

6.1 Gender ratio (0-6 yrs) change over time . . . . . . . . . . . . . . . . . . . . 38

ASM Angular Second Moment.

CAD Computer Aided Diagnosis.

GLCM Grey-Level Co-occurrence Matrix.

HMAF Homogeneous Mask Area Filtering.

kNN K Nearest Neighbor.

MSE Mean Square Error.

PCA Principal Component Analysis.

RBF Radial Basis Function.

SFTA Segmentation-based Fractal Texture Analysis.

1.3 Organization of the report

Fundamentals of Ultrasound Imaging

Ultrasound imaging is a popular medical imaging modality which is routinely used in

2.1 Frequency range of ultrasound wave

2.2 Converse piezoelectric effect

2.3 Different ultrasound imaging modes

2.4 Ultrasound image generation

2.5 Advantages of ultrasound imaging

1. Ultrasound is considered safe than other medical imaging techniques.

2. It is relatively inexpensive in comparison to other medical imaging techniques.

3. It is painless and easily tolerated by the patients.

4. It produces image in real time.

2.6 Drawbacks of ultrasound imaging

3. Good image cannot be obtained from obese patients.

4. More presence of noise components.

5. Images are of low resolution.

2.7 Future of ultrasound imaging

Ultrasound Image Preprocessing

2. Speckle removed image is easy to segment.

2. Speckle removal algorithm should not blur the edges.

Y (x, y) = I(x, y)W (x, y) + I 0 (x, y)(1 − W (x, y)) (3.1)

3.3 Estimation of statistical parameters

Lower RMSE indicates despeckling algorithm is better.

PSNR = 20 log10 (255/RMSE) (3.5)

Higher PSNR indicates despeckling algorithm is better.

3.3.5 Correlation coefficient

3.3.6 Speckle index

(g) Adaptive bilateral

Explicitly Programmed CAD

1. Explicitly programmed CAD (no use of machine learning)

2. Machine learning based CAD

• Carotid artery localization

• Gallbladder stone detection