Ultrasound Classification of Breast Masses

Using a Comprehensive Nakagami Imaging and

Machine Learning Framework
Ahmad Chowdhury1, Rezwana R. Razzaque1, Ahmad Shafiullah1,
Sabiq Muhtadi1, Brian S. Garra2 and S. Kaisar Alam3,4
1 Department of Electrical and Electronic Engineering, Islamic University of Technology,
Gazipur, Bangladesh
2 Division of Imaging, Diagnostics and Software Reliability, Office of Science and

Engineering Laboratories, Center for Devices and Radiological Health, Food and Drug
Administration, Silver Spring, MD, United States
3Imagine Consulting LLC, Dayton, NJ, United States
4The Center for Computational Biomedicine Imaging and Modelling (CBIM), Rutgers

University, NJ, Piscataway, United States

E-mail: [email protected]

Abstract
Ultrasound imaging is establishing itself as a popular screening technique for breast cancer due to its non-
invasive nature. In this study, we have utilized statistical parameters from the well-known Nakagami
distribution to construct several different types of Nakagami and derived Nakagami parametric images
from ultrasound B-mode scans, and investigated their potential to improve the non-invasive identification
of breast cancer. The dataset we used contains 130 biopsy-proven patients consisting of 104 benign and
26 malignant cases. For each of these patients, we generated seven types of Nakagami and derived
Nakagami images from basic and as well as derived parameters of the Nakagami distribution through a
sliding-window technique. To determine the suitable window size for image generation, we conducted an
empirical analysis using three windows of width 0.2 mm and lengths 0.1875 mm, 0.45 mm and 0.75mm.
From each patient image, we extracted a total of 72 features that consisted of morphometric, elemental
and hybrid features. Feature selection using Recursive Feature Elimination with Cross Validation (RFE-CV)
was performed to find the optimum subset of features for the best classification performance.
Incorporating the selected subset of features with the Support Vector Machine (SVM) classifier, and by
tuning the decision threshold, we obtained a maximum classification accuracy of 93.08% and an Area
Under the Curve (AUC) of 0.9712, along with a False Negative Rate of 0%, and a very low False Positive
Rate of 8.65%. Our results indicate that the high accuracy of such a procedure may assist in the diagnostic
process associated with detection of breast cancer, as well as help to reduce false positive diagnosis.

Keywords: Breast Cancer, Ultrasound, Classification, Quantitative Ultrasound, Nakagami Distribution,

Feature Selection, Recursive Feature Elimination
1. Introduction
Breast cancer is the most commonly diagnosed cancer in females around the globe by a significant
margin, with an estimated 2.3 million new cases of breast cancer being reported in 2020 (Sung et al.,
2021). It also accounted for 6.9% (approximately 700,000) of all cancer-related deaths worldwide (Sung
et al., 2021). In the United States, about 1 in 8 women (13%) are expected to develop invasive breast
cancer over the course of their lifetime, and a staggering 43,600 deaths are expected in 2021 alone
(American Cancer Society, 2021). The survival of patients is fundamentally associated with rapid and
accurate early diagnosis, as it plays a vital role in treatment selection as well as predicting the patient's
response to therapy.

Improvements in the quality of biomedical imaging have resulted in increased sensitivity to potentially
harmful tissue features, such as cancer. However, this improvement in sensitivity has not been
accompanied by improved specificity, or the ability to determine whether a cancer is benign or malignant.
As a result, an ‘overdiagnosis’ problem has occurred, which is leading to ‘overtreatment’ (Esserman et al.,
2014; Welch and Black, 2010), and current screening methods for breast cancer are associated with a high
rate of false positives. In the United States, false positive recalls and screenings associated with breast
cancer cost around $4 billion a year (Vlahiotis et al., 2018). This not only amounts to excessive anxiety for
the patient, but can also lead to unnecessary biopsy (Vlahiotis et al., 2018), which is still considered to be
the gold standard technique for pathological confirmation of malignancy and characterization of tumor
grade (Tadayyon et al., 2014). According to the National Cancer Institute, breast biopsies in the United
States have a false-positive rate of up to 71% (Mayo et al., 2019), resulting in an annual expense of $2.18
billion in biopsy procedures, which can be avoided with more accurate prior diagnosis. There is an inherent
need to reduce false positive diagnosis associated with breast cancer treatment.

X-ray Mammography is the recommended screening technique associated with breast cancer, and has
helped reduce mortality by upto 45% (Duffy et al., 2002). However, according to reports, patients with
dense breasts have an increased likelihood of receiving a False Negative result for lesion detection using
mammography (Oelze, 2012), with sensitivity for this patient group reported to be as low as 48% (Kolb et
al., 2002). Ultrasound (US) imaging techniques have been principally applied in a supporting role for the
diagnosis of breast cancer (Sainsbury, 2013). It has since become an important tool in the non-invasive
diagnosis of breast cancer (Kolb et al., 2002; Sehgal et al., 2006), and several studies have shown that
diagnostic accuracy of ultrasound approaches or exceeds that of mammography (Flobbe et al., 2003; Kolb
et al., 2002; Leconte et al., 2003).

Conventional Ultrasound (US) imaging procedures are qualitative in nature, and involve assessment of
tumor features according to the Breast Imaging Reporting and Data System (BI-RADS) criteria developed
by the American College of Radiology (ACR) (Mendelson et al., 2013). This is reported to be faster, cheaper
and more accurate than mammography (Förnvik et al., 2010; Shoma et al., 2006). However, ultrasound is
inherently operator and device-dependent (Berg et al., 2012; Brem et al., 2015), which poses limitations
on reproducibility.

On the other hand, Quantitative Ultrasound (QUS) methods are highly reproducible, and independent of
device and operator related factors (Boote et al., 1988; Nam et al., 2012; Yao et al., 1990). QUS techniques
allow for the evaluation of tissue in terms of structural and mechanical properties, and provide numerical
data related to tissue features, which have the potential to improve the specificity of biomedical imaging
techniques (Mamou and Oelze, 2013; Oelze and Mamou, 2016). The use of quantitative parameters allows
clinicians to reduce the number of biopsies, whilst maintaining the same level of detection in terms of
malignant cases (Trop et al., 2015). Several different quantitative parameters have been explored by
researchers with regards to characterization of breast tissue. This includes textural parameters (Klimonda
et al., 2019; Sadeghi-Naini et al., 2017), parameters related to entropy (Tsui, 2015; Tsui et al., 2017), and
parameters derived from the power spectrum (Alam et al., 2011; Sadeghi-Naini et al., 2017) as well as
related to the backscatter co-efficient (Nam et al., 2013; Oelze and Mamou, 2016).

The statistics of the envelope signal obtained from ultrasound scanners may also be modelled as a
probability density function (PDF) in order to quantitatively analyze the scattering properties of soft tissue.
Several distributions are utilized in this regard to model scattered signals in soft tissue (Destrempes and
Cloutier, 2010). Among them the Nakagami distribution and the Homodyned K distribution have been
frequently utilized by researchers to model scattered signals in the breast.

The homodyned K model for ultrasound echoes was initially proposed by Dutt and Greenleaf (Dutt and
Greenleaf, 1994), and later modified by Hruska (Hruska, 2009), and Hruska and Oelze (Hruska and Oelze,
2009). The homodyned K parameter along with BI-RADS classifiers was later applied by Trop et al. (Trop
et al., 2015), who demonstrated their ability to reduce the number of biopsies performed on breast cancer
patients by 25%, whilst maintaining 100% sensitivity. Byra et al. (Byra et al., 2016) utilized parameters
extracted from segmented parametric maps of homodyned K parameters in order to classify breast
tumors, and obtained a maximum AUC of 0.84.

The Nakagami distribution is a simple and generalized model proposed for ultrasonic backscatter by
Shankar (Mohana Shankar, 2000), and is one of the most important statistical distributions for modelling
soft tissue (Nakagami, 1960). The Nakagami distribution is very popular in QUS due to its reduced
computational complexity and its ability to adequately describe tissue scattering statistics. Many research
groups have attempted to utilize characteristics of the Nakagami distribution in order to analyze
ultrasound backscattered echo from the breast. Shankar et al. (Shankar et al., 2001) utilized two Nakagami
parameters, m and α, to classify breast masses from a total of 52 patients (14 malignant, 38 benign), and
reported AUC values of 0.79±0.11 and 0.828±0.10 respectively for the two parameters individually. In an
attempt to enhance classification performance, Shankar, et al.(Shankar et al., 2002) proposed the concept
of compounding the Nakagami parameter m, and observed an AUC of 0.8316 using the proposed
technique. In another study, Shankar, et al. (Shankar et al., 2003) performed a five-parameter analysis for
99 patients (29 malignant, 70 benign) based on parameters derived from both Nakagami and K
distribution at the site, boundary, spiculated region and shadow of the breast mass, and a combination of
these features lead to an AUC of 0.96 ±0.02. Tsui et al. (Tsui et al., 2009) found that traditional Nakagami
parameters extracted with an unfocused transducer cannot be utilized to distinguish between different
scatterer qualities. They proposed a novel approach for tissue characterization using a non-focused
transducer termed noise-assisted Nakagami parameter based on empirical mode decomposition of noisy
backscattered echoes, and proved its feasibility for ultrasound tissue characterization. In a later study,
Tsui et al. (Tsui et al., 2010) used backscattered signals from breast scans to form Nakagami parametric
images in order to classify breast lesions for 100 patients (50 malignant, 50 benign) and obtained an AUC
of 0.81 ± 0.04 along with an accuracy of 82%. Liao et al. (Liao et al., 2011) performed 2-D analysis by
integrating Nakagami parametric images and four texture feature parameters from B-scan images, and
obtained an accuracy of 88.2% for 130 breast cancer patients. Liao et al. (Liao et al., 2012) also introduced
a novel method to classify breast lesions by incorporating elasticity properties obtained from the strain-
compounding imaging method, based on the acquisition of multiple frames under various strain
conditions with the scatterer characteristics of Nakagami images. They conducted their study on a total
of 50 patients and reported an AUC of 0.92. Dobruch-Sobczak, et al. (Dobruch-Sobczak et al., 2017)
combined the BI-RADS assessment with Nakagami statistics and found it to be the most appropriate
method to distinguish between benign and malignant breast lesions of BI-RADS category 4a. According to
their study, in comparison to BI-RADS alone, using B-mode and the mLmin parameter from Nakagami
image together gave an AUC of 0.978. Byra, et al. (Byra et al., 2017) utilized Nakagami parametric images
from 458 RF data matrices in order to train a convolutional neural network with 5-fold cross-validation to
classify breast lesions, and obtained an ROC area of 0.912. Hsu, et al. (Hsu et al., 2019) utilized
morphological and texture features from B-scan images and Nakagami parametric images to characterize
breast lesions, and reported an AUC of 0.96 and accuracy of 89.4%. Although quantitative parametric
studies are conducted in the area within the lesion in most cases, Klimonda, et al. (Klimonda et al., 2018)
proved that the two parameters, weighted entropy and the Nakagami shape parameter (m), calculated
from the surrounding tissue of the lesion better classifies the lesion than values from within the lesion
(AUC values were on average 18% higher for the surrounding tissue than within the tissue). In another
study, Klimonda, et al. (Klimonda et al., 2019) proposed a multi-parametric approach by combining
Nakagami shape parameter, entropy and texture parameters. They studied these parameters for both
inside the lesion and the tissue surrounding the lesion, and obtained AUC of 0.92 for the surrounding
tissue and 0.94 from a combination of both of them.

The paper proposes a novel framework for classification of breast masses using numerous types of
Nakagami parametric images. Seven different parametric images are generated based on Nakagami alpha,
omega and mu parameters from ultrasound radio-frequency (RF) data. Different types of morphometric,
elemental and hybrid features are extracted from each of the parametric images, some of which are
excellent for classification when grouped together. To our knowledge, no other research work has been
conducted in such a comprehensive manner on Nakagami parameters and their features. While our
research proposes a large number of features, it is also quite an impractical approach to incorporate all
the features for classification. In a clinical scenario, generation of all the features for each patient would
be quite cumbersome, and incorporation of a high number of features would increase the computation
time of the classification algorithm. Thus, our framework further broadens in prospect by adopting a
technique for feature selection which provides us with an optimum feature subset for the best
classification performance. Incorporating the selected subset of features with the Support Vector
Machine (SVM) classifier, and by tuning the decision threshold, we obtained a False Negative Rate of 0%,
along with a False Positive Rate as low as 8.65%, and a maximum classification accuracy of 93.08%.
2. Methodology
This study analyzes features derived from Nakagami parametric images. The process begins with the
generation of envelope images from patient ultrasound RF data. 7 different Nakagami parametric images
are then generated from each envelope image using 3 different window sizes. A total of 72 features are
extracted from these Nakagami parametric images. Feature selection is performed using the RFE-CV
algorithm in order to select the subset of features that would result in the best classification, and to
eliminate redundant features. The selected subset of features is then utilized for classification purposes.
Figure 1 depicts the proposed framework.

Figure 1: Proposed Framework

2.1 Description of the Dataset
The dataset utilized for this study is from ATL’s PMA study (IRB approved) undertaken in 1994 (Alam et
al., 2011). The dataset consists of RF data comprises biopsy proven analysis results of 130 patients, of
which 104 are benign and 26 are malignant. The data were acquired at three clinical sites (Thomas
Jefferson University, University of Cincinnati, and Yale University), at the time of routine ultrasonic
examinations of patients scheduled for biopsy. Informed consent was obtained from each patient
participating in the study. Participating patients had mammographically-visible lesions, that were either
palpable or non-palpable, and which were discovered through screening, physical examination or both
methods. Exclusion factors for patients consisted of: age (less than 18 years, due to legal consent
limitations), prior breast carcinoma, biopsy or mastectomy, breast implant, simple cyst, pregnancy,
microcalcifications not associated with a mass on sonography, and males or transsexuals. Before
performing biopsy, the masses were examined in a supine position by an experienced radiologist or
sonographer, using a Philips (ATL at the time) Ultrasound (Bothell, WA) UM-9 HDI scanner, with an L10-5
(7.5-MHz nominal frequency) linear-array transducer, during which standard ultrasonic breast
examination procedures were employed. Multiple views were selected for every lesion by the radiologist,
which included at least a radial and an anti-radial view, and in some cases additional views. Some patients
had multiple masses and each mass was assigned a number, and the radiologist marked the images with
the mass number. The file name convention also included the mass number.

In order to define the vascularity of each mass compared to that of the mirror image location in the
opposite breast, color-flow and pulse-Doppler examination were performed, although the study
described in this paper excludes the Doppler results, as these data were not saved. The mammographer
scored each mammogram using a level of suspicion (LOS) on a scale of 1 to 5, where 1 denotes benign, 2
probably benign, 3 indeterminate, 4 probably malignant and 5 definitely malignant.

The RF echo-signal data was digitally obtained by interfacing a Spectrasonics Inc. (King of Prussia, PA)
acquisition module with the scanner to allow acquisition of RF data. (The UM-9 HDI scanner didn’t have
an RF data output.) The L10-5 transducer was used at a default power level and a single transmit focal
length, as selected by the operator. The RF data was sampled at 20 MHz, with an effective dynamic range
of 14 bits. The acquired data was corrected for time-gain-control (TGC) before processing; TGC data was
acquired before every scan.
2.2 Nakagami Parametric Imaging
2.2.1 The Nakagami Distribution
It is possible to model the backscattered echo of acoustic pulse as the algebraic sum of the contributions
of individual scatterers with the tissue (Shankar, 1995; Wagner et al., 1983). Consider a total of M
scatterers in the target cell, and let 𝑋𝑋𝑚𝑚 and 𝜑𝜑𝑚𝑚 represent the amplitude and phase of the 𝑚𝑚𝑡𝑡ℎ scatterer
respectively. Then, the backscattered echo may be represented as

𝑀𝑀 (1)
𝐵𝐵 (𝑡𝑡) = � 𝑋𝑋𝑚𝑚 𝑐𝑐𝑐𝑐𝑐𝑐 (𝜔𝜔𝜔𝜔 − 𝜑𝜑𝑚𝑚 )
𝑚𝑚=1

The backscattered echo can also be represented in terms of the in-phase and quadrature components.

𝐵𝐵(𝑡𝑡) = 𝑃𝑃 𝑐𝑐𝑐𝑐𝑐𝑐 𝑐𝑐𝑐𝑐𝑐𝑐 (𝜔𝜔𝜔𝜔) + 𝑄𝑄 𝑠𝑠𝑠𝑠𝑠𝑠 𝑠𝑠𝑠𝑠𝑠𝑠 (𝜔𝜔𝜔𝜔) (2)

Where,

𝑀𝑀 (3)
𝑃𝑃 = � 𝑋𝑋𝑚𝑚 𝑐𝑐𝑐𝑐𝑐𝑐 (𝜑𝜑𝑚𝑚 )
𝑚𝑚=1

𝑀𝑀 (4)
𝑄𝑄 = � 𝑋𝑋𝑚𝑚 𝑠𝑠𝑠𝑠𝑠𝑠 (𝜑𝜑𝑚𝑚 )
𝑚𝑚=1

The envelope of the backscattered echo may be represented as

𝐸𝐸 = �𝑃𝑃2 + 𝑄𝑄2 (5)

𝑃𝑃 and 𝑄𝑄 may be considered Gaussian distributed with zero mean and equal variance if the target cell
contains a large number of randomly located scatterers. Under this condition, the envelope 𝐸𝐸 will obey
Raleigh statistics. However, this will not be the case if the target cell contains scatterers with randomly
varying scattering cross-sections and a comparatively high degree of variance (Molthen et al., 1995;
Narayanan et al., 1994; Shankar, 1995; Shankar et al., 2001). The envelope can also be considered Rician
or Post Rayleigh if in addition to randomly located scatterers, the target cell contains periodically located
scatterers spaced corresponding to integral multiples of wavelength at demodulation frequency (Wagner
et al., 1987). All these scattering conditions exist in the Nakagami distribution (Nakagami, 1960).

The Nakagami distribution (Nakagami, 1960) is one of the simplest models for modelling backscatter
envelope. It is a two-parameter distribution first introduced by Nakagami (1943, 1960) in the context of
wave propagation. The Nakagami probability density function is given by,

2𝑚𝑚𝑚𝑚 2𝑚𝑚−1 −𝑚𝑚𝑚𝑚2 (6)

𝑁𝑁 (𝑚𝑚, 𝛺𝛺) = 𝐴𝐴 𝑒𝑒 𝛺𝛺
𝛤𝛤𝛤𝛤𝛺𝛺𝑚𝑚

Here, 𝛤𝛤 represents the Euler gamma function.

The cumulative distribution of the Nakagami distributed envelope 𝐹𝐹(𝑟𝑟) is given by,

𝑟𝑟
2𝑚𝑚𝑚𝑚 2𝑚𝑚−1 −𝑚𝑚𝑚𝑚2 (7)
𝐹𝐹 (𝑟𝑟) = � 𝐴𝐴 𝑒𝑒 𝛺𝛺 𝑑𝑑𝑑𝑑
0 𝛤𝛤𝛤𝛤𝛺𝛺𝑚𝑚

The following expressions are for the mean intensity of the Nakagami distribution and its SNR:

𝐸𝐸 [𝐼𝐼 ] = 𝛺𝛺 (8)

𝑆𝑆𝑆𝑆𝑅𝑅2 = 𝑚𝑚 (9)

This intensity SNR parameter should not be confused with the amplitude SNR.

The Nakagami distribution has two parameters, which are expressed as

[𝐸𝐸 (𝑅𝑅2 )]2 (10)

𝑚𝑚 =
𝐸𝐸[𝑅𝑅2 − 𝐸𝐸 (𝑅𝑅2 )]2

𝛺𝛺 = 𝐸𝐸[𝑅𝑅2 ] (11)
The parameter 𝑚𝑚 is referred to as the shape parameter, and contains information about the envelope
statistics. In the case of the Nakagami distribution, it is constrained such that 𝑚𝑚 ≥ 0.5 (Nakagami, 1960),
in which case it is referred to as the Nakagami parameter. The case of 𝑚𝑚 = 1 corresponds to Rayleigh
scattering (Shankar et al., 2001), while 𝑚𝑚 in the range of 0.5 to 1 corresponds to pre-Rayleigh scattering
(Mohana Shankar, 2000; Nakagami, 1960; Shankar, 1995). A case of 𝑚𝑚 > 1 corresponds to post-Rayleigh
or Rician (Mohana Shankar, 2000). Hence, the Nakagami distribution incorporates all these scattering
conditions(Shankar et al., 2001).

The parameter 𝛺𝛺 is a scaling parameter.

A similarity also exists between the Nakagami parameter 𝑚𝑚 and the effective number 𝑀𝑀 of the K
distribution (Narayanan et al., 1994; Shankar, 1995). The K distribution was initially proposed to account
for the high degree of variability in scattering cross sections of the scatterers in the target cell (Dutt and
Greenleaf, 1994; Shankar, 1995).

The cumulative distribution of the K distributed envelope 𝐹𝐹𝐾𝐾 (𝑟𝑟) can be expressed as

2𝑏𝑏 𝑏𝑏𝑏𝑏 𝑀𝑀 (12)

𝐹𝐹𝐾𝐾 (𝑟𝑟) = � � 𝐾𝐾𝑀𝑀−1 (𝑏𝑏𝑏𝑏) 𝑟𝑟 ≥ 0 𝑀𝑀 ≥ 0
𝛤𝛤(𝑀𝑀) 2

Here, the parameter 𝑏𝑏 is a scaling factor, and 𝑀𝑀 represents the effective number of scatterers in the target
cell.

There is evidence that the K distribution can be expressed as the product of two Nakagami-distributed
random variables (Nakagami, 1960). Hence, the parameters of the K distribution, namely 𝑀𝑀 and 𝑏𝑏 may be
expressed in terms of the parameters of the Nakagami distribution (Shankar et al., 2001) as follows

2𝑚𝑚 (13)
𝑀𝑀 =
1 − 𝑚𝑚

And
 (14)
2𝑚𝑚
𝑏𝑏 = 2�
𝛺𝛺(1 − 𝑚𝑚)

Hence, the Nakagami parameter 𝑚𝑚 is similar to the K distribution parameter 𝑀𝑀, and may provide an
indication of the effective number of scatterers in the target cell with a compressed dynamic range
(Shankar et al., 2001).

1
We may define a new parameter, 𝛼𝛼, where 𝛼𝛼 = and is a measure of effective cross section of scatterers
𝑏𝑏

in the tissue under examination. Thus,

(15)
1 𝛺𝛺(1 − 𝑚𝑚)
𝛼𝛼 = �
2 2𝑚𝑚

It is defined as the effective cross section because of the dependance of 𝛼𝛼 on the number density of
scatterers through 𝑚𝑚, the signal-to-noise ratio of the cross sections through 𝑚𝑚, and the attenuation
through both 𝛺𝛺 or 𝛼𝛼 (Narayanan et al., 1994; Shankar, 1995; Shankar et al., 1996). Furthermore, 𝛼𝛼 can
provide information about the scattering characteristics within the target cell through 𝛺𝛺. Hence, the
Nakagami distribution is expected to be effective in tissue characterization, as the parameters 𝑚𝑚 and 𝛼𝛼
provide information about the number density of scatterers, the level of attenuation present and the
degree of homogeneity in the scattering cross-sections (Shankar et al., 2001).

2.2.2 Generation of Nakagami Parametric Images

Tsui and Chang (Tsui and Chang, 2007) found that Nakagami images constructed with an optimal window
size have the ability to impart both the global and local backscattered statistics of the ultrasonic signals in
a tissue, leading it to be a very efficient local scatterer concentration detector. They found that the optimal
window for generating Nakagami images is a square with side length equal to three times the pulse length
of the incident ultrasound beam, and concluded that Nakagami images can effectively assist B-mode
image in medical diagnosis (Tsui and Chang, 2007).

Usually, a smaller window size is chosen for the parametric imaging to achieve a good spatial resolution
of the image, but in that case most estimators become unstable and the tissue property map becomes
inaccurate (Larrue and Noble, 2011). Again, larger window size is chosen to obtain a smooth and accurate
map of tissue properties, decreasing spatial resolution of the map i.e., making smaller structures and
lesion boundaries undetectable (Larrue and Noble, 2011). Especially, the Nakagami parameter 𝑚𝑚 is
sensitive to such problems of estimation, window size and boundary effects (Larrue and Noble, 2011). To
solve these problems, a trade-off between the resolution and the stability of the estimator i.e., the
smoothness of the parametric image has to be maintained (Larrue and Noble, 2011). Though our study
focuses on automated quantitative analysis (resolution of the image is inconsequential), we tried to
maintain the trade-off between the aforementioned quantities and choose an optimum window size
through empirical analysis.

A Nakagami Image is the local map of Nakagami parameters generated from the envelope image. The
following steps were followed to generate Nakagami images:

1) Local backscattered envelopes are collected using a window within the envelope image. The local
backscattered envelope values are used to estimate local Nakagami parameters, 𝑚𝑚 and 𝛺𝛺, which
are allocated as new pixels in the corresponding m and omega matrices.

2) Step 1 is repeated with the window moving in 0.0385mm increments through the entire envelope
image, yielding the Nakagami images as maps of local 𝑚𝑚 and 𝛺𝛺 values.

Derived Nakagami images, namely Pre-alpha, Real alpha, Imaginary alpha, Phase alpha and Absolute alpha
are generated from the 𝑚𝑚 and 𝛺𝛺 images by estimating the derived Nakagami values corresponding to
each 𝑚𝑚 and 𝛺𝛺 value.

In order to conduct empirical analysis, three different datasets are generated using three different
window sizes: 0.1875 𝑚𝑚𝑚𝑚, 0.45 𝑚𝑚𝑚𝑚 and 0.75𝑚𝑚𝑚𝑚 . For these 3 different window sizes an overlap of 80%
for 0.1875 𝑚𝑚𝑚𝑚, 91.67% for 0.45 𝑚𝑚𝑚𝑚 and 95% for 0.75𝑚𝑚𝑚𝑚 was maintained in step 2. Empirical analysis
was conducted to determine the best size of the window for classification by observing accuracy and area
under curve (AUC) for the varying window sizes.

2.2.3 Description of Nakagami Images

The parameters that have been used in this study for analysis were derived from two different types of
Nakagami parametric images:
1) Nakagami Basic Images
2) Nakagami Derived Images

Nakagami Basic Images:

1. 𝑚𝑚 Image: map of the Nakagami parameter 𝑚𝑚 which provides information about envelope statistics as
well as effective number of scatterers (Shankar et al., 2001). Its value is constricted such that,

𝑚𝑚 ≥ 0.5

2. 𝛺𝛺 Image: map of the parameter 𝛺𝛺 which is a Scaling parameter.

Nakagami Derived Images:

The derived Nakagami images were formed using the parameter 𝛼𝛼, which is a measure of effective cross-
section of scatterers. It is given by:

(15)
1 𝛺𝛺(1 − 𝑚𝑚)
𝛼𝛼 = �
2 2𝑚𝑚

It is a complex value. 5 types of images are formed using the α parameter:

1. Pre-𝛼𝛼 Image: map of the parameter pre-alpha which we define as follows:

𝛺𝛺(1 − 𝑚𝑚) (16)

𝑃𝑃𝑃𝑃𝑃𝑃 − 𝑎𝑎𝑎𝑎𝑎𝑎ℎ𝑎𝑎 =
2𝑚𝑚

2. 𝛼𝛼-absolute Image: map of the absolute value of the 𝛼𝛼 parameter.

3. 𝛼𝛼-phase Image: map of the phase value of the 𝛼𝛼 parameter.
4. 𝛼𝛼-real Image: map of the real part of the 𝛼𝛼 parameter.
5. 𝛼𝛼-imaginary Image: map of the imaginary part of the 𝛼𝛼 parameter.
 Figure 2: Overview of the 7 types of Nakagami Parametric Images

2.3 Feature Extraction

All feature processing software have been built in MATLABTM (The MathWorks, Inc., Natick, MA). 3 types
of features have been extracted from each Nakagami parametric image:

1) Morphometric Features
2) Elemental Features
3) Hybrid Features

2.3.1 Morphometric Features

These features describe the shape or boundary of the lesion. They have been calculated based on the
lesion boundaries traced on the Nakagami images. These features are described below:

a) Aspect Ratio: Ratio of maximum vertical lesion dimension to maximum horizontal lesion
dimension.

𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉𝑉 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 (17)

𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 =
𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻 𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷
b) Compactness: Ratio of square root of lesion area and its maximum diameter. It represents the
compactness of the shape of the traced boundary.

√𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 (18)

𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 =
𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷

c) Roundness: Ratio of the lesion area and its maximum diameter squared. It represents the
roundness of the shape of the traced boundary.

𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 (19)

𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 =
(𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷)2

d) Convexity: Ratio between convex perimeter and actual lesion perimeter which quantifies the
border property. This parameter is very sensitive to lesion spiculation and thus it is adept at
quantifying border property.

𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 (20)

𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 =
𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖

e) Form Factor: Ratio of the lesion area to the square of lesion perimeter.

𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 (21)

𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹 =
(𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃)2

f) Solidity: Ratio between lesion area and convex area. It is derived as,

𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿𝐿 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 (22)

𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 =
𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴

g) Fractal Dimension for morphometric feature analysis: A fractal dimension is a ratio providing a
statistical index of complexity comparing how details in a pattern (strictly speaking, a fractal
pattern) change with the scale at which it is measured. It has also been characterized as a measure
of the space-filling capacity of a pattern that describes how a fractal scales differently from the
space it is embedded in. The fractal dimension of a closed contour is also an indicator of boundary
 irregularity or roughness (Alam et al., 2011). Fractal dimension has correlation with the quantity
roughness, and as spiculation increases fractal dimension increases. We calculated fractal
dimensions in three different ways for analyzing the boundary roughness:

● Kolmogorov Fractal Dimension: Computes the slope of the line that plots the number of grid
squares through which the lesion boundary passes versus the grid size on log-log axes.

● Minkowski Fractal Dimension: Computes slope of the line that plots the area swept out by
circles versus their diameter on log-log axes.

● Hausdorff Fractal Dimension: The Hausdorff dimension is an integer for sets of points that
define a smooth shape or a shape that has a small number of corners, and for irregular shapes
it becomes non-integer. Here, we used the perimeter to calculate this fractal dimension.

2.3.2 Elemental Features

Elemental features are related to the quantitative measures of the parametric images. One of the
elemental features, Co-occurrence Contrast, has been extracted from the entire parametric image. This is
defined as follows:

• Co-occurrence Contrast: Contrast for the co-occurrence matrix of an image to estimate texture of
the image.

The remaining features have been extracted from different regions-of-interest (ROI) of Nakagami
parametric images. Nine different ROIs were considered for parameter extraction in this study: (1) left-
anterior, (2) left-lateral, (3) left-posterior, (4) tumor-anterior, (5) tumor, (6) tumor-posterior, (7) right-
anterior, (8) right-lateral and (9) right-posterior. These are illustrated in Figure 3.
 Figure 3: Nakagami mu image with analysis region traces superimposed: (1): left-anterior, (2): left-
lateral, (3): left-posterior, (4): tumor-anterior, (5): tumor, (6): tumor-posterior, (7): right-anterior, (8):
right-lateral, and (9): right-posterior

These features are described below:

a) Echogenicity: Mean of the pixel intensity values within the lesion, and it is represented by 𝜇𝜇𝐿𝐿 . For
this feature no attenuation correction is necessary.
b) Heterogeneity: Standard deviation of pixel intensity values within the lesion and it is represented
by 𝜎𝜎𝐿𝐿 .
c) FNPA: Four-neighborhood pixel algorithm, that computes 4-neighborhood-pixels texture of any
parameter within the lesion area (Alam et al., 2011). FNPA for an image of size 𝑚𝑚 × 𝑛𝑛 with pixel
values 𝑥𝑥(𝑘𝑘, 𝑙𝑙) is given by,

𝐹𝐹𝐹𝐹2 = 𝐹𝐹𝐹𝐹1 /𝜇𝜇 (23)

where,

𝑛𝑛 𝑚𝑚
1 (24)
𝐹𝐹𝐹𝐹1 = � � [|𝑥𝑥(𝑘𝑘, 𝑙𝑙) − 𝑥𝑥(𝑘𝑘 − 1, 𝑙𝑙)| + |𝑥𝑥(𝑘𝑘, 𝑙𝑙) − 𝑥𝑥(𝑘𝑘 + 1, 𝑙𝑙)| + |𝑥𝑥(𝑘𝑘, 𝑙𝑙) − 𝑥𝑥(𝑘𝑘 + 1, 𝑙𝑙)|]
4
𝑙𝑙=1 𝑘𝑘=1
And, 𝜇𝜇 is the mean value of 𝐹𝐹𝐹𝐹1 .

d) Hurst Coefficient Fractal Dimension: The texture of the lesion in an image can be defined using
this feature, that is, a Hurst coefficient is a fractal dimension coefficient that characterizes the
surface roughness within the lesion. It uses 7 × 7 sub-images for the computation.
e) Shadow-Normal: Difference between mean Nakagami parameter values in comparable shadowed
and non-shadowed regions posterior to the lesion (normalized by lesion thickness). The average
of the variations between left-lateral and left-posterior, and right-lateral and right-posterior is
compared to the difference between tumor and tumor-posterior. It is defined as follows:

1 (25)
𝑆𝑆ℎ𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = 𝑀𝑀𝑝𝑝𝑝𝑝 − (𝑀𝑀𝑝𝑝𝑝𝑝𝑝𝑝 + 𝑀𝑀𝑝𝑝𝑝𝑝𝑝𝑝 )
2

where, 𝑀𝑀𝑝𝑝𝑝𝑝 is the mean Nakagami parameter posterior to lesion, 𝑀𝑀𝑝𝑝𝑝𝑝𝑝𝑝 and 𝑀𝑀𝑝𝑝𝑝𝑝𝑝𝑝 are the mean
Nakagami parameters in normal tissue right and left lateral posterior to lesion respectively (Alam
et al., 2011).
f) Relative-Absorption: Compounded feature and that is given by:

1 1 (26)
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 = �𝑀𝑀𝑝𝑝𝑝𝑝 − 𝑀𝑀𝑎𝑎𝑎𝑎 � − (𝑀𝑀 − 𝑀𝑀𝑎𝑎𝑎𝑎 )
𝑑𝑑1 𝑑𝑑2 𝑝𝑝𝑝𝑝

where, 𝑀𝑀𝑎𝑎𝑎𝑎 is the mean Nakagami parameter inside the lesion, 𝑀𝑀𝑝𝑝𝑝𝑝 is the mean Nakagami
parameter posterior to the lesion, 𝑀𝑀𝑎𝑎𝑎𝑎 is the mean Nakagami parameter of the normal tissue next
to the lesion, 𝑀𝑀𝑝𝑝𝑝𝑝 is the mean Nakagami parameter in normal tissue lateral posterior to the lesion,
𝑑𝑑2 is the distance between 𝑀𝑀𝑝𝑝𝑙𝑙 and 𝑀𝑀𝑎𝑎𝑎𝑎 centroids, and 𝑑𝑑1 is the distance between 𝑀𝑀𝑝𝑝𝑝𝑝 and 𝑀𝑀𝑎𝑎𝑎𝑎
centroids (Alam et al., 2011).
 Figure 4: Relative absorption

2.3.3 Hybrid Features

These features are a combination of both morphometric and elemental features. They are calculated using
both the lesion contour and elemental features. To eliminate dependence on contour length and
Nakagami parameter values, normalization is necessary. These features are described below:

a) Margin definition Area: Ratio of the gradient around the lesion boundary of one pixel width to the
area of the boundary region.

b) Margin definition Gradient: Sum of the magnitudes of the lesion contour's gradients normalized
by the sum of the magnitudes of the lesion contour's gradients.

For each patient 7 Nakagami parametric images have been generated, from which 7 Elemental and 2
Hybrid features can be obtained for each image. Alongside these features, 9 additional morphometric
features are generated for each patient. Thus, in total, 72 features were extracted from each patient scan.
 Figure 5: Breakdown of all the features used in this study

2.4 RFE-CV Feature Reduction

A dataset with a large number of features is not suitable for machine learning classifiers. Feature Selection
(FS) provides a solution to this problem. It encompasses selecting a subset of relevant features from the
input feature set that are the most significant in terms of the target concept. The selected subset of
features has the ability to improve learning accuracy and reduce training time (Langley, 1995, 1994; Zhao
et al., 2010) by selecting only those features which have the most predictive power and removing
irrelevant and redundant features (John et al., 1994). Furthermore, reducing the dimensionality of data
through feature selection improves the generalization of the developed model by minimizing the risk of
overfitting, and hence results in better predictability (Gutlein et al., 2009; Piramuthu, 2004).
In this study, we have utilized Recursive Feature Elimination with Cross Validation (RFE-CV), a wrapper-
type feature selection algorithm, to select the subset of most relevant features for our dataset. It is based
on the Recursive Feature Elimination (RFE) algorithm, which is very suitable for feature selection of
datasets with a small sample size (Guyon et al., 2002). RFE works by eliminating the least important
features whose exclusion would have the least effect on the chosen performance metric, for our case it
was accuracy. It starts by searching for a subset of features taking the whole training dataset as input and
discards less important features until a desired number of features remain. To accomplish this, RFE uses
different machine learning models like Decision Tree, Random Forest and Logistic Regression at its core,
to compute the redundancy of a feature and rank features by importance. With each iteration the model
is refitted with fewer and fewer features and performance is evaluated in terms of the accuracy. For
accuracy prediction, the RFE can fit a classifier model predefined by the user to the selected subset of
features. In our study, the algorithm, Multilayer Perceptron (MLP) led us towards obtaining the best
combination of the features for classification with Linear SVM.

The main advantage of RFE is flexibility by proper tuning of its hyperparameters, which can be tuned to
achieve desired model performance. The main drawback of RFE is that the minimum number of features
required for the optimum result has to be predefined as an input. But it is impossible to determine this
minimum number and an input has to be given based on speculation which may not actually lead towards
the optimum result. To avoid this problem, RFE-CV uses cross validation to determine the minimum
number of features for the most optimum result, and no feature number has to be specified as input. RFE-
CV scores different subsets of features based on accuracy. For every subset of different sizes, it keeps
track of the mean values of the scores obtained from all the test sets while performing cross validation.
From these subsets of different sizes, the one with the best mean score is chosen as the best combination
of minimum number of features for the optimum result. So, RFE-CV is more accurate than RFE as it makes
the feature selection process automated, and leaves no subset untested by trying out every possible
feature combination.
 Figure 6: RFE-CV Feature Selection Algorithm

3. Results
Three sets of data were prepared by extracting features from parametric images formed using three
windows of different size: 0.1875mm, 0.45mm and 0.75mm. Feature selection using RFE-CV generated a
subset of the same five features for data generated using all three window sizes. These five features are,
Aspect Ratio (morphometric), Convexity (morphometric), Mu Co-occurrence contrast (Elemental), Mu
Hurst Coefficient (Elemental) and Phase Echogenicity (Elemental), shown in Table 1. These five features
from all three sets of data were used for classification, using Linear SVM classifier with 5-fold cross-
validation (from MATLAB Classification Learner Application). An empirical analysis among the three
different window sizes (Table 2) clearly delineated an improvement in performance metrics: AUC and
Accuracy, with increasing window size, and a conclusion could be drawn that the dataset of five features
for images generated using a window size 0.75mm yields the best Area Under the ROC (AUC) value of
0.9712, which is shown in Figure 7.
 Selected Features

Aspect Ratio

Convexity

Mu Co-occurrence contrast

Mu Hurst Coefficient

Phase Echogenicity

Table 1. Selected Features for Window Size of 0.75mm

Window Size AUC Accuracy

0.1875mm 0.87 86.9%

0.45mm 0.93 87.7%

0.75mm 0.97 89.2%

Table 2: Empirical Analysis of Three Window Sizes
 Figure 7: Area under the ROC Curve for the Dataset of Window Size of 0.75mm

After the selection of the best window size i.e., 0.75mm, the individual performance of the 5 selected
features derived from parametric images generated using this window were analyzed. The mean and
standard deviation values of these 5 features are provided in Table 3. The distribution of feature values
for benign and malignant cases for each of the 5 features are depicted in Figure 8.

Feature Benign Cases Malignant Cases

Aspect Ratio 0.634 ± 0.207 0.887 ± 0.222
Convexity 0.993 ± 0.011 0.961 ± 0.029
Phase Echogenicity 0.909 ± 0.123 1.005 ± 0.086
Mu Co-occurrence Contrast 3.232 ± 2.843 3.279 ± 2.197
Mu Hurst Coefficient 0.5626 ± 0.0470 0.538 ± 0.030
Table 3: Mean and standard deviation values of the 5 selected features for a window of 0.75mm
 (a) (b)

(c) (d)

(e)
Figure 8: Boxplot of the 5 selected features for 0.75mm window size: (a) Aspect Ratio. (b) Convexity. (c)
Mu Co-occurrence Contrast. (d) Mu Hurst Coefficient Fractal Dimension. (e) Phase Echogenicity.
Correlation coefficients between the features were also calculated and all the values were found to be
below 0.5. Table 4 depicts correlation among the features.

Correlation Aspect Convexity Phase Mu Co- Mu Hurst

Coefficients Ratio Echogenicity occurrence Coefficient
Contrast

Aspect Ratio 1.00 -0.312 0.263 0.062 -0.287

Convexity -0.312 1.00 -0.102 -0.013 0.123

Phase 0.263 -0.102 1.00 0.496 0.068

Echogenicity

Mu Co-
occurrence 0.062 -0.013 0.496 1.00 -0.044
Contrast

Mu Hurst
Coefficient -0.287 0.123 0.068 -0.044 1.00

Table 4. Correlation among the Selected Features for Window Size of 0.75mm

Table 5 denotes the performance parameters obtained using the 5 selected features extracted from
parametric images formed using a window size of 0.75mm. A Classification accuracy of 89.2% was
obtained using the selected features, along with a Sensitivity of 0.577 and a Specificity of 0.971.
 False True False True Sensitivity Specificity Classification
Negative Negative Positive positive Accuracy
11 101 3 15 0.577 0.971 89.2%
Table 5. Performance Parameters before the Tuning of Decision Threshold

To classify all the malignant cases accurately (i.e., to make False Negatives, FN=0) and to maintain a
reasonable value of the cases requiring further biopsy (i.e., False Positives), hyper parameter tuning was
conducted for the threshold value of AUC, keeping the AUC constant to 0.9712. While performing the 5-
fold cross validation with Linear SVM Classifier, each data from every 5-test dataset was assigned a unique
threshold value on which they were later classified. All these threshold values were obtained using the
performance curve function from MATLAB. These threshold values ranged from a maximum to a minimum
value and any specific value could be chosen as the divider, above which all the data were assigned a
certain class, and all the data below the opposite class. While the MATLAB classifier application chose a
default divider threshold value and deduced an output based on it, our attempt was to perform
classification choosing different threshold values and using that to determine the output. Our goal was to
obtain the most optimal threshold in terms of Sensitivity, Specificity as well as Accuracy. Table 6 denotes
the accuracy scores and other quantities at different decision thresholds.

Threshold False True False True Sensitivity Specificity Classification

Negative Negative Positive positive Accuracy
0.888330 0 87 17 26 1 0.837 86.92%
0.849946 0 89 15 26 1 0.856 88.46%
0.726128 0 95 9 26 1 0.913 93.08%
0.578317 3 97 7 23 0.885 0.933 92.30%
0.019403 11 100 4 15 0.577 0.962 88.46%
Table 6. Tuning of Decision Threshold

From Table 6, it can be determined that a threshold value of 0.726128 provides the optimal circumstances,
yielding an accuracy score of 93.08% along with a Sensitivity value of 1 and a Specificity value of 0.913.
Furthermore, a threshold of 0.726128 provides an FPR (False Positive Rate) of 8.65%, which is significantly
small, along with a FNR (False Negative Rate) of 0%.
4. Discussion
Ultrasound instruments are becoming more effective in distinguishing between cancerous and non-
cancerous solid breast tumors visually, and this has moved to a point where many radiologists recommend
periodic follow-ups based on breast ultrasound findings, without performing a biopsy. We studied the
classification performance of multiple features to determine whether such an approach might play a role
in reducing misdiagnoses or unnecessary biopsies. The results obtained from this study indicate that our
approach has the potential to distinguish between benign and malignant breast lesions with a very high
accuracy. Furthermore, upon tuning the decision threshold, not only were we able to reduce the number
of False Negatives to 0, the optimum threshold value resulted in a very minimal False Positive Rate (FPR)
of 8.65%. The final output of our proposed system yielded a classification accuracy of 93.08%, along with
a Sensitivity of 1 and a Specificity of 0.913. Thus, we believe that our proposed system has the potential
to assist clinicians in diagnosis, as well as significantly reduce the number of unnecessary biopsies.

Our system began with a total of 72 features extracted from 7 different types of Nakagami parametric
images. A mixture of morphometric features, elemental features and hybrid features were extracted from
each type of parametric image. Feature selection was then performed to identify the subset of features
that are the most effective in distinguishing between benign and malignant masses. This final subset
constituted 5 features: Aspect Ratio, Convexity, Mu Co-occurrence contrast, Mu Hurst Coefficient and
Phase Echogenicity.

The Aspect Ratio and Convexity are morphometric features that remain the same across all 7 images. The
remaining three features selected by our feature selection algorithm are elemental. The elemental
features examined in this study may be affected by breast composition (e.g., fatty, fibrous or glandular
breast types), as this varies between different people, and are likely to influence RF echoes. In this case,
the contralateral breast, which is presumed to be healthy, may be used to provide a baseline to compare
with the diseased breast (Kaplan, 2001). Two out of the three elemental features are from the 𝑚𝑚
parametric image. The Nakagami 𝑚𝑚 parameter has been found to be effective towards the classification
of benign and malignant masses in previous studies (Klimonda et al., 2019; Shankar et al., 2003, 2002,
2001). In our study, both features that were selected for classification from the 𝑚𝑚 Parametric images are
related to image texture. Thus, texture seems to be a good indicator for characterizing benign and
malignant masses, and future studies may analyze other textural features derived from the 𝑚𝑚 parametric
image. The third elemental feature is derived from the 𝛼𝛼 - phase image, or related to the phase of effective
cross-section of scatterers. As discussed previously, the complex parameter 𝛼𝛼 is dependent on the
number density of scatterers, the signal-to-noise ratio of the cross section as well as attenuation. Hence,
this parameter has the potential to be very effective in tissue characterization, and the significance of the
phase of this parameter may be explored in greater detail in future studies.

Hybrid features were not found to be effective in distinguishing between tumor types in this study. The
algorithm we used for feature selection didn’t consider the hybrid features as a part of the best subset of
features for classification, although this may differ with other FS algorithms.

In Figure 8, boxplot pairs corresponding to values of the selected 5 features determined for benign and
malignant lesions are analyzed. The morphometric features seem to be the most effective in
distinguishing between lesion types. From Figure 8(b) it is seen that the feature Convexity has the highest
distinguishing ability, followed closely by Aspect Ratio. Overlapping is observed for the elemental features,
although the median values are clearly separable between benign and malignant cases. Correlation
between the selected features is analyzed in Table 4, and no features were found to be significantly
correlated.

The dataset utilized for this study contained an imbalance between positive and negative classes (104
positive, 26 negative). Although we achieved very satisfactory performance, completely eliminating False
Negatives and reducing the False Positive Rate to 8.65%, future studies are recommended to utilize a
balanced dataset to avoid classifier biasing.

5. Conclusion
In this study, we propose a framework to classify cancerous and non-cancerous tumors of the breast
using a combination of comprehensive statistical features and machine learning. Our results indicate that
such a technique can play a significant role in the diagnostic process associated with detection of breast
cancer, as we were able to distinguish between tumor types with a high level of accuracy, as well as
provide a very low percentage of false positives. The scope of this study can be further enhanced by
incorporating statistical distributions such as the homodyne-K distribution. The current process remains
semi-automated because of the need to manually segment the lesion, however, automated boundary
detection using sophisticated image-segmentation methods may be implemented in the future to make
the process automated. Furthermore, a larger dataset may be used to validate the results obtained in this
study. Deep learning models such as the Convolutional Neural Network may also be incorporated to
extend the scope of this study further.

References
Alam, S.K., Feleppa, E.J., Rondeau, M., Kalisz, A., Garra, B.S., 2011. Ultrasonic Multi-Feature Analysis
Procedure for Computer-Aided Diagnosis of Solid Breast Lesions. Ultrasonic Imaging 33.
https://doi.org/10.1177/016173461103300102

American Cancer Society, 2021. How Common Is Breast Cancer? [WWW Document].
https://www.cancer.org/cancer/breast-cancer/about/how-common-is-breast-cancer.html.

Berg, W.A., Blume, J.D., Cormack, J.B., Mendelson, E.B., 2012. Training the ACRIN 6666 Investigators and
Effects of Feedback on Breast Ultrasound Interpretive Performance and Agreement in BI-RADS
Ultrasound Feature Analysis. American Journal of Roentgenology 199.
https://doi.org/10.2214/AJR.11.7324

Boote, E.J., Zagzebski, J.A., Madsen, E.L., Hall, T.J., 1988. Instrument-independent acoustic backscatter
coefficient imaging. Ultrasonic Imaging 10. https://doi.org/10.1016/0161-7346(88)90054-5

Brem, R.F., Lenihan, M.J., Lieberman, J., Torrente, J., 2015. Screening Breast Ultrasound: Past, Present,
and Future. American Journal of Roentgenology 204. https://doi.org/10.2214/AJR.13.12072

Byra, M., Nowicki, A., Wróblewska-Piotrzkowska, H., Dobruch-Sobczak, K., 2016. Classification of breast
lesions using segmented quantitative ultrasound maps of homodyned K distribution parameters.
Medical Physics 43. https://doi.org/10.1118/1.4962928

Byra, M., Piotrzkowska-Wroblewska, H., Dobruch-Sobczak, K., Nowicki, A., 2017. Combining Nakagami
imaging and convolutional neural network for breast lesion classification, in: 2017 IEEE
International Ultrasonics Symposium (IUS). IEEE. https://doi.org/10.1109/ULTSYM.2017.8092154

Destrempes, F., Cloutier, G., 2010. A Critical Review and Uniformized Representation of Statistical
Distributions Modeling the Ultrasound Echo Envelope. Ultrasound in Medicine & Biology 36.
https://doi.org/10.1016/j.ultrasmedbio.2010.04.001

Dobruch-Sobczak, K., Piotrzkowska-Wróblewska, H., Roszkowska-Purska, K., Nowicki, A., Jakubowski, W.,
2017. Usefulness of combined BI-RADS analysis and Nakagami statistics of ultrasound echoes in the
diagnosis of breast lesions. Clinical Radiology 72. https://doi.org/10.1016/j.crad.2016.11.009

Duffy, S.W., Tabár, L., Chen, H.-H., Holmqvist, M., Yen, M.-F., Abdsalah, S., Epstein, B., Frodis, E.,
Ljungberg, E., Hedborg-Melander, C., Sundbom, A., Tholin, M., Wiege, M., Åkerlund, A., Wu, H.-M.,
Tung, T.-S., Chiu, Y.-H., Chiu, C.-P., Huang, C.-C., Smith, R.A., Rosén, M., Stenbeck, M., Holmberg, L.,
2002. The impact of organized mammography service screening on breast carcinoma mortality in
seven Swedish counties. Cancer 95. https://doi.org/10.1002/cncr.10765

Dutt, V., Greenleaf, J.F., 1994. Ultrasound Echo Envelope Analysis Using a Homodyned K Distribution
Signal Model. Ultrasonic Imaging 16. https://doi.org/10.1006/uimg.1994.1016
Esserman, L.J., Thompson, I.M., Reid, B., Nelson, P., Ransohoff, D.F., Welch, H.G., Hwang, S., Berry, D.A.,
Kinzler, K.W., Black, W.C., Bissell, M., Parnes, H., Srivastava, S., 2014. Addressing overdiagnosis and
overtreatment in cancer: a prescription for change. The Lancet Oncology 15.
https://doi.org/10.1016/S1470-2045(13)70598-9

Flobbe, K., Bosch, A.M., Kessels, A.G.H., Beets, G.L., Nelemans, P.J., von Meyenfeldt, M.F., van
Engelshoven, J.M.A., 2003. The Additional Diagnostic Value of Ultrasonography in the Diagnosis of
Breast Cancer. Archives of Internal Medicine 163. https://doi.org/10.1001/archinte.163.10.1194

Förnvik, D., Zackrisson, S., Ljungberg, O., Svahn, T., Timberg, P., Tingberg, A., Andersson, I., 2010. Breast
tomosynthesis: Accuracy of tumor measurement compared with digital mammography and
ultrasonography. Acta Radiologica 51. https://doi.org/10.3109/02841850903524447

Gutlein, M., Frank, E., Hall, M., Karwath, A., 2009. Large-scale attribute selection using wrappers, in:
2009 IEEE Symposium on Computational Intelligence and Data Mining. IEEE.
https://doi.org/10.1109/CIDM.2009.4938668

Guyon, I., Weston, J., Barnhill, S., Vapnik, V., 2002. Gene selection for cancer classification using support
vector machines. Machine Learning 46. https://doi.org/10.1023/A:1012487302797

Hruska, D.P., 2009. Improved Techniques for Statistical Analysis of the Envelope of Backscattered
Ultrasound Using the Homodyned K Distribution . Urbana, Illinois.

Hruska, D.P., Oelze, M.L., 2009. Improved parameter estimates based on the homodyned K distribution.
IEEE Transactions on Ultrasonics, Ferroelectrics and Frequency Control 56.
https://doi.org/10.1109/TUFFC.2009.1334

Hsu, S.-M., Kuo, W.-H., Kuo, F.-C., Liao, Y.-Y., 2019. Breast tumor classification using different features of
quantitative ultrasound parametric images. International Journal of Computer Assisted Radiology
and Surgery 14. https://doi.org/10.1007/s11548-018-01908-8

John, G.H., Kohavi, R., Pfleger, K., 1994. Irrelevant features and the subset selection problem, in:
Proceedings of the Eleventh International Conference on Machine Learning: ICML’94. San
Francisco, pp. 121–129.

Kaplan, S.S., 2001. Clinical Utility of Bilateral Whole-Breast US in the Evaluation of Women with Dense
Breast Tissue. Radiology 221. https://doi.org/10.1148/radiol.2213010364

Klimonda, Z., Dobruch-Sobczak, K., Piotrzkowska-Wróblewska, H., Karwat, P., Litniewski, J., 2018.
Quantitative Ultrasound of Tumor Surrounding Tissue for Enhancement of Breast Cancer Diagnosis,
in: Rojas, I., Ortuño, F. (Eds.), Bioinformatics and Biomedical Engineering. Springer, Cham.
https://doi.org/10.1007/978-3-319-78759-6_18

Klimonda, Z., Karwat, P., Dobruch-Sobczak, K., Piotrzkowska-Wróblewska, H., Litniewski, J., 2019. Breast-
lesions characterization using Quantitative Ultrasound features of peritumoral tissue. Scientific
Reports 9. https://doi.org/10.1038/s41598-019-44376-z

Kolb, T.M., Lichy, J., Newhouse, J.H., 2002. Comparison of the Performance of Screening Mammography,
Physical Examination, and Breast US and Evaluation of Factors that Influence Them: An Analysis of
27,825 Patient Evaluations. Radiology 225. https://doi.org/10.1148/radiol.2251011667
Langley, P., 1995. Elements of Machine Learning. Morgan Kaufmann, San Francisco.

Langley, P., 1994. Selection of relevant features in machine learning.

Larrue, A., Noble, J.A., 2011. Nakagami imaging with small windows, in: 2011 IEEE International
Symposium on Biomedical Imaging: From Nano to Macro. IEEE.
https://doi.org/10.1109/ISBI.2011.5872546

Leconte, I., Feger, C., Galant, C., Berlière, M., Berg, B. vande, D’Hoore, W., Maldague, B., 2003.
Mammography and Subsequent Whole-Breast Sonography of Nonpalpable Breast Cancers: The
Importance of Radiologic Breast Density. American Journal of Roentgenology 180.
https://doi.org/10.2214/ajr.180.6.1801675

Liao, Y.-Y., Li, C.-H., Tsui, P.-H., Chang, C.-C., Kuo, W.-H., Chang, K.-J., Yeh, C.-K., 2012. Strain-
compounding technique with ultrasound Nakagami imaging for distinguishing between benign and
malignant breast tumors. Medical Physics 39. https://doi.org/10.1118/1.3700167

Liao, Y.-Y., Tsui, P.-H., Li, C.-H., Chang, K.-J., Kuo, W.-H., Chang, C.-C., Yeh, C.-K., 2011. Classification of
scattering media within benign and malignant breast tumors based on ultrasound texture-feature-
based and Nakagami-parameter images. Medical Physics 38. https://doi.org/10.1118/1.3566064

Mamou, J., Oelze, M.L. (Eds.), 2013. Quantitative Ultrasound in Soft Tissues. Springer Netherlands,
Dordrecht. https://doi.org/10.1007/978-94-007-6952-6

Mayo, R.C., Kent, D., Sen, L.C., Kapoor, M., Leung, J.W.T., Watanabe, A.T., 2019. Reduction of False-
Positive Markings on Mammograms: a Retrospective Comparison Study Using an Artificial
Intelligence-Based CAD. Journal of Digital Imaging 32. https://doi.org/10.1007/s10278-018-0168-6

Mendelson, E., Böhm-Vélez, M., Berg, W., 2013. Ultrasound, in: ACR BI-RADS® Atlas, Breast Imaging
Reporting and Data System.

Mohana Shankar, P., 2000. A general statistical model for ultrasonic backscattering from tissues. IEEE
Transactions on Ultrasonics, Ferroelectrics and Frequency Control 47.
https://doi.org/10.1109/58.842062

Molthen, R.C., Shankar, P.M., Reid, J.M., 1995. Characterization of ultrasonic B-scans using non-rayleigh
statistics. Ultrasound in Medicine & Biology 21. https://doi.org/10.1016/S0301-5629(94)00105-7

Nakagami, M., 1960. The m-Distribution—A General Formula of Intensity Distribution of Rapid Fading,
in: Statistical Methods in Radio Wave Propagation. Elsevier. https://doi.org/10.1016/B978-0-08-
009306-2.50005-4

Nam, K., Rosado-Mendez, I.M., Wirtzfeld, L.A., Ghoshal, G., Pawlicki, A.D., Madsen, E.L., Lavarello, R.J.,
Oelze, M.L., Zagzebski, J.A., O’Brien, W.D., Hall, T.J., 2012. Comparison of Ultrasound Attenuation
and Backscatter Estimates in Layered Tissue-Mimicking Phantoms among Three Clinical Scanners.
Ultrasonic Imaging 34. https://doi.org/10.1177/0161734612464451

Nam, K., Zagzebski, J.A., Hall, T.J., 2013. Quantitative Assessment of In Vivo Breast Masses Using
Ultrasound Attenuation and Backscatter. Ultrasonic Imaging 35.
https://doi.org/10.1177/0161734613480281
Narayanan, V.M., Shankar, P.M., Reid, J.M., 1994. Non-Rayleigh statistics of ultrasonic backscattered
signals. IEEE Transactions on Ultrasonics, Ferroelectrics and Frequency Control 41.
https://doi.org/10.1109/58.330265

Oelze, M.L., 2012. Quantitative ultrasound techniques and improvements to diagnostic ultrasonic
imaging, in: 2012 IEEE International Ultrasonics Symposium. IEEE.
https://doi.org/10.1109/ULTSYM.2012.0058

Oelze, M.L., Mamou, J., 2016. Review of Quantitative Ultrasound: Envelope Statistics and Backscatter
Coefficient Imaging and Contributions to Diagnostic Ultrasound. IEEE Transactions on Ultrasonics,
Ferroelectrics, and Frequency Control 63. https://doi.org/10.1109/TUFFC.2015.2513958

Piramuthu, S., 2004. Evaluating feature selection methods for learning in data mining applications.
European Journal of Operational Research 156. https://doi.org/10.1016/S0377-2217(02)00911-6

Sadeghi-Naini, A., Suraweera, H., Tran, W.T., Hadizad, F., Bruni, G., Rastegar, R.F., Curpen, B., Czarnota,
G.J., 2017. Breast-Lesion Characterization using Textural Features of Quantitative Ultrasound
Parametric Maps. Scientific Reports 7. https://doi.org/10.1038/s41598-017-13977-x

Sainsbury, R., 2013. Breast Surgery, 5th ed. Saunders Ltd, London.

Sehgal, C.M., Weinstein, S.P., Arger, P.H., Conant, E.F., 2006. A Review of Breast Ultrasound. Journal of
Mammary Gland Biology and Neoplasia 11. https://doi.org/10.1007/s10911-006-9018-0

Shankar, P.M., 1995. A model for ultrasonic scattering from tissues based on the K distribution. Physics
in Medicine and Biology 40. https://doi.org/10.1088/0031-9155/40/10/006

Shankar, P.M., Dumane, V.A., George, T., Piccoli, C.W., Reid, J.M., Forsberg, F., Goldberg, B.B., 2003.
Classification of breast masses in ultrasonic B scans using Nakagami and K distributions. Physics in
Medicine and Biology 48. https://doi.org/10.1088/0031-9155/48/14/313

Shankar, P.M., Dumane, V.A., Piccoli, C.W., Reid, J.M., Forsberg, F., Goldberg, B.B., 2002. Classification of
breast masses in ultrasonic b-mode images using a compounding technique in the nakagami
distribution domain. Ultrasound in Medicine & Biology 28. https://doi.org/10.1016/S0301-
5629(02)00617-8

Shankar, P.M., Dumane, V.A., Reid, J.M., Genis, V., Forsberg, F., Piccoli, C.W., Goldberg, B.B., 2001.
Classification of ultrasonic B-mode images of breast masses using Nakagami distribution. IEEE
Transactions on Ultrasonics, Ferroelectrics and Frequency Control 48.
https://doi.org/10.1109/58.911740

Shankar, P.M., Molthen, R., Narayanan, V.M., Reid, J.M., Genis, V., Forsberg, F., Piccoli, C.W.,
Lindenmayer, A.E., Goldberg, B.B., 1996. Studies on the use of non-Rayleigh statistics for ultrasonic
tissue characterization. Ultrasound in Medicine & Biology 22. https://doi.org/10.1016/0301-
5629(96)00080-4

Shoma, A., Moutamed, A., Ameen, M., Abdelwahab, A., 2006. Ultrasound for Accurate Measurement of
Invasive Breast Cancer Tumor Size. The Breast Journal 12. https://doi.org/10.1111/j.1075-
122X.2006.00249.x
Sung, H., Ferlay, J., Siegel, R.L., Laversanne, M., Soerjomataram, I., Jemal, A., Bray, F., 2021. Global
Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers
in 185 Countries. CA: A Cancer Journal for Clinicians 71. https://doi.org/10.3322/caac.21660

Tadayyon, H., Sadeghi-Naini, A., Wirtzfeld, L., Wright, F.C., Czarnota, G., 2014. Quantitative ultrasound
characterization of locally advanced breast cancer by estimation of its scatterer properties. Medical
Physics 41. https://doi.org/10.1118/1.4852875

Trop, I., Destrempes, F., el Khoury, M., Robidoux, A., Gaboury, L., Allard, L., Chayer, B., Cloutier, G., 2015.
The Added Value of Statistical Modeling of Backscatter Properties in the Management of Breast
Lesions at US. Radiology 275. https://doi.org/10.1148/radiol.14140318

Tsui, P.-H., 2015. Ultrasound Detection of Scatterer Concentration by Weighted Entropy. Entropy 17.
https://doi.org/10.3390/e17106598

Tsui, P.-H., Chang, C.-C., 2007. Imaging Local Scatterer Concentrations by the Nakagami Statistical
Model. Ultrasound in Medicine & Biology 33. https://doi.org/10.1016/j.ultrasmedbio.2006.10.005

Tsui, P.-H., Chang, Chien-Cheng, Ho, M.-C., Lee, Y.-H., Chen, Y.-S., Chang, Chien-Chung, Huang, N.E., Wu,
Z.-H., Chang, K.-J., 2009. Use of Nakagami Statistics and Empirical Mode Decomposition for
Ultrasound Tissue Characterization by a Nonfocused Transducer. Ultrasound in Medicine & Biology
35. https://doi.org/10.1016/j.ultrasmedbio.2009.08.003

Tsui, P.-H., Chen, C.-K., Kuo, W.-H., Chang, K.-J., Fang, J., Ma, H.-Y., Chou, D., 2017. Small-window
parametric imaging based on information entropy for ultrasound tissue characterization. Scientific
Reports 7. https://doi.org/10.1038/srep41004

Tsui, P.-H., Yeh, C.-K., Liao, Y.-Y., Chang, C.-C., Kuo, W.-H., Chang, K.-J., Chen, C.-N., 2010. Ultrasonic
Nakagami Imaging: A Strategy to Visualize the Scatterer Properties of Benign and Malignant Breast
Tumors. Ultrasound in Medicine & Biology 36. https://doi.org/10.1016/j.ultrasmedbio.2009.10.006

Vlahiotis, A., Griffin, B., Stavros MD, F.A.T., Margolis, J., 2018. Analysis of utilization patterns and
associated costs of the breast imaging and diagnostic procedures after screening mammography.
ClinicoEconomics and Outcomes Research Volume 10. https://doi.org/10.2147/CEOR.S150260

Wagner, R.F., Insana, M.F., Brown, D.G., 1987. Statistical properties of radio-frequency and envelope-
detected signals with applications to medical ultrasound. Journal of the Optical Society of America
A 4. https://doi.org/10.1364/JOSAA.4.000910

Wagner, R.F., Smith, S.W., Sandrik, J.M., Lopez, H., 1983. Statistics of Speckle in Ultrasound B-Scans. IEEE
Transactions on Sonics and Ultrasonics 30. https://doi.org/10.1109/T-SU.1983.31404

Welch, H.G., Black, W.C., 2010. Overdiagnosis in Cancer. JNCI Journal of the National Cancer Institute
102. https://doi.org/10.1093/jnci/djq099

Yao, L.X., Zagzebski, J.A., Madsen, E.L., 1990. Backscatter coefficient measurements using a reference
phantom to extract depth-dependent instrumentation factors. Ultrasonic Imaging 12.
https://doi.org/10.1016/0161-7346(90)90221-I
Zhao, Z., Morstatter, F., Sharma, S., Anand, A., 2010. Advancing Feature Selection Research . ASU
Feature Selection Repository 1–28.