Machine Learning-Based Approaches For Breast

Cancer Detection in Microwave Imaging

Humza Sami, Mahnoor Sagheer, Kashif Riaz, Muhammad Qasim Mehmood, and Muhammad Zubair
Department of Electrical Engineering
Informationt Technology University
Lahore, Pakistan
email: {humza.sami, mahnoor.sagheer, kashif.riaz, qasim.mehmood, muhammad.zubair}@itu.edu.pk

Abstract — Detection of breast cancer at an earlyhumza.sami@itu.edu.pk

stage can the breast into a binary class classification problem. Among
significantly reduce the mortality rate. Microwave imaging is a these algorithms, the support vector machine (SVM) is one of
promising detection tool for harmless and non-ionizing the famous classification algorithms for binary and multi-
screening of breast cancer. In this work, a fast and accurate class problems. In this work, SVM is investigated using linear
machine learning algorithm is proposed for the prediction of the
breast lesion using microwave signals. Machine learning has
and high degree polynomial kernels.
proved itself reliable in the field of biomedical application where II. METHODOLOGY
the diagnosis of the disease is desired. The support vector
machine (SVM) algorithm with the linear and polynomial Since transmitted and backscattered signals are
kernel is trained and tested on raw backscattered signals data. considered to be nonlinear in space. The kernel is a mapping
SVM with third-degree polynomial kernel obtained 99.7% function that takes an inseparable data and maps it to a high
accuracy that outperforms the existing conventional machine dimensional feature space in which data is separable. The
learning binary classification algorithms. Thus, the prediction SVM with polynomial kernels classify the nonlinear data in
of tumor presence would help the radiologist to diagnose tumor
high dimensional feature space by maximizing the gap of two
correctly at early stages.
classes. The main objective of the classifier is to separate the
Keywords— Breast cancer, Support vector machine, given N-dimensional data into the correct classes by finding
Machine learning, Kernels a hyperplane. SVM proves itself a computationally fast and
optimal classifier for the prediction of breast lesions. In this
I. INTRODUCTION work, open-source datasets generated by the University of
In the last two decades, microwave imaging has Manitoba are used for the training of the model [7]. A
been demonstrated as a safe, non-ionizing, and cost-efficient working flow graph of our approach is shown in Fig. 1.
solution in comparison to other state-of-the-art imaging Here, a total of 1008 data examples are used for
modalities in breast cancer screening [1]. Recently, many testing and training the model. From 1008, 800 examples are
machine learning based approaches have been proposed for used for training the model and 208 examples are used for the
the improvement and liability of microwave imaging for cross-validation of the model. The dataset is available in the
breast cancer. Machine learning algorithms can be trained on .pickle format. The dataset contains 1008 examples of
available datasets to predict the hidden pattern and create a 1001x72 matrices. Here 1001 are the frequency points while
decision base model in medical applications. Breast cancer 72 are antenna locations. The data is preprocessed before the
detection is considered to be a binary class classification implementation of the algorithm. The dataset samples are
problem. Given the backscattered signals as a dataset, a consisting of frequency domain s-parameters (S21). These
machine learning algorithm can classify the sample breast frequency domain samples are then converted to time-domain
into a lesion breast (positive class) or a healthy breast signals using an inverse Fourier transform. After that absolute
(negative class) [2]. Several approaches have been proposed value is taken of each complex time domain signal sample.
for the prediction of malignant and benign tumors in [3-5] Then a time windowing filter is applied to make sure that only
either from obtained images or signals. Recently in 2019, signal response due to the phantom is used for the training of
different machine learning approaches have been tested on the model. For the minimization of the chance of overfitting,
unprocessed clinical data [6]. Clinical data were obtained each data sample has a different tumor size and position.
from Perugia Hospital, Italy, using a non-ionizing imaging
device, MammoWave. More than 500 training examples
were applied to train the models and a 98% accuracy was
achieved.
In this work, we have investigated the performance
of the ML algorithms on unprocessed experimental breast
phantom data. The implementation of ML algorithms can
help in deciding between non-healthy and the healthy breast
by learning the patterns of backscattered signals. Various ML
classifiers modalities have been proposed for pattern Fig. 1. Working flow graph of the proposed methodology
identifications and future predictions, but K-NNs, Linear
Discriminant Analysis, Naïve Bayes, Multi-layer perceptron, In the model testing phase, the model was tested on
and support vector machine are recognized to be efficient our own generated dataset. For the data generation, a 3D
classification algorithms. These algorithms are practice to printed breast phantom was used to mimic the electrical
transform qualitative backscattered microwave signals from properties of human breast. The details of measurement setup

978-1-946815-10-1/21/$31.00 ©2021 IEEE 72 URSI 2021

Authorized licensed use limited to: KDU - Gen. Sir John Kotelawala Defence University. Downloaded on January 16,2023 at 16:44:07 UTC from IEEE Xplore. Restrictions apply.
 and phantom are given in [8]. Two Vivaldi antennas were
used as transmitter and receiver. The operating bandwidth of
the system is 2-4 GHz. NI PXIe 5630 vector network analyzer
was used for the transmission and reception of signals.
Antennas were placed around a circle and phantom was
placed between them. Motors rotate the antennas in a way
that they scan complete circle. Two scans were measured to
remove the noise and signal clutter. The second scan was
done with some offset angle. After that, the first scan was
subtracted from the second scan to remove the artifacts.
III. RESULTS AND DISCUSSION Fig.2. Actual data distribution
After the model training, the model was tested with Due to the coherence in signals data, we can train our
our measured dataset. The performance of SVM was model even with 200 training samples data. These results are
measured using linear and polynomial kernels. Further on, encouraging that the SVM algorithm with polynomial kernels
feature extraction and dimensionality reduction techniques gives promising predictions to identify lesions and healthy
were also applied to analyze the model. The Principal breasts.
Component Analysis (PCA) algorithm was used for
dimensionality reduction. At first, the SVM algorithm was IV. CONCLUSION
trained on the given data. Then the best performing kernel In this work, the SVM algorithm with different
was applied to reduced data. Meanwhile, the specificity and kernels has been implemented on raw backscattered signals
sensitivity of SVM were also analyzed for the given dataset. data. An SVM model is trained using 25-100% training data.
In the polynomial kernel, 2nd, and 3rd-degree polynomials The experimental results show that SVM with a polynomial
were tested for training. The SVM model was trained using kernel with a 3rd degree provides above 99% accuracy.
25% to 100% data. The associated classification performance Furthermore, due to the coherence present in the
is mentioned in table 1. The best performing hyperparameters backscattered signals, only 25% of data can also produce a
obtained are, learning parameter = 0.01 and soft margin = good enough prediction with 94% accuracy. A successful
3500. SVM with polynomial kernel has shown 99.7% implementation of machine learning approaches on breast
accuracy on K-fold cross-validation. A comprehensive cancer data can help radiologists to diagnose tumors
comparison of recent work is shown in table 2. correctly.
TABLE 1. SVM POLYNOMIAL KERNEL RESULTS ACKNOWLEDGMENTS
Kernel Data (%) Accuracy Specificity Sensitivity This work was supported in part by the Higher Education
25 0.9356 0.945 0.902 Commission (HEC), Pakistan under Grant TDF-03-160.
50 0.980 0.988 0.970
𝑺𝑽𝑴𝑷𝒐𝒍𝒚
75 0.985 0.990 0.980 REFERENCES
100 0.997 0.999 0.992
[1] L. Wang, “Early diagnosis of breast cancer,” Sensors, vol. 17, no. 7,
pp. 1572, 2017.
TABLE 2. COMPARISON WITH RECENT WORK
[2] T. Reimer, J. Sacristan and S. Pistorius, “Improving the diagnostic
Ref Algorithm Data Type Accuracy (%) capability of microwave radar imaging systems using machine
[3] LDA Raw signals 87.7 learning,” 2019 13th European Conference on Antennas and
[4] CNN Images 93.44 Propagation (EuCAP), Krakow, Poland, pp. 1-5, 2019.
[5] 𝑆𝑉𝑀!"#$%& Raw signals 98 [3] R. C. Conceição, H. Medeiros, M. O'Halloran, D. Rodriguez-Herrera,
D. Flores-Tapia and S. Pistorius, “SVM-based classification of breast
tumour phantoms using a UWB radar prototype system,” 2014 XXXIth
TABLE 3. PCA ANALYSIS RESULTS URSI General Assembly and Scientific Symposium (URSI GASS),
Beijing, pp. 1-4, 2014.
Kernel Data dimension Accuracy
1001 x 60 0.960 [4] V. Oleksyuk, F. Saleheen, D. F. Caroline, S. A. Pascarella and C. Won,
1001 x 54 0.954 “Classification of breast masses using Tactile Imaging System and
𝑺𝑽𝑴𝑷𝒐𝒍𝒚 machine learning algorithms,” 2016 IEEE Signal Processing in
1001 x 35 0.952
Medicine and Biology Symposium (SPMB), Philadelphia, PA, pp. 1-4,
1001 x 18 0.940
2016.
[5] B. Gerazov and R. C. Conceicao, “Deep learning for tumour
The SVM 3rd degree polynomial kernel provided the best classification in homogeneous breast tissue in medical microwave
imaging,” IEEE EUROCON 2017 -17th International Conference on
testing accuracy of 99.7% on given data. Since the original Smart Technologies, Ohrid, pp. 564-569, 2017.
data had 72 feature vectors, so PCA was applied to reduce its [6] S. P. Rana, M. Dey, G. Tiberi et al. “Machine learning approaches for
dimensionality. A comprehensive analysis is done on reduced automated lesion detection in microwave breast imaging clinical data,”
data in descending order and results are reported in table 3. Sci Rep, vol. 9, pp. 10510, 2019.
Since data is a low rank due to the coherence in time-series [7] “Breast Cancer Dataset,” UM-BMID - Google Drive. [Online].
signals. Therefore, the reported algorithm has given 94% Available: https://bit.ly/UM-bmid. [Accessed: 23-Jun-2020].
accuracy when it was applied to reduced dataset. Fig. 2 [8] H. Sami, M. Sagheer, M. A. Altaf, J. Iqbal and M. Zubair,
depicts the data distribution in the reduced two-dimensional “Development of prototype microwave radar-based imaging system for
breast cancer detection”, Journal of Engineering and Applied Sciences,
feature space.
vol. 39, no. 2, pp. 170-179, 2021.

73

Authorized licensed use limited to: KDU - Gen. Sir John Kotelawala Defence University. Downloaded on January 16,2023 at 16:44:07 UTC from IEEE Xplore. Restrictions apply.