You are on page 1of 16

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/326621132

Comparison of rotation invariant local frequency, LBP and SFTA methods for
breast abnormality classification

Article  in  International Journal of Signal and Imaging Systems Engineering · January 2018


DOI: 10.1504/IJSISE.2018.093266

CITATIONS READS

0 141

3 authors, including:

Spandana Paramkusham
Birla Institute of Technology and Science Pilani
9 PUBLICATIONS   29 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Spandana Paramkusham on 26 November 2018.

The user has requested enhancement of the downloaded file.


136 Int. J. Signal and Imaging Systems Engineering, Vol. 11, No. 3, 2018

Comparison of rotation invariant local frequency,


LBP and SFTA methods for breast abnormality
classification

Spandana Paramkusham*, Kunda M.M. Rao


and B.V.V.S.N. Prabhakar Rao
Department of Electrical Engineering,
BITS Pilani – Hyderabad Campus,
R.R. Dist. Hyderabad – 500078, India
Email: p2011013@hyderabad.bits-pilani.ac.in
Email: kundammrao@gmail.com
Email: budhiraju@hyderabad.bits-pilani.ac.in
*Corresponding author

Abstract: Breast cancer is the second most prominent cancer diagnosed among women. Digital
mammography is one of the effective imaging modalities used to detect breast cancer in early
stages. Computer-aided detection systems help radiologists to detect and diagnose abnormalities
earlier and faster in a mammogram. In this paper, a comprehensive study is carried out on
different feature extraction methods for classification of abnormal areas in a mammogram.
The prominent techniques used for feature extraction in this study are local binary pattern (LBP),
rotation invariant local frequency (RILF) and segmented fractal texture analysis (SFTA).
Features extracted from these techniques are then fed to a support vector machine (SVM)
classifier for further classification via 10-fold cross-validation method. The evaluation is
performed using image retrieval in medical applications (IRMA) database for feature extraction.
Our statistical analysis shows that the RILF technique outperforms the LBP and SFTA
techniques.

Keywords: breast cancer; mammograms; masses; microcalcification; feature extraction; SVM;


support vector machine.

Reference to this paper should be made as follows: Paramkusham, S., Rao, K.M.M. and
Prabhakar Rao, B.V.V.S.N. (2018) ‘Comparison of rotation invariant local frequency, LBP and
SFTA methods for breast abnormality classification’, Int. J. Signal and Imaging Systems
Engineering, Vol. 11, No. 3, pp.136–150.

Biographical notes: Spandana Paramkusham is presently a research scholar at BITS Pilani,


Hyderabad working on early detection of breast cancer using image processing techniques.
Her research interests include medical image processing, pattern recognition and machine
learning.

Kunda M.M. Rao joined Indian Institute of Science, Bangalore in 1973. He was responsible
for the design and development of Black and White and Colour Drum Scanner Imager,
the first time in India, for digitising and reproducing various types of photographic prints.
He received his PhD degree in Image Processing from SK University, Ananthapur.
Currently, he is an Adjunct Professor, BITS Pilani, Hyderabad campus. His areas of interests
include satellite data processing, image processing, medical imaging, data mining and photo
writing.

B.V.V.S.N. Prabhakar Rao is working as an Associate Professor in the Department of


Electrical Engineering, BITS Pilani, Hyderabad campus. He has 18 years of teaching
and research experience. He received his PhD degree from Indian Institute of Technology,
Delhi. His current research interest is in the area of nonlinear optics and biomedical signal
processing.

Copyright © 2018 Inderscience Enterprises Ltd.


Comparison of rotation invariant local frequency, LBP and SFTA methods for breast abnormality classification 137

1 Introduction have been applied on ROIs to detect breast cancer (Lladó et


al., 2009; Beura et al., 2015). However, LBP compromises
Breast cancer is the second most leading cause of death in
some textural information due to thresholding and SFTA
India and most common cancer in the world. Breast cancer
loses some textural information due to thresholding. Among
has been on a steady rise in this new period. About 155,000
the LBP and SFTA techniques, RILF is more robust to noise
new cases of breast cancer have been diagnosed in the year
and has a relatively small number of features (Rouzbeh et
2015 in India and 76,000 women were expected to die due
al., 2013). Hence, we made the first attempt to study the use
to breast cancer. Early detection of breast cancer helps in
of RILF technique and extended version of SFTA for
increasing survival rates of breast cancers. Digital
detecting normal and abnormal ROIs in mammograms.
mammography is a key imaging modality for breast
Then, we compared its performance with LBP and SFTA.
screening using x-rays. It detects cancer before physical
changes develop. However, dense mammogram shows a
small difference between normal and abnormal tissues due
2 Related works
to low X-ray attenuation (Suneeta et al., 2017). Computer-
aided diagnosis plays a vital role to assist radiologists in the Over the last few decades, researchers have proposed many
detection and analysis of abnormalities in a mammogram. feature extraction techniques for abnormality detection in
CAD system requires good image processing methods to mammograms. These techniques are used to classify
increase sensitivity and specificity (Hadjiiski et al., 2006). ROIs of mammograms into normal/abnormal and benign/
Four abnormal signs of breast cancer are: malignant. Haralick descriptors, correlogram function and
shape descriptors were proposed by de Nazare Silva et al.
• microcalcifications
(2015) to classify the mammogram ROIs into normal/
• masses abnormal. Sharma and Khanna (2015) used Zernike
moments of different orders as features for the classification
• bilateral asymmetry
of malignant and normal regions in a mammogram. Features
• architectural distortion (Salvado and Roque, 2005). are extracted from approximation bands of curvelet
transform by Gedik and Atasoy (2013) to distinguish
Microcalifications are tiny calcium deposits and appear as a mammograms into normal and abnormal regions. The ROIs
small bright region with different shapes on the breast of mammograms are subdivided into blocks and LBP is
(Tiedeu et al., 2012). Masses are the most common applied on each block by Lladó et al. (2009) to form a
asymmetric signs of cancer and appear brighter than the feature vector. This feature vector is submitted to the
surrounding tissue (Timp et al., 2007). Most benign masses classifier to characterise mammogram regions. Mass in
possess well-defined sharp borders, while malignant masses mammograms was detected using LBP variance and shape
often have ill-defined, microlobulated, or spiculated borders descriptors by Masmoudi et al. (2015). The 2D discrete
(Rangayyan et al., 2010). Bilateral asymmetry is an wavelet transforms has been applied by Edward et al. (2013)
asymmetry of the breast parenchyma between left and right to create a set of maps at different size scales. Then
breast and this asymmetry indicates breast cancer in its early statistical features were computed from each map, and a
stage. An architectural distortion on a mammogram is subset of these features was submitted to Bayesian
basically a disruption of the normal ‘random’ pattern of classifiers to detect mammogram abnormal areas. The
curvilinear and fine linear radiopaque structures. There is no rotation and scale invariant features are extracted by Ergin
visible mass, but the distortion often appears as a ‘stellate’ and Kilinc (2014) using a histogram of oriented gradients
shape or with radiating spiculation (Beheshti et al., 2016). (HOG), dense scale invariant feature transform (DSIFT),
Among all other image processing techniques that are used and local configuration pattern (LCP) methods to classify
for CAD, feature extraction is an important task. This breast tissues in mammogram into normal and abnormal.
process should properly delineate normal and abnormal Ericeira et al. (2013) proposed cross-variogram function to
areas in mammograms and the size of the extracted features detect structural variations (asymmetric regions) in left and
should be less so that the CAD system is computationally right breasts. Then, features are extracted using variogram
efficient. function to classify asymmetric regions into mass/non-mass.
In this paper we have extracted features from ROIs, Wajid and Hussain (2015) extracted Local energy-based
taken from mammogram images, using rotation invariant shape histogram (LESH) features and fed to support vector
local frequency (RILF), local binary pattern (LBP) and machine to classify ROIs into benign and malignant
segmented fractal texture analysis (SFTA) techniques. abnormalities. Wavelet and scaling coefficients are
However, these techniques have been predominantly used in extracted using spherical wavelet transform by Gorgel et al.
real-world texture analysis problems. LBP and its variants (2013) from breast abnormal regions to separate them into
are proposed by Ojala et al. (2001). It has been used in mass/non-mass and benign/malignant. The taxonomic
different fields, which include texture classification, diversity index (∆) and the taxonomic distinctness (∆n)
neonatal face expression recognition, stem cell were calculated by de Oliveira et al. (2015) from
classification, detection of cervical cancer etc. (Liao et al., phylogenetic trees obtained from breast ROIs. Then, these
2007; Loris Nanni et al., 2013; Yimo Guo et al., 2012). parameters are fed to classifiers to describe the ROIs into
SFTA is proposed by Costa et al. (2012). LBP and SFTA mass/non-mass. Beheshti et al. (2016) proposed new fractal-
138 S. Paramkusham et al.

based asymmetric features to classify breast into In LBP(P,R), gc represents centre pixel and g0, g1, …, gP–1
mass/normal, microcalcification/normal and architectural represent neighbouring pixels, P corresponds to quantisation
distortion/normal. Features are computed from grey level of angular space (number of neighbours) and R (Radius of
cooccurrence matrix and grey level run length matrix by the neighbourhood) corresponds to the spatial resolution of
Mohanty et al. (2013) from ROIs of mammograms to the operator. The exact coordinates of neighbour pixels are
distinguish benign masses from malignant. Reyad et al.
(2014) compared statistical, blockwise LBP, wavelet and
given by ( − R sin ( ) , Rcos ( )) ,
2π p
P
2π p
P
if the centre pixel gc
curvelet features for the detection of normal/abnormal cases have coordinates (0, 0). The LBP descriptors are calculated
in a mammogram. These are some feature extraction from the histogram of the labelled image. The histogram
techniques that are used to describe mammograms. After H(k) for k = 1,2,3, …, K is given as.
extracting the features, selecting a better classifier helps in M N
increasing the efficiency of CAD system. As SVM classifier H ( k ) = ∑∑I ( LBP( P , R ) (i, j )) (2)
performs well for two class study, we have applied SVM i =1 j =1

classifier and evaluated the performance using 10-fold


where, K is the maximum label value in the LBP(P, R) image.
cross-validation.
Figure 1(a) shows the circularity symmetric neighbour sets
The present work is organised as follows: In Section 3,
for P = 8 and R = 1. Figure 1(b) shows the basic example of
we briefly explain the methodology of different feature
LBP operator. There are many variants of LBP based on
extraction techniques. In Section 4, we describe how these
encodings, the shape of neighbourhood and transitions in
techniques are applied on ROIs. The performance of LBP,
binary code generated for each pixel. LBP variants which
RILF and SFTA techniques is evaluated and comparison
are based on transitions and rotation invariance include
of these techniques was discussed in Section 5. The
uniform LBP, rotation invariant LBP, rotation invariant
conclusions were written in Section 6.
uniform LBP (Ojala et al., 2001). The LBP variants which
are based on the shape of the neighbourhood include
elliptical binary pattern (EBP) and hyperbolic binary pattern
3 Feature extraction techniques (HBP), these variants are considered to detect anisotropic
In this following section, we briefly explain the overview of features (Nanni et al., 2013). The LBP variants which are
all feature extraction techniques that are used in this work. based on encoding include local ternary pattern (LTP) and
These techniques are tested on the mammogram ROIs local quinary pattern (LQP). These variants have extra
for validation. The ROIs have been extracted from discriminative power and less sensitive to noise in uniform
mammogram using segmentation algorithm. Our focus is regions when compared to LBP. LTP was used for image
mainly on the classification of ROIs into normal/abnormal. retrieval by Vipparthi and Nagar (2015) and LQP was used
The abnormal ROIs can be either microcalcification or for stem cell classification by Paci et al. (2013). In order to
mass. study the application of LBP and its variants on
mammograms, we have extracted features from ROIs using
3.1 LBP and its variants uniform LBP, rotation invariant uniform LBP, rotation
uniform LBP, uniform LTP, rotation invariant uniform LTP,
3.1.1 LBP rotation uniform LTP, uniform LQP, rotation invariant
uniform LQP and rotation uniform LQP. LBP variants
The local binary pattern is an efficient rotation invariant
based on encodings are discussed below.
method for texture analysis. LBP has been presented by
many authors in many applications because of its Figure 1 (a) Circular symmetric neighbour for P = 8 and R = 1
computational efficiency, high discriminative power and and (b) Illustration of LBP operator
less vulnerability to illumination changes. This method is
first introduced by Ojala et al. (2001). They have considered
statistics of grey level differences. LBP operator assigns the
binary code i.e, 0 or 1 to each pixel based on the grey level
difference of that particular pixel and its neighbourhood.
The neighbouring pixel values which are not at the centre
are obtained by interpolation. For a centre pixel gc, LBP
code is defined as
p = P −1
LBP( P , R ) = ∑ s( g
p =0
p − g c )2 p (1)

where
s(x) = 1 for, x >= 0
= 0 for, x < 0
Comparison of rotation invariant local frequency, LBP and SFTA methods for breast abnormality classification 139

3.1.2 Local ternary pattern where |fn(x, y)| is the magnitude of the nth(1 ≤ n ≤ 2)
frequency component of local function LCF(N, R) at (x, y),
The difference between the centre pixel and neighbourhood
and W and H are width and height of the image respectively.
pixel is encoded into three values with –τ 2 ≤ user
When the 2D spectrum is circularly shifted, range changes
threshold (τ). LTP coding strategy can be given as from [0: W–1, 0: H–1] to [–W/2: W/2–1, –H/2: H/2–1]. To
p = P −1 have rotation invariance property for these features, circular
LTP( P , R ) = ∑ s( g
p =0
p − g c )2 p (3) bandpass disc shape filters are applied to the spectrum of
frequency channels. These filters are defined as follows
where s(x) = 1 for x >= τ
1 if r1≤ x 2 + y 2 ≤ r 2
= 0 for –τ < x < τ Dr1, r 2 ( x, y ) =  (7)
0 otherwise
=–1 otherwise.
where r1 and r2 are the inner and outer radius of disc shape
filters. The rotation invariant local frequency magnitude
3.1.3 Local quinary pattern
descriptors (RILFMD) are computed as
The difference between the centre pixel and the W H

∑ ∑ CH ( k , l ) .D
−1 −1
neighbourhood pixel is encoded into three values with user 2 2 (k , l )
RILFMD ( r1, r 2, n ) = k =−W / 2 −H /2 n r1, r 2
thresholds (τ 1,τ 2). The definition for LQP is given as W H

∑ ∑ D (k , l )
−1 −1
2 2
k =−W / 2 −H /2 r1, r 2
p = P −1
LQP( P , R ) = ∑ s( g
p =0
p − g c )2 p
(4) (8)
where |CHn| is the magnitude of the Fourier spectrum
where s(x) = 2 for x >= τ
=1 for τ1 < x < τ2 3.3 SFTA
SFTA method involves two steps: In the first step, the input
= 0 for –τ1 ≥ x < τ1 grey level images are decomposed into a set of binary
= –1 for x < –τ1 images using two-threshold binary decomposition (TTBD).
Then in the second step, features like fractal dimensions,
= –2 for other wise. mean and size are computed for each binary image.
Step 1: Decomposition of image into binary images using
3.2 Rotation invariant local frequency
two threshold binary decomposition (TTBD)
RILF is a rotation invariant texture classification method
TTBD takes input as the grey level image I(x, y) and uses
based on local frequency components. In this method, local
multilevel Otsu’s algorithm to return binary images. This
circular function (LCF(N, R)(x, y)) is defined at each pixel to
algorithm calculates a set of threshold values (T) based on
obtain features. This local circular function considers N
minimisation of intra class variance and is applied to each
neighbouring samples on a circle with radius R at each pixel
image recursively until the desired nt (number of threshold
(x, y) as in LBP. In LBP, thresholding the neighbourhood
values) is reached. Initially, first set of binary images are
pixels compromises some textural information. To
obtained using the thresholds calculated from Otsu’s
overcome this problem 1D Fourier transform is applied on
algorithm. So, the total number of binary images is nt in the
(LCF(N, R)(x, y) = (t0, t1, t2, …, tN-1)). 1D Fourier transform of
first set. The second set of binary images are obtained by
LCF is given as
selecting pairs of thresholds from T.
2π i ( n −1) (5)
N −1 − I(x, y) is then decomposed into binary images by
fn = ∑tk e N
, ( n = 1, 2, … , N ) applying adjacent thresholds from T = {t, nl} where t ∈ T
k =0
and nl maximum grey level intensity in the image I(x, y).
The low-frequency components f1 and f2 are considered to The first set of binary images Ib1(x, y) is given by
obtain features. The reason for considering low-frequency
components is that they have 90% of the texture energy and 1 if I ( x, y ) >= t1 ∈ T
I b1 ( x, y ) =  (9)
they do not contain noise information which appears in high 0 otherwise
frequencies. Then, the magnitude based features are
The second set of binary images computed from pairs of
extracted from two low-frequency channels (f1 and f2).
thresholds (T = {t1, ... tnt, nl}) as follows
These features are obtained by computing circularly shifted
2D Fourier transform. The 2D spectrum of each frequency 1 if t1 < I ( x, y ) ≤ tu
channel CHn, (n = 1, 2) is given as I b 2 ( x, y ) =  (10)
0 otherwise
W −1H −1  xk yl 
−2 π i  + 
CH n ( k , l ) = ∑∑ f n ( x, y ) .e W H 
(6)
where tl and tu are upper and lower thresholds.
x =0 y =0
140 S. Paramkusham et al.

Step 2: Feature extraction from binary decomposed images: 4.3 Method 3


After applying TTBD algorithm, SFTA feature vector is 4.3.1 Segmental fractal texture analysis (SFTA)
extracted from the decomposed binary images. It includes
the mean of the grey level values, number of pixels and In this method, SFTA feature vectors have been extracted
fractal dimension from the boundary of the binary image. from ROIs with different thresholds of 3, 4, 5, 6, 7 and 8.
These features are then fed to SVM classifier for
verification. Figure 2 shows the methodology followed to
detect abnormalities in ROIs using LBP and its variants,
4 Implementation of feature extraction
RILF and SFTA.
techniques
In this section, we have explained the implementation of Figure 2 Block diagram for the ROI classification using LBP,
different feature extraction techniques as mentioned in the SFTA and RILF feature extraction techniques
above section on ROIs to extract features. These extracted
features are then fed to a SVM classifier (linear kernel) via
10 fold cross-validation method to assess the significance of
the feature extraction techniques.

4.3.2 Extended version of SFTA


4.1 Method 1
• Block wise SFTA (BSFTA): ROI is further sub divided
4.1.1 LBP into N X N blocks as SFTA features are extracted for
In this method, different variants of LBP were applied to the each block and concatenated to form feature vector and
ROIs. The histograms of all these LBP variants are taken as fed to SVM classifier for further validation.
feature vectors for further classification.
• Pyramidal SFTA (PSFTA): In this method, ROI is
decomposed into image pyramid via the Gaussian filter
4.1.2 Extended version of LBP’s which is also known as scale space representation of the
image. For each level, SFTA features are calculated and
• Block wise LBP (BLBP): In this method, ROI is
concatenated to form a feature vector. Further, each
subdivided into N × N blocks. LBP histogram features
level feature vectors were given to classifier for
are calculated for each block and concatenated together
validation. The block diagram shown in Figure 3
to form complete feature vector of that particular ROI.
explains the process of detection of normal and
Further, this feature vector is given to SVM classifier
abnormal ROIs using BLBP and BSFTA. The block
for validation.
diagram in Figure 4 shows the procedure to classify
• Pyramidal LBP (PLBP): In this method, ROI is ROIs into normal and abnormal using PLBP and
decomposed into image pyramid via the Gaussian filter PSFTA.
(Wang et al., 2011) which is also known as scale space
representation of the image. Let I and Gk denote the Figure 3 Block diagram for the ROI classification using BLBP
and BSFTA
original image and the kth level of the image pyramid,
respectively. Then the image pyramid can be obtained
by
Gk ( x, y ) = I k = 0
I (11)
= ∑∑W ( m, n ) G ( 2 x + m, 2 y + n ) , k > 0
m n
k −1

where x and y are spatial coordinates of the image, W(m, n)


is as a Gaussian filter of size m X n.
LBP histogram features are then extracted from each
level of image pyramid and concatenated to form a feature
Figure 4 Block diagram for the classification of ROIs using
vector for further evaluation BLBP and BSFTA

4.2 Method 2
4.2.1 RILF
In this method RILF feature extraction technique,
magnitude descriptors are calculated from the low-
frequency channels using 2D Fourier transform. These
magnitude descriptors are considered as feature vectors and
passed to SVM classifier.
Comparison of rotation invariant local frequency, LBP and SFTA methods for breast abnormality classification 141

5 Results and discussion Figure 5 ROIs of IRMA database: (a) mass;


(b) microcalcification and (c) normal
This section gives the results and discussion of RILF,
LBP and SFTA feature extraction techniques. Then, the
results of all types of feature extraction techniques
are compared.

5.1 Experimental setup


Evaluation of our study is carried out using 10 fold cross-
validation. In this validation method, feature data is divided
into 10 groups randomly. Among these groups, one group is
selected for test and other groups are for training. In this
procedure, each group will undergo testing and the mean of
all evaluation parameters from 10 groups is calculated to get
final accuracy, sensitivity and specificity. Sensitivity
measures the percentage of positive breast cancer cases that
are correctly classified. Specificity (true negative rate) is the
percentage of negative breast cancer cases that are correctly
classified. The expressions for sensitivity, specificity and
accuracy are given as.
Sensitivity ( Se) = TP / (TP + FN )

Specificity ( Sp ) = TN / (TN + FP )

Accuracy ( A) = (TP + TN ) / (TP + TN + FP + FN ) 5.3 Results of Method 1


where, TP, FP, TN, FN are true positive, false positive, true Method 1 was tested using ROIs with different LBP variants
negative and false negative respectively. such as uniform LBP, uniform rotation invariant LBP,
rotation invariant LBP, uniform LTP, uniform rotation
5.2 Database invariant LTP, rotation invariant LTP, uniform LQP,
uniform rotation invariant LQP and rotation invariant LQP.
The dataset used in this work is taken from image retrieval
We used three different neighbourhoods (8, 1), (12, 2) and
in medical applications (IRMA) project database. The
(16, 2) for testing.
database comprises of ROIs taken from digital mammogram
images of selected patients from the Radiological Diagnosis
Department of the University of Aachen. The datasets in the 5.3.1 LBP and its variants
IRMA project have three different cases of breast cancer: • Results for mass/normal classification
Normal (no cancer), benign and malignant. The abnormal
(benign and malignant) cases have three different types of In LBP, uniform LBP resulted in best values (accuracy of
abnormalities: Masses, microcalcifications and others 88.17%, sensitivity of 93.50% and specificity of 83.87%)
(Architectural distortion). IRMA database contains 2777 with (8, 1) neighbourhood as shown in Figure 6. With LTP,
ROIs of which 932 are normal, 1157 are mass and 688 are uniform LTP gave best values (accuracy of 91.47%,
microcalcification. The images in the dataset are of sensitivity of 99.22% and specificity of 85.23%) with (8, 1)
128 × 128 pixel size and at 256 grey levels. neighbourhood as shown in Figure 7. In Figure 8, we can
These ROIs undergo feature extraction process using observe that in LQP, rotational invariant LQP resulted in
RILF, LBP and SFTA techniques. Then, these features are best values (accuracy of 91.83%, sensitivity of 99.16% and
given to classifiers to detect abnormal ROIs. Classification specificity of 85.92%) with (12,2).
consists of two components. First one distinguishes • Results for microcalcification/normal classification
ROIs into mass/normal and second one delineates ROIs
into microcalcification/normal. To carry out these In LBP, uniform LBP resulted in best values (accuracy of
classifications we have considered two datasets from IRMA 82.52%, sensitivity of 86.50% and specificity of 77.12%)
database. Dataset 1 contains 1157 mass and 932 normal with (16, 2) neighbourhood as shown in Figure 9. In LTP,
ROIs. These ROIs are used for both testing and training uniform LTP obtained the best accuracy of 87.80%, with
in mass/normal classification. Dataset 2 encompasses 97.02% sensitivity and of 75.31% specificity using (8, 1)
688 microcalcification and 932 normal ROIs for neighbourhood as shown in Figure 10. In LQP, uniform
microcalcification/normal classification. Figure 5(a)–(c) rotation invariant LQP gave best values (accuracy of
show normal, mass and microcalcification ROIs taken from 86.40%, sensitivity of 94.44% and specificity of 75.55%)
IRMA database. with (12,2) neighbourhood as shown in Figure 11.
142 S. Paramkusham et al.

Figure 6 Results of LBP for mass/normal

Figure 7 Results of LTP for mass/normal

Figure 8 Results of LQP for mass/normal

Figure 9 Results of LBP for microcalcification/normal


Comparison of rotation invariant local frequency, LBP and SFTA methods for breast abnormality classification 143

Figure 10 Results of LTP for microcalcification/normal

Figure 11 Results of LQP for microcalcification/normal

Figure 12 Results of BLBP for mass/normal

Among all the LBP variants, rotational invariant LQP gave 1 Results for mass/normal
the highest accuracy of 91.83% for mass/normal
The best accuracy of 87.66% is achieved by BLBP with
classification. Uniform LTP gave the highest accuracy of
2 × 2 blocks per ROI and neighbourhood of (8, 1). Figure 12
87.80% for microcalcification/normal.
depicts that, sensitivity and specificity with BLBP are
94.73% and 75.17% respectively.
5.3.2 Extended versions of LBP
2 Results for microcalcification/normal
We have tested our dataset using the extended version of
LBP that are BLBP and PLBP. In this classification, best accuracy of 79.24% is achieved
with 4 × 4 blocks per ROI and neighbourhood (8, 1). From
• BLBP Figure 13 we can observe that BLBP features achieved
Uniform rotation invariant blockwise LBP was tested by sensitivity of 82.25% and specificity of 75.17%,
dividing ROIs or patches into N X N blocks, where N = 2, 4, respectively.
8 and 16.
144 S. Paramkusham et al.

Figure 13 Results of BLBP for microcalcification/normal

Figure 14 Results of PLBP for mass/normal

Figure 15 Results of PLBP for microcalcification/normal

• PLBP 2 Results for microcalcification/normal


PLBP features are extracted from ROIs using three In PLBP, the best accuracy of 76.93% is obtained when
decomposition levels 2,3and 4 with three neighbourhoods ROIs are decomposed into three levels and with
(8, 1), (12, 2) and (16, 2). We decomposed ROI the neighbourhood of (8, 1). PLBP method achieved
into different levels and at each level uniform rotation, sensitivity of 79.37% and specificity of 73.62% as shown in
invariant features are calculated. Then, these features at Figure 15.
each level are concatenated to form a complete feature
vector. 5.4 Results of Method 2
1 Results for mass/normal The rotation invariant local frequency magnitude descriptors
PLBP resulted in the best accuracy of 83.35% with the are calculated by considering 2D spectrum of two low-
neighbourhood (8, 1) and three-level decomposition. From frequency channels. This method is tested on ROIs by
Figure 14, we can observe that it has sensitivity of 86.52% considering five neighbourhoods [(N, R) = (8, 1), (16, 2),
and specificity of 80.79%. (24, 3), (28, 4), (32, 5)]. We have extracted features from
Comparison of rotation invariant local frequency, LBP and SFTA methods for breast abnormality classification 145

both channel 1 and channel 2. Figure 16 shows that RILF for microcalcification/normal classification. It depicts that
method gives best values (accuracy of 93.53%, sensitivity of the RILF gives best values (accuracy of 91.10%, sensitivity
99.51% and specificity of 88.71%) with neighbourhood of 98.04% and specificity of 81.17%) with the
(8, 1) for classifying mass/normal. Figure 17 shows results neighbourhood (8, 1).

Figure 16 Results of RILF for mass/normal

Figure 17 Results of RILF for microcalcification/normal

Figure 18 Results of SFTA for mass/normal

5.5 Results of Method 3 5.5.1 SFTA


Method 3 includes SFTA and extended versions of • Results for mass/normal: Figure 18 shows that SFTA
SFTA for extracting features. SFTA method is applied to method achieved the best accuracy of 89.70%,
all ROIs with a different number of thresholds (nt = 3, 4, 5, sensitivity of 95.31% and specificity of 41.09% with
6, 7, 8). nt = 3.
146 S. Paramkusham et al.

• Results for microcalcification/normal: Figure 19 shows features. These features extracted from each block are
that SFTA method achieved the best accuracy concatenated to form enhanced feature vector.
of 86.44%, sensitivity of 88.94% and specificity
1 Results for mass/normal
of 81.27% for microcalcification/normal with nt = 3.
The results shown in Figure 20 makes it clear that BSFTA
5.5.2 Extended version of SFTA gives the best accuracy of 88.47%, sensitivity of 89.16%,
and specificity of 42.54% with N = 2 and nt = 3 for
We have tested our ROIs with extended versions of SFTA mass/normal.
which includes BSFTA and PSFTA.
2 Results for microcalcification/normal
• BSFTA
From Figure 21, we can observe that BSFTA achieved
In BSFTA method, ROI is divided into N × N blocks and accuracy of 88.54%, sensitivity of 84.88% and specificity of
SFTA operator is applied to each block of ROI to extract 77.24% with N = 2 and nt = 3 for microcalcification/normal.
Figure 19 Results of SFTA for microcalcification/normal

Figure 20 Results of BSFTA for mass/normal

Figure 21 Results of BSFTA for microcalcification/normal


Comparison of rotation invariant local frequency, LBP and SFTA methods for breast abnormality classification 147

Figure 22 Results of PSFTA for mass/normal

Figure 23 Results of PSFTA for microcalcification/normal

• PSFTA 5.6 Comparison of all methods


PSFTA features are extracted by decomposition of ROIs 5.6.1 Quantitative comparison of all methods
using pyramidal filters into images. Then SFTA is applied at
each level of image pyramid to extract the features and these We have compared RILF method with best methods of LBP
features at each level are concatenated to form enhanced and SFTA. Table 1 shows mean accuracy, sensitivity and
feature vector. specificity calculated with RILF method and the best
methods of LBP and SFTA for mass/normal and
1 Results for mass/normal microcalcification/normal classification respectively. In
Figure 22 shows that PSFTA gives best accuracy of LBP methods, LQP attained the best accuracy of 91.83% for
91.98%, sensitivity of 97.29%, and specificity of 87.69% by mass/normal, and LTP attained the best accuracy of 87.80%
decomposing the images into two levels and selecting nt = 3 for microcalcification/normal, where as in SFTA methods,
for SFTA feature extraction. pyramidal SFTA achieved best accuracy of 91.98% and
89.81% for mass/normal and microcalcification/normal
2 Results for microcalcification/normal respectively. These values show that our method RILF
From Figure 23, we can observe that PSFTA achieved resulted in the best accuracy of 93.53% and 91.11% using
accuracy of 89.81%, sensitivity of 91.60% and specificity of (8, 1) neighbourhood with ch1 and ch2 for mass/normal and
83.90% at decomposition level 3 and number of threshold microcalcification/normal classification respectively. For
nt = 4 for microcalcification/normal. evaluation of these feature vectors, we have used linear
From the statistics, we observed that specificity in case SVM classifier via 10-fold cross-validation. Table 1 shows
of mass/normal is very less because SFTA method gave accuracy sensitivity and specificity per fold attained by
inconclusive results for many normal ROIs for all the RILF and best methods of LBP and SFTA.
thresholds (nt = 3, 4, 5, 6, 7, 8). The reason for such results
is because segmentation is performed globally based on 5.6.2 Statistical comparison between all methods
grey level intensity values and it does not consider In this section, we have selected the best case in each
illumination and local changes in the image. method and statistically compared using T-test. We applied
148 S. Paramkusham et al.

10-fold cross-validation method and calculated mean p-values for mass/normal and microcalcification/normal.
accuracy, mean sensitivity and mean specificity. Method 2 These values imply that the best case methods of method 1
attained the best accuracy of 93.53% and 91.11% for i.e., LTP and method 3 i.e., PSFTA are significantly
mass/normal and microcalcification/normal using RILF different from method 2 at 95% confidence level for
with (8,1) neighbourhood and two frequency channels. mass/normal. In case of microcalcification/normal, method
T-test is used with a confidence level of 90% to determine 1 is significantly different from method 2 in calculating
the two methods are significantly different. Any two accuracy and specificity but it is indifferent for sensitivity.
methods can be significantly different if p-value is less than Method 3 is significantly different from method 2 in
0.05. P-value is calculated by considering method 2 as calculating accuracy and sensitivity but it is indifferent for
reference using T-test. From Tables 2 and 3, we can observe specificity.

Table 1 Results of best methods of RILFMD, LBP and SFTA for mass/normal and microcalcification/normal

RILFMD with (8,1) neighbourhood and two low-frequency channels for microcalcification normal
Fold1 Fold 2 Fold 3 Fold 4 Fold 5 Fold 6 Fold 7 Fold 8 Fold9 Fold 10 Avg
Accuracy 93.78 92.87 91.15 91.67 91.24 90.54 89.95 89.89 89.91 90.06 91.11
Sensitivity 100 98.84 98.50 98.86 98.85 98.14 96.17 96.38 96.78 97.99 98.04
Specificity 82.19 82.13 82.52 82.27 80.26 80.59 80.53 80.12 80.66 80.66 81.17
RILFMD with (8,1) neighbourhood with two low frequency channels for mass/normal
Accuracy 93.75 93.05 93.29 93.53 93.87 93.93 93.98 93.72 93.30 92.92 93.53
Sensitivity 98.92 99.46 99.64 99.46 99.57 99.64 99.54 99.60 99.64 99.68 99.52
Specificity 89.57 87.88 88.18 88.74 89.27 89.34 89.49 88.97 88.18 87.47 88.71
LBP variant : uniform LTP with (8,1) neighbourhood for microcalcification/normal
Accuracy 86.42 88.85 88.25 87.79 87.27 87.02 88.01 88.36 88.13 87.90 87.80
Sensitivity 98.92 97.85 96.06 96.51 96.34 96.77 97.24 97.32 96.78 96.46 97.02
Specificity 69.57 76.64 77.67 76.00 75.00 73.85 75.52 76.23 76.41 76.31 75.31
LBP variant : uniform rotation invariant LQP with (12,2) neighbourhood for mass/normal
Accuracy 94.76 93.08 92.19 91.86 91.57 91.46 91.10 90.84 90.69 90.81 91.83
Sensitivity 98.94 99.47 99.64 99.20 99.36 99.28 99.23 99.19 98.69 98.61 99.16
Specificity 91.38 87.93 86.17 85.93 85.29 85.16 84.55 84.11 84.25 84.53 85.93
SFTA method: Pyramidal SFTA with nt=4 for microcalcification/normal
Accuracy 92.31 91.17 90.74 90.27 89.72 88.95 88.65 88.61 89.04 88.75 89.82
Sensitivity 90.32 91.94 91.76 91.69 92.70 91.41 91.26 91.55 91.54 91.85 91.60
Specificity 88.24 86.13 84.95 84.73 82.56 82.81 82.37 82.21 83.06 81.98 83.90
SFTA method: Pyramidal SFTA for mass/normal
Accuracy 92.43 92.41 92.56 91.99 91.86 91.74 91.35 91.59 91.94 92.12 91.98
Sensitivity 98.64 98.16 98.14 97.05 96.78 96.78 96.78 96.92 96.78 96.89 97.29
Specificity 87.48 87.98 87.73 87.50 87.29 87.89 87.70 87.70 87.82 87.85 87.69

Table 2 P-values for method 1 and best methods of LBP and Table 3 P-values for method 1 and best methods of LBP and
SFTA for mass/normal SFTA for microcalcification/normal

Method1: Method1:
RILFMD LQP SFTA pyramidal RILFMD LTP SFTA Pyramidal
Accuracy 2.147950902387617e-04 5.878239724561006e-10 Accuracy 2.565648424536226e-07 0.024439300047089
Sensitivity 0.005190864320234 5.259538042536624e-10 Sensitivity 0.292571618575745 3.882984228280946e-13
Specificity 5.125426209348066e-04 3.090083848066811e-06 Specificity 4.459507642243632e-08 0.094282813640476
Comparison of rotation invariant local frequency, LBP and SFTA methods for breast abnormality classification 149

5.7 Discussion Costa, A.F., Humpire-Mamani, G. and Machado Traina, A.J.


(2012) ‘An efficient algorithm for fractal analysis of textures’,
Experiment results of three well-known techniques show 25th SIBGRAPI Conference on Graphics, Patterns and
that RILF i.e., method 2 gives a highest accuracy. RILF not Images, Washington DC, USA, pp.39–46.
only dominates in the number of features compared to LBP de Nazare Silva, J., Carvalho Filho, A.O. Silva, A.C., De Paiva,
and SFTA methods, but is also more robust to noise. This A.C. and Gattass, M. (2015) ‘Automatic detection of masses
robustness to noise is achieved by selection of low- in mammograms using quality threshold clustering,
correlogram function and SVM’, Journal of Digital Imaging,
frequency channels which has maximum texture energy and
Vol. 28, No. 3, pp.323–337.
less noise information. In LBP methods, thresholding and
de Oliveira, F.S.S., de Carvalho Filho, A.O., Silva, A.C.,
binomial factor 2n on neighbours compromise textural
de Paiva, A.C. and Gattass, M. (2015) ‘Classification of
information. To preserve this information in RILF method breast regions as mass and non-mass based on digital
the grey level values surrounding the centre pixel are mammograms using taxonomic indexes and SVM’,
converted into the frequency domain. In LBP methods by Computers in Biology and Medicine, Vol. 57, pp.42–53.
changing R and N values, the number of features Edward, J.K., Barnett, M.G. and Chytyk-Praznik, K. (2013)
exponentially grows. However, in RILF method features ‘Automatic detection of anomalies in screening
can be extracted at different scales with the same number of mammograms’, BMC Medical Imaging, Vol. 13, No. 1,
features. In the case of SFTA method, local features of the p.43.
image are not considered, as the thresholding is performed Ergin, S. and Kilinc, K. (2014) ‘A new feature extraction
globally using Otsu’s method and a number of thresholds framework based on wavelets for breast cancer diagnosis’,
Computers in Biology and Medicine, Vol. 51, pp.171–182.
are to be manually selected. We observed that in SFTA
method as the number of thresholds increases the Ericeira, D.R., Silva, A.C., De Paiva, A.C. and Gattass, M. (2013)
‘Detection of masses based on asymmetric regions of digital
performance has been reduced drastically. Many cases were bilateral mammograms using spatial description with
unpredicted in SFTA method as no extra texture patterns variogram and cross-variogram functions’, Computers in
were identified with the increase in the number of Biology and Medicine, Vol. 43, No. 8, pp.987–999.
thresholds. Gedik, N. and Atasoy, A. (2013) ‘A computer-aided diagnosis
system for breast cancer detection by using a curvelet
transform’, Turkish Journal of Electrical Engineering and
6 Conclusion Computer Sciences, Vol. 21, No. 4, pp.1002–1014.
Gorgel, P., Sertbas, A. and Ucan, O.N. (2013) ‘Mammographical
In this paper, we have extracted features using LBP, SFTA, mass detection and classification using local seed region
and RILF methods to distinguish normal and abnormal growing–spherical wavelet transform (LSRG–SWT) hybrid
(mass and microcalcifications) ROIs. Among these scheme’, Computers in Biology and Medicine, Vol. 43, No. 6,
methods, RILF works in the frequency domain and it does pp.765–774.
not consider noise part of the image. This property adds to Guo, Y., Zhao, G. and PietikäInen, M. (2012) ‘Discriminative
RILF features to get high classification accuracy when features for texture description’, Pattern Recognition, Vol. 45,
compared to LBP and SFTA methods. The evaluation of No. 10, pp.3834–3843.
these methods has been carried out using IRMA database. Hadjiiski, L., Sahiner, B. and Chan, H-P. (2006) ‘Advances in
RILF features achieved accuracy 93.53% and 91.11%, CAD for diagnosis of breast cancer’, Current Opinion in
Obstetrics and Gynecology, Vol. 18, No. 1, p.64.
sensitivity 99.51% and 98.04% and specificity 88.71% and
81.17% for mass/normal and microcalcification/normal Liao, S., Zhu, X., Lei, Z., Zhang, L. and Li, S.Z. (2007) ‘Learning
multi-scale block local binary patterns for face recognition’,
classification respectively. In summary, RILF has given best International Conference on Biometrics 2007, 27 August,
accuracy, sensitivity and specificity when compared to other Springer, Berlin, Heidelberg, pp.828–837.
methods. Further, in future, we would like to develop a new Lladó, X., Oliver A, Freixenet J, Martí R, Martí J. 2009)
feature vector which does three class study that classifies ‘A textural approach for mass false positive reduction
normal, benign and malignant ROIs irrespective of the type in mammography’. Computerized Medical Imaging and
of abnormalities present in mammogram Graphics, Vol. 33, No. 6, 1 September, pp.415–422.
Lladó, X., Oliver, A., Freixenet, J., Martí, R. and Martí, J. (2009)
‘A textural approach for mass false positive reduction in
References mammography’, Computerized Medical Imaging and
Graphics, Vol. 33, No. 6, pp.415–422.
Beheshti, S.M.A., Noubari, H.A., Fatemizadeh, E. and Khalili, M. Masmoudi, A.D., Ayed, N.G.B., Masmoudi, D.S. and Abid, R.
(2016) ‘Classification of abnormalities in mammograms by (2015) ‘Robust mass classification–based local binary pattern
new asymmetric fractal features’, Biocybernetics and variance and shape descriptors’, International Journal of
Biomedical Engineering, Vol. 36, No. 1, pp.56–65. Signal and Imaging Systems Engineering, Vol. 8, Nos. 1–2,
Beura, S., Majhi, B. and Dash, R. (2015) ‘Automatic pp.20–27.
characterization of mammograms using fractal texture Mohanty, A.K., Senapati, M.R., Beberta, S. and Lenka, S.K.
analysis and fast correlation based filter method’, (2013) ‘Texture-based features for classification of
Proceedings of the 2nd International Conference on mammograms using decision tree’, Neural Computing and
Perception and Machine Intelligence, 26–27 February, Applications, Vol. 23, Nos. 3–4, pp.1011–1017.
Kolkata, West Bengal.
150 S. Paramkusham et al.

Nanni, L., Paci, M., Brahnam, S., Ghidoni, S. and Menegatti, E. Sharma, S. and Khanna, P. (2015) ‘Computer-aided diagnosis of
(2013) ‘Virus image classification using different texture Malignant mammograms using Zernike moments and SVM’,
descriptors’, The 14th International Conference on Journal of Digital Imaging, Vol. 28, No. 1, pp.77–90.
Bioinformatics and Computational Biology (BIOCOMP’13), Suneeta, M., Lewis, S., Brennan, P., Noakes, J. and
Las Vegas, NV. Mello-Thoms, C. (2017) ‘The role of digital breast
Ojala, T., Valkealahti, K. and Oja, E. and Pietikäinen, M. (2001) tomosynthesis in the breast assessment clinic: a review’,
‘Texture discrimination with multidimensional distributions Journal of Medical Radiation Sciences, Vol. 64, 1 March,
of signed gray-level differences’, Pattern Recognition, pp.203–211.
Vol. 34, No. 3, pp.727–739. Tiedeu, A., Daul, C., Kentsop, A., Graebling, P. and Wolf, D.
Paci, M., Nanni, L., Lahti, A., Aalto-Setala, K., Hyttinen, J. and (2012) ‘Texture-based analysis of clustered
Severi, S. (2013) ‘Non-binary coding for texture descriptors microcalcifications detected on mammograms’, Digital Signal
in sub-cellular and stem cell image classification’, Current Processing, Vol. 22, No. 1, pp.124–132.
Bioinformatics, Vol. 8, No. 2, pp.208–219. Timp, S., Varela, C. and Karssemeijer, N. (2007) ‘Temporal
Rangayyan, R.M., Banik, S. and Leo Desautels, J.E. (2010) change analysis for characterization of mass lesions in
‘Computer-aided detection of architectural distortion in prior mammography’, IEEE Transactions on Medical Imaging,
mammograms of interval cancer’, Journal of Digital Imaging, Vol. 26, No. 7, pp.945–953.
Vo. 23, No. 5, pp.611–631. Vipparthi, S.K. and Nagar, S.K. (2015) ‘Directional local ternary
Reyad, Y.A., Berbar, M.A. and Hussain, M. (2014) ‘Comparison patterns for multimedia image indexing and retrieval’,
of statistical, LBP and multi-resolution analysis features for International Journal of Signal and Imaging Systems
breast mass classification’, Journal of Medical Systems, Engineering, Vol. 8, No. 3, pp.137–145.
Vol. 38, No. 9, pp.1–15. Wajid, S.K. and Hussain, A. (2015) ‘Local energy-based shape
Rouzbeh, M., Kalra, S. and Yang, Y-H. (2013) ‘Rotation invariant histogram feature extraction technique for breast cancer
local frequency descriptors for texture classification’, IEEE diagnosis’, Expert Systems with Applications, Vol. 42, No. 20,
Transactions on Image Processing, Vol. 22, No. 6, pp.6990–6999.
pp.2409–2419. Wang, W., Chen, W. and Xu, D. (2011) ‘Pyramid-based multi-
Salvado, J. and Roque, B. (2005) ‘Detection of calcifications in scale LBP features for face recognition’, International
digital mammograms using wavelet analysis and contrast Conference on Multimedia and Signal Processing (CMSP),
enhancement’, Intelligent Signal Processing, 2005 IEEE Vol. 1, pp.151–155.
International Workshop on, 1–3 September, Faro, Portugal.

View publication stats

You might also like