Detection of Breast Cancer using Curvelet

Transform and Adaptive Particle Swarm

Optimization Technique

Meena L. C1 Joe Prathap P. M2 Sankara Narayanan S3

2023 12th International Conference on Advanced Computing (ICoAC) | 979-8-3503-1821-0/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICoAC59537.2023.10249367

Research Scholar, Department of IEEE Senior Member & Professor, Department of Computing
Computer Science and Engineering, Department of Computer Science and Technologies, SRM Institute of Science
R.M.D. Engineering College, Engineering, and Technology,
Tiruvallur, India. R.M.D. Engineering College, Kattankulathur,India
[email protected] [email protected]
Tiruvallur, India.
[email protected]

Abstract— The breast cancer is the most prevalent automatic mammogram classification aids to provide proper
malignancy. Better chance of curing breast cancer is early treatment. An approach to feature extraction using
detection, which can also lower mortality rates. The best multiresolution representation was introduced by Mohamed
technique for early breast disease detection is the Meselhy Eltoukhy et al. Curvelet generates an accuracy of
mammography. In the suggested approach, curvelet transform 91%, according to experimental results, which is superior to
is utilized to extract features, and adaptive particle swarm wavelet [1]. Using several wavelet types, Prakash et al. [2]
optimization helps to choose the eminent features. Adaptive devised a multilayer wavelet decomposition technique.
Particle Swarm Optimization has been devised to speed up and According to the experimental results, Daubechies wavelet
simplify the process of feature selection and Support Vector
with fourth level decomposition had the better detection
Machine (SVM) aids in breast cancer classification. We present
performance [2]. Multiresolution study of digital
an Adaptive Particle Swarm Optimization (APSO) that
outperforms Particle Swarm Optimization (PSO) regarding mammograms is suggested by Essam A. Rashed et al. [3].
search efficiency. The suggested model is examined using a This structure is built on the multilevel decomposition of a
collection of 332 images from the Mammographic Image small number of the largest wavelet coefficients. According
Analysis Society (MIAS) database. The executed findings are to the results, the structure has a good level of categorization
compared with the old transforms, and the results demonstrate precision [3]. The effectiveness of Curvelet in representing
that the suggested model has higher detection accuracy rates discontinuities along curves or edges in pictures or objects
than the earlier approaches. has been shown in several investigations [4]. Mohamed
Meselhy Eltoukhy et al. compare wavelet and curvelet for the
Keywords— Adaptive Particle Swarm Optimization, Breast identification of breast cancer in their study [5] and found that
cancer, Curvelet Transform, Mammogram, Segmentation. the curvelet transform performance is superior to the wavelet
[5]. Dheeba et al. [6] proposed PSOWNN (Particle Swarm
I. INTRODUCTION Optimized Wavelet Neural Network), a new technique for
Breast cancer is the most common illness among women, detecting breast disease in mammograms, based on texture
causing 1.3 million new cases annually. The advances in extraction and pattern classification. It has a receiver
screening and treatment, as well as computer-assisted operating characteristics curve area of 0.96853 and
diagnosis, account for 28% to 65% of the mortality decrease. sensitivity of 94.167%. In order to diagnose breast cancer, the
Many breast cancer instances in the United States have not PSO algorithm was applied after extracting curvelet
yet been diagnosed as of 2017, according to data. Breast coefficients from the area of interest (ROI) Mammogram
tumor can spread to other parts of the body or infect pictures [9].
neighbouring tissues. With the computer-assisted diagnosis,
To analyze images with various structures, the wavelet
radiologists can read images more consistently and
transform is the best tool. It provides a technique for space
accurately. Due to the mammogram's high sensitivity to tiny
scale representations by differentiating between various
lesions and relatively low cost, it is widely utilized in the
spatial localities and breaking up images into various scale
early breast cancer screening process. There have been
directions. The wavelet transform makes use of inner
numerous approaches for identifying breast cancer developed
products, such as the Fourier transform, to assess the
during the last 30 years. Our proposed model produces a
similarity between two signals. Wavelets do not provide
higher accuracy rate and less computation time.
better edge representation than curvelet transform. It offers a
Curvelet transform is utilized in the proposed model to aid lower classification accuracy rate and so the curvelet
in the diagnosis of breast cancer in mammograms, and transform is the better choice [5]. PSO is a population-based
adaptive particle swarm optimization is employed to boost optimization approach mimicking bird swarming behavior
accuracy. [7]. PSO represents potential solutions by particles learning
how the swarm participates in finding the global optimum,
II. LITERATURE SURVEY with fitness values and velocities. A collection of random
Digital mammography aids in diagnosis of breast cancer. particles is used to initialize PSO. Then, using gbest and pbest
Early breast cancer detection reduces mortality, and values to update velocity and position, it searches for optima.

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 20,2023 at 02:41:52 UTC from IEEE Xplore. Restrictions apply.
 The optimization procedure is continued up until the least
error condition is met or the limited number of iterations is
achieved. Using a hybrid convolutional neural network and
predation optimization algorithm, Kaladevi et al. [11]
suggested a method for diagnosing breast cancer.
III. METHODOLOGY
The proper interpretation of mammogram images can be
aided by digital image processing techniques. From the
MIAS database, 332 mammogram pictures are taken. The
primary approach used to improve the mammography images Fig.2. Curvelet tiling
is contrast limited adaptive histogram equalization (CLAHE),
which is subsequently used as an input to bilateral filtering in The curvelet transform is executed in two dimensions in
the frequency domain [8]. CLACHE is used to improve local the form of R2, where x serves the spatial variable, ω serves
contrast and enhance image edges. A bilateral filter aids in as the frequency domain variable, and r1 and θ as polar
smoothing images and reduces noise while keeping edges. coordinates. The terms "radial window" and "angular
window" refer to two windows, W1(r1) and V1(t),
respectively. W1 accepts genuine arguments that are in the
Feature affirmative and are backed by evidence because they are
MIAS Extraction by
Dataset Preprocessing smooth, nonnegative, and real-valued. The following
Curvelet
Transform
conditions should be satisfied:
∑ 1 2 1 = 1, 1 ∈ 3/4,3/2 (1)

∑ 1 1− = 1, 1 ∈ −1/2,1/2 (2)
Dimension
SVM Classifier Reduction using The frequency window U1i is defined by
[ /!⌋ #
APSO
1 1, =2 1 2 1 1 % (3)
$

Where [i/2] is the integer component of i/2. The W1 and

V1 polar wedges define the Ui support. The symmetrized
Normal Cancer version 1 1, + 1 1, + π is employed to find
curvelet.
The waveform ( ) is defined by its Fourier transform
Fig. 1. Proposed breast cancer detection model

An 80/20 split was then used to separate the preprocessed (ˆ + = 1 + . All curvelets are produced by rotations
dataset into training and testing data. In order to extract and translations of mother curvelet ( . A sequence of
features, a curvelet transform is employed, and the eminent translation parameters , = ,- , , ∈ . and rotation angles
features are selected using adaptive particle swarm = 2/ ⋅ 2 ⌊ / ⌋ are introduced, with = 0,1,2 … such that
optimization. Breast cancer is categorized using support 0 ≤ ≤ 2/ (the spacing between consecutive angles is
vector machines. Fig. 1 shows the proposed approach scale-dependent). The curvelet functions are functions of ) =
combining curvelets and APSO. The preprocessed images are )- , ) defined at scale 2 , orientation angle and
,
inputted to extract the features, and then pick the features position )5 = 6#7 8,- ⋅ 2 , , ⋅ 2
- /
9 by
using APSO. Breast cancer can be identified using the SVM
,
classifier, a supervised learning system. ( , ,5 ) = ( 6#7 8) − )5 9% (4)
A. Curvelet Transform
where 6# is the rotation by radians and 6# - is its inverse,
The curvelet transform is a variant of the wavelet
cos sin
6# = % , 6# - = 6#? = 6 #.
transform that has demonstrated to be particularly effective
in identifying picture activity along curves. Curvelet −sin cos
transforms record structural activity along the radial "wedge"
in the frequency domain, offering a more effective method A curvelet coefficient is represented by
HHHHHHHHHH
A B, , , : = DE! F ) ( G, ,5 ) I)
for coding image edges compared to wavelet transforms. The
originality of the curvelet transform is based on the idea of (5)
anisotropic scaling [5]. The discrete curvelet transform
representation of images is based on the concept of There are two ways to perform fast digital curvelet
superimposing multiple length and width functions. Fig. 2 transforms: unequispaced FFTs or wrapping. Wrapping is
displays the curvelet analysis scheme and the wedge samples utilized in this investigation and the method has the following
are shaded. steps
1. Perform a 2D FFT to obtain Fourier samples.
F̂[J1- , J1 K, −J1/2 ≤ J1- , J1 < J1/2.
2. Determine the product for each scale B and angle ℓ,
˜ ,ℓ [J1- , J1 KF̂[J1- , J1 K.

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 20,2023 at 02:41:52 UTC from IEEE Xplore. Restrictions apply.
 3. Wrap this product results around the origin to get the Therefore, a dynamic velocity of particles may result. The
results proposed APSO algorithm is depicted in the Fig. 3.
F̃ ,ℓ [J1- , J1 K = 8 ˜1 ,ℓ F̂9[J1- , J1 K
where the range for J1- and J1 is
Algorithm 1 Prominent feature selection using APSO
0 ≤ J1- < O-, and 0 ≤ J1 < O ,
Input : Population number, maximum iteration number
lies between −//4 to //4
Output: Prominent features
1. Start
4. Use inverse 2D FFT to get the discrete coefficients
A Q B, ℓ, , .
2. Initialize population and set the parameters for
each particle
B. Dimension reduction of curvelet coefficients using 3. Find velocity and position of each particle using
Adaptive Particle Swarm Optimization Equation (6) & (7)
The wrapping approach produces a relatively big total 4. Calculate each particle’s fitness
curvelet coefficients. If all these coefficients are applied, the 5. If the fitness value is greater than the best (ppbest),
computation becomes high, raising the temporal complexity. the current value is assigned as the new ppbest.
To reduce time complexity and pick significant 6. Identify the particle with the highest fitness value
characteristics, Adaptive Particle Swarm Optimization has among all the particles as ggbest
been proposed and the characteristics that have been chosen 7. Update inertia weight using Equation (8)
are subsequently given into the SVM classifier. 8. Calculate the new modified velocity using
Equation (9)
A swarm is a collection of particles moving through the
9. Change particle position using Equation (7)
search space, each of which represents a potential solution
(fitness). Every particle has a velocity vector (v1i), a position 10. Check for stopping condition
vector (x1i), and the location where the particle encountered If the condition is met, stop
the best fitness (ppbesti) and the location of the top particle Else repeat iteration to the step 3
(ggbest) within the swarm. Every generation updates each 11. Stop
particle's velocity using their best position as in (6).
PSO is a successful global search technique and improved
R1 = R1 − 1 + S- × - × ppbest − )1 %+ analytical accuracy is attained by APSO within a manageable
processing time. The feature selection is to pick the fewest
S × × 8 ggbest − )1 9 6 characteristics that will allow the classifier to discriminate
Where r1(t) and r2(t) are assigned between 0 and 1 and are between normal and malignant.
not common for all iterations. The values s1 and s2 will be
constant and taken as 2. Every generation, each particle's
position is updated. This is accomplished by adding the
velocity vector to the position vector.
)1 = )1 − 1 + R1 (7)
In the normal PSO, there was no actual control
mechanism for the velocity. This might make the algorithm
behave inefficiently, particularly in the area around the global
minimum. In APSO, the inertia weight [10] is defined as a
function of the fitness function of a minimization problem's
local best (ppbest) and global best (ggbest) values.
((ppbesti)average) is the term used to describe the average of all
the individual’s best qualities within that generation.

Inertia weight, X1 = Y1.1 − Z

ggbest
(8)
8 ppbest 9
average

In the initial PSO, there was no velocity control

mechanism. This could lead to inefficient algorithm
behaviour, particularly in the proximity to the global
minimum. Inertia weight is introduced to alter the velocities
and it is modified in (9).
R1 = X1 R1 − 1 + S- × - × 8 ppbest − )1 9
+S × × 8 ggbest − )1 9 9
Fig. 3. APSO algorithm flowchart

C. Support Vector Machine

After iteration, inertia weight is updated with the best SVM creates a hyper-plane to separate dimensional data
fitness values from both the global and local levels. The into two classes. It is used for classifying mammogram
greater the difference between the global and local fitness images into two distinct subsets for training and the testing
values, the higher the inertia weight value. A rising ratio the classifiers. SVM, contrast to other machine learning
value shows that the mean best fitness value of the particles models, it learns from the support vectors rather than the
is approaching the global best value, necessitating a more correct and incorrect data. SVM typically does not experience
targeted search by reducing the particles' inertia weight.

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 20,2023 at 02:41:52 UTC from IEEE Xplore. Restrictions apply.
 the overfitting problem. SVM provides faster prediction TABLE I. PERFORMANCE METRICS COMPARISION
along with better accuracy.
Sl. Method Classification Sensitivity Specificity
IV. RESULTS AND DISCUSSIONS No. Accuracy
There are numerous models for feature selection and
1 Wavelet+ 88% 84.9% 81.4%
classification. To accurately identify breast cancer, it is SVM
crucial to choose the right methods 2 Curvelet + 92% 90.8% 89.8%
SVM
A. Input and Preprocessed Mammogram Images 3 Curvelet+ 96% 92.1% 91.7%
Fig. 4 shows the mammography input image. MIAS PSO+SVM
dataset has 1024x1024 images with 50% noisy background. 4 Curvelet+ 96.5% 93.8% 91.8%
Dark portions are removed using cropping operation and it APSO+SVM
eliminates the label and black background. The images are
enhanced using CLAHE and bilateral filtering. Cropped and
enhanced image is shown in Fig. 5. 98 96 96.5
96
B. Curvelet Transform

Percentage value
94 92
The Fast Discrete Curvelet Transform (FDCT) extracts 92
curvelet coefficients using the wrapping method, with a total 90 88
scale of log2 (min (M, N)) -3 and a total angle of 16 for 88
256x256 images.
86
84
82
Accuracy
Wavelet+SVM Curvelet+SVM
Curvelet+PSO+SVM Curvelet+APSO+SVM

Fig. 6. Comparision based on accuracy

95 93.8
92.1
Fig.4. Input image 90.8
Percentage value

90

84.9
85

80
Sensitivity

Wavelet+SVM Curvelet+SVM
Curvelet+PSO+SVM Curvelet+APSO+SVM

Fig. 5. Cropped and Enhanced Image Fig. 7. Comparision based on sensitivity

C. Performance Measures
Proposed system's performance was quantitatively 95
91.7 91.8
assessed against ground truth database. Table I shows the 89.8
Percentage value

performance measures of existing methods and the proposed 90

method. Curvelet outperforms wavelet in classification
accuracy at 92%. 85
81.4
The curvelet approach with PSO produces a 92.1% 80
sensitivity and a 91.7% specificity. APSO produces more
accuracy than the PSO model. APSO offers better 75
optimization ability and selects suitable features within an
Specificity
acceptable time. The performance comparision is shown in
Wavelet+SVM Curvelet+SVM
the Figures 6 to 8. In Fig. 9, the suggested algorithm's
convergence is compared with the other models. Curvelet+PSO+SVM Curvelet+APSO+SVM

Fig. 8. Comparision based on specificity

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 20,2023 at 02:41:52 UTC from IEEE Xplore. Restrictions apply.
 [7] R. C. Eberhart, and J. Kennedy, “Particle swarm optimization,” IEEE
0.2 International Conference on Neural Networks, pp. 1942–1948, 1995.
[8] R.C. Eberhart, M. Ravikumar, P. G. Rachana, and B. J. Shivaprasad,
0.15 “Enhancement of Mammogram Images Using CLAHE and Bilateral
Filter Approaches,” Cybernetics, Cognition and Machine Learning
MSE

0.1 Applications. Algorithms for Intelligent Systems, pp. 261–271, 2021.

[9] D. Saraswathi, D. Dharani and E. Srinivasan, “An efficient feature
0.05 extraction technique for breast cancer diagnosis using curvelet
transform and swarm intelligence,” IEEE International Conference on
0 Wireless Communications, Signal Processing and Networking, pp.
441-445, 2016.
[10] M. Senthil Arumugam and M. V. C. Rao, “On the improved
performances of the particle swarm optimization algorithms with
Iterations adaptive parameters, cross-over operators and root mean square (RMS)
Wavelet+SVM Curvelet+SVM variants for computing optimal control of a class of hybrid systems,”
Applied Soft Computing, vol. 8, pp. 324–336, 2008.
Curvelet+PSO+SVM Curvelet+APSO+SVM [11] R. Praveen, P. Kaladevi, V. V. Punitha, and D. Muthusankar, “Breast
cancer diagnosis using orca predation optimization algorithm,” Journal
Fig. 9. Convergence curve of Intelligent & Fuzzy Systems, 2023, in press.

V. CONCLUSION
We provide a brand-new technique for automatically
spotting breast cancer in mammography. The proposed
Computed Aided Diagnosis (CAD) procedure involves
mammography image transformation using curvelet
algorithm, followed by feature selection using APSO. The
final stage is classification with SVM. The suggested model
was executed on images from the MIAS database that has
both normal and abnormal (332). According to the findings,
the proposed APSO algorithm outperforms the existing
method in classification accuracy. The proposed APSO
algorithm reduces time complexity and selects the most
important features. The executed approach yields a good
compromise between the number of chosen characteristics
and classification accuracy. Future research will be focused
on swarm intelligence and also to compare with other existing
cancer diagnosis methods. Early-stage recovery and
treatment of breast cancer benefit greatly from accurate and
effective diagnosis in internet of things (IoT) healthcare.

REFERENCES

[1] F. Ibrahima and M. E. Mohamed, “An optimized feature selection

method for breast cancer diagnosis in digital mammogram using
multiresolution representation,” International Journal of Applied
Mathematics and Information Sciences, vol. 8, no. 6, pp. 2921-2928,
2014.
[2] B. L. Prakash, and M. Thirumalesh, “Multilevel wavelet transformation
based 3d-breast mammographic image classification”, International
Journal of Engineering Research and Technology, vol.2, no.12, pp.
888-893, 2013.
[3] Ismail A. Ismail, Essam A. Rashed, and Sherif I. Zaki, “Multiresolution
mammogram analysis in multilevel decomposition,” Expert Systems
with Applications, pp. 286-292, 2007.
[4] I. Faye, M. M. M. Eltoukhy, and Brahim Belhaouari Samir, “Using
curvelet transform to detect breast cancer in digital
mammogram,” 2009 5th International Colloquium on Signal
Processing & Its Applications, Kuala Lumpur, Malaysia, pp. 340-345,
2009.
[5] B. B. Samir, M. Meselhy Eltoukhy, and I. Faye, “A Comparision of
Wavelet and Curvelet for Breast Cancer Diagnosis in Digital
Mammogram,” Expert Systems with Applications, pp. 384-391, 2010.
[6] S. Tamil Selvi, J. Dheeba, and N. Albert Singh, “Computer-Aided
Detection of Breast Cancer on Mammograms: A Swarm Intelligence
Optimized Wavelet Neural Network Approach”, Journal of biomedical
informatics, vol. 49, pp. 45-52, 2014.

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on September 20,2023 at 02:41:52 UTC from IEEE Xplore. Restrictions apply.