Automated Breast Cancer Detection Using Machine Learning Techniques by Extracting Different Feature Extracting Strategies

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

2018 17th IEEE International Conference On Trust, Security And Privacy In Computing And Communications/ 12th

IEEE International Conference On Big Data Science And Engineering

Automated Breast Cancer Detection using Machine


Learning Techniques by Extracting Different Feature
Extracting Strategies
Lal Hussain*, Wajid Aziz*, Sharjil Saeed*, Saima Rathore†, Muhammad Rafique††
*Department of Computer Science & IT, University of Azad Jammu and Kashmir, Muzaffarabad, Pakistan
†Department of Radiology, University of Pennsylvania, Philadelphia, PA 19104
††Department of Physics, University of Azad Jammu and Kashmir, Muzaffarabad, Pakistan
lall_hussain2008@live.com, kh_wajid@yahoo.com, rajasharjeel@gmail.com

Abstract—This Breast Cancer in women is the most frequency genetics and childhood nutrition is accounted for 1/3 of 6
diagnosed and second leading cause of cancer deaths. Due to differences in cancer risk [5], [6].
complex nature of microcalcification and masses, radiologist fail The early diagnosis and detection of breast cancer
to properly diagnose breast cancer. In past researchers can decrease the death rate and provide means for prompt
developed Computer aided diagnosis (CAD) systems that help the
radiologist to detect abnormalities in an efficient manner. In this
treatment. The breast cancer is diagnosed and detected using
research, we have employed robust Machine learning combination of approaches including imaging, physical
classification techniques such as Support vector machine (SVM) examination and biopsy [7]. The imaging techniques used to
kernels and Decision Tree to distinguish cancer mammograms detect the breast cancer are mammography and ultrasound. In
from normal subjects. Different features are proposed such as this method X-ray is used to create images of the breast. These
texture, morphological entropy based, scale invariant feature images are known as mammograms. Radiologists who are
transform (SIFT), and elliptic Fourier descriptors (EFDs). These trained in to read mammograms to detect sign of breast
features are passed as input to ML classifiers. Jack-knife 10-fold cancer. The effectively of screening process can rely on
cross validation was used and performance evaluated in term of radiologists’ explanation [8]. The patients affected by palpable
specificity, sensitivity, Positive predive value (PPV), negative
predictive value (NPV), false positive rate (FPR) and receive
breast cancer may have sonogram and mammogram
operating curve (ROC). The highest performance based on single examination with both normal and benign or nonspecific
feature extracting strategy was obtained using Bayesian appearance [9]. Biopsy is used to confirm the symptoms of
approach with texture and EFDs features, and SVM RBF and breast cancer, but it is invasive surgical operation causing
Gaussian kernels with EFDs features whereas highest AUC with psychological and physical impact on patients.
single feature was obtained using Bayesian approach by Moreover, the research evidences indicate that
extracting texture, morphological, EFDs and entropy features radiologists may miss up to 30 % of breast cancer depending
and SVM RBF and Gaussian kernels with EFDs features up on the density of breasts [10]. The mammograms in breast
cancer have been evaluated using two powerful indicators
I. INTRODUCTION known as masses and micro-calcifications. Mass detection is
more challenging than the micro-calcification, not due to the
Breast cancer is one of the most frequently diagnosed large variation in size and shape in which masses can appear
cancer in women worldwide. In developing countries, breast in mammogram but also because masses often exhibit poor
cancer accounts for 23% of the total cancer cases and 1.6 image contrast [11]. Radiologists read the mammograms
million new cases of the breast cancer are estimated based on judgement on experience, training and subjective
worldwide affecting the women [1]–[3] The breast cancer criteria. There may be 65-75% inter-observe variations rate
accounts for nearly one in three cancers among US women even by the trained experts [12] Thus, CAD may help the
excluding skin cancer and second leading cause of cancer radiologists to interpret the mammograms to detect and
death among women after lung cancer [4]. In 2016, about 29% classify masses. The literature also reveals that about 65-90%
deaths were accounted in female due to breast cancer in the of the biopsies of suspected cancers turned out to be benign,
United State. In 2016, it was estimated that 595,690 American thus it is very important to develop such techniques that can
will die from cancer corresponding to 1,600 deaths per day distinguish the malignant and benign lesions. The combination
[5]. The most common cause of cancer deaths are lung and of CAD, expert knowledge and Machine learning techniques
bronchus, prostate, and colorectum in men and lung and would greatly improve the detection accuracy. The detection
bronchus, breast, and colorectum in women. The invasive accuracy without CAD was obtained below 80% and with
cancer lifetime probability of being diagnosed in men (42%) is CAD above 90% [13].
higher than for women (38%) which may be due to some In the previous studies, the researchers extracted
external reflect differences in environment exposure, different features from colon biopsy images to detect the
endogenous hormones and complex interaction between these cancer and distinguish the normal and malignant subjects.
influences. The cancer incidence and deaths in both men and [14], [15] analyzed distinctiveness of six texture features
women are associated with adult height determined by

2324-9013/18/31.00 ©2018 IEEE 327


DOI 10.1109/TrustCom/BigDataSE.2018.00057

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 19,2022 at 04:07:42 UTC from IEEE Xplore. Restrictions apply.
(contrast, angular second moment, correlation, entropy, used mass instance images digitized by LUMYSIS. This
inverse difference moment, and dissimilarity) for database contains approximately 2,500 studies. We used the
classification of colon biopsy images and obtained overall latest volumes of DDMS database i.e. of Normal volume 12
accuracy of 90.2% [14]. [16] used probabilistic neural network and for that of Cancer volume 15 containing a total of 899
to diagnose the breast cancer by extracting multi-scale texture images including 500 cancer images having 105 cases and 399
properties of the tissue surrounding the micro-calcification normal subject images having 100 cases.
(MCs). Moreover, [17] proposed a colon biopsy image based
classification technique (CBIC) by extracting hybrid feature B. Methods
set consisting of traditional histogram of oriented gradients The proposed system consists of four stages namely
based features and variant of statistical moments and Haralick pre-processing, features extraction, classification, training/
texture features to classify the colon biopsy images by using testing of data, and to classify images into normal and
SVM classifier and obtained a 98.85% of classification malignant images using SVM, Decision Tree and Bayesian
accuracy. In the past few decades, different other approaches classifier. The extracted individual and combination of
have also been used to detect and diagnose the breast colon features are then passed as input to the Machine learning
images including probabilistic algorithm and radial gradient classifiers such as SVM polynomial, RBF and Gaussian
index based algorithm [18], Convolution Neural Network kernels, Bayesian classification approach and Decision Tree.
(CNN) classifier [19] and mixed feature based neural network The detailed algorithms are described in the classifiers section.
[20], computerized method for the automated segmentation of Finally, the training and target data split was made using
individual micro-calcifications in a region of interest (ROI) Jackknife 10-fold cross validation to classify the normal and
[21], artificial intelligent techniques, including fractal cancer subjects. In this study, we extracted morphological,
dimension analysis, with the discrete wavelet transform [20]. texture, entropy, SIFT and EFDs features as detailed in
In the past researchers employed different features Hussain et al. [26]. In this study, we have employed robust
extracting strategies such as texture, morphological, SIFT, Support Vector Machine and its kernel, Naïve Bayes and
EFDs and combination of other features [22]–[24], [24]–[27]. Decision Tree algorithms as detailed by Hussain et al. [26],
The existing techniques have some limitations, the graph based [32].
techniques are competitively expensive. The other CAD
techniques based on texture features exploited general texture III. RESULTS AND DISCUSSIONS
features for classification and fail to provide the background The classification performance was measured using
knowledge of morphological features. Moreover, most of the
different classifiers such as Bayesian approach, SVM with
information hidden in the masses and micro-calcifications of
linear, polynomial, RBF (Radial Base Function) and Gaussian
colon biopsy images was extracted using complexity based
sample entropy and wavelet entropy features [28], [29], [29]– kernels and Decision Tree. The performance was evaluated
[32]. The machine learning methods based on different features by extracting various features such as texture, morphological,
extracting strategies have limitations as different researchers SIFT, EFDs and Entropy based features as shown in table 3.
employed different features extracting methods, however, these The highest 100 % performance in term of specificity,
classifiers are not fine-tuned. In contract the deep learning sensitivity, PPV, NPV, TA, AUC was obtained using Bayes,
convolution neural network models with transfer learning SVM polynomial and SVM Gaussian with EFDs feature set,
approaches are fine-tuned to optimize the parameters by texture features using Bayes classifier and entropy and
minimizing the error. To cope up the problem, we have morphological features using SVM RBF followed by Bayes
employed following transfer learning approaches using Deep using morphological and entropy features reflecting TA
learning methods. Based on the limitations of existing features, (99.89% and 99.67%) respectively. The total accuracy
we extracted different features from Breast Cancer Images. obtained using Decision tree classifiers with EFDs
(TA=97.22%), Gaussian with texture feature as (TA=97.44%).
II. MATERIAL AND METHODS The total accuracy with SVM RBF kernel using texture
features was obtained as (TA=96.49%) and DT with Texture
A. Data Sets (TA=95.88%), Morphological (TA=95.88%) was obtained. A
Data Set were taken from publicly available database good total accuracy was obtained using SVM polynomial
provided by the University of South Florida [33] and used in kernel with texture features (TA=93.77%), Morphological
[34] available online at (TA=90.66%), SVM Gaussian with morphological features
(http://marathon.csee.usf.edu/Mammography/Database.html). (TA=90.66%), and Decision Tree with SIFT features
The Digital Database for Screening Mammography (DDMS) (TA=91.70%). The sensitivity, specificity, PPV, NPV and
database contains the mammographic images for research FPR are obtained accordingly using different classifiers and
purposes which is the Using BI-RADS Descriptors and set of features as reflected in Table 1. The PPV and NPV
Ensemble Learning for Classifying Masses 71 largest publicly denote the proportion of breast cancer with positive and
available resource for the mammogram analysis research negative results diagnosed correctly for each case respectively
community. In DDSM images, suspicious regions of interest and these results are depended on the prevalence of disease.
(ROIs; including masses and microcalcifications) are marked The findings also reveal that individual parameters could
by experienced radiologists, and BI-RADS information is also provide either good sensitivity or good specificity.
annotated for each abnormal region. In our experiment, we

328

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 19,2022 at 04:07:42 UTC from IEEE Xplore. Restrictions apply.
The AUC values using single feature and different Machine
learning classifiers are obtained as reflected in Table 1 and
Figure 2. The highest separation (AUC=100%) was obtained
using Bayesian classifier with texture, morphological, EFDs,
and Entropy. Moreover, SVM polynomial with EFDs, RBF
with morphological and entropy feature and SVM Gaussian
with EFDs features also exhibit highest (AUC=100%).
Likewise, the Decision Tree showed also highest
(AUC=100%) using texture, morphological, EFDs, SIFT and
entropy features followed by SVM Gaussian with texture
features (AUC=99.97%), SVM RBF with SIFT feature
(AUC=99.02%), and SVM polynomial with texture feature
(AUC=97.81%), and morphological feature (AUC=97.16%). (c)
Likewise, AUC values obtained using SVM polynomial with
entropy feature (AUC=95.26%). SVM Gaussian with
morphological feature give (AUC=96.52%), and with entropy
feature (AUC=94.20%) and SIFT (AUC=93.91%).
The performance was evaluated using receiver operating curve
(ROC) analysis. The proximity of a measured value to the
true value is indicated by accuracy measure. The proportion of
positive results are measured using sensitivity such as percent
of correctly identified cancer patients and specificity denote
the proportion of correctly identified normal subjects and
negative. The sensitivity in the ROC curve is measured by
plotting the function of (1-specificity) for different operating
points. The specificity/ sensitivity pair on ROC plot is denoted (d)
by each operating point corresponding to a decision threshold.

(a) (e)
Fig. 1. ROC Analysis based on Signle feature set using a) Texture b)
Morphological c) Entropy d) EFDs e) SIFT

(b)

329

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 19,2022 at 04:07:42 UTC from IEEE Xplore. Restrictions apply.
TABLE 1: CLASSIFICATION ON USING ML TECHNIQUES FOR INDIVIDUAL FEATURES EXTRACTING STRATEGY

FEATURE Sensitivity Specificity PPV NPV TA FPR AUC


Bayes
Texture 1 1 1 1 1 0 1
Morphologi 0.9989 0.9991 0.9989 0.9986 0.9989 0.0008 1
SIFT 0.7263 0.6816 0.7307 0.7359 0.7263 0.3184 0.8033
EFDs 1 1 1 1 1 0 1
Entropy 0.9967 0.9958 0.9967 0.9974 0.9967 0.0042 1
SVM Polynomial
Texture 0.9377 0.9316 0.9381 0.9406 0.9377 0.06844 0.9781
Morphologi 0.9066 0.8981 0.9072 0.9098 0.9066 0.1019 0.9716
SIFT 0.7977 0.7721 0.7988 0.8014 0.7977 0.2279 0.8693
EFDs 1 1 1 1 1 0 1
Entropy 0.8921 0.8795 0.8941 0.8992 0.8921 0.1205 0.9526
SVM RBF
Texture 0.7531 0.7265 0.7564 0.7607 0.7531 0.273 0.8389
Morphologi 1 1 1 1 1 0 1
SIFT 0.9649 0.9542 0.9667 0.9741 0.9649 0.04577 0.9902
EFDs 0.5717 0.479 0.5797 0.5822 0.5717 0.521 0.6897
Entropy 1 1 1 1 1 0 1
SVM Gaussian
Texture 0.9744 0.972 0.9745 0.9755 0.9744 0.0280 0.9977
Morphologi 0.911 0.9057 0.911 0.9113 0.911 0.0943 0.9652
SIFT 0.8596 0.8468 0.8596 0.8593 0.8596 0.1532 0.9391
EFDs 1 1 1 1 1 0 1
Entropy 0.8843 0.8722 0.8858 0.8899 0.8843 0.1278 0.942
Decision Tree
Texture 0.9588 0.955 0.959 0.9605 0.9588 0.0449 1
Morphologi 0.9588 0.9606 0.9592 0.9556 0.9588 0.03942 1
SIFT 0.917 0.9147 0.9171 0.913 0.917 0.08533 1
EFDs 0.9722 0.9712 0.9722 0.9717 0.9722 0.02877 1
Entropy 0.9455 0.9489 0.9466 0.9409 0.9455 0.05109 1

References
IV. CONCLUSION [1] M. H. Forouzanfar, K. J. Foreman, A. M. Delossantos, R.
In this research, the robust Machine learning classification Lozano, A. D. Lopez, C. J. L. Murray, and M. Naghavi,
techniques such as SVM kernels, Bayesian approach and “Breast and cervical cancer in 187 countries between 1980
Decision Tree are employed to distinguish the cancer and 2010: A systematic analysis,” Lancet, vol. 378, no.
mammograms from that of normal subjects. The masse 9801, pp. 1461–1484, 2011.
detection due to low image contrast and microcalsification due [2] A. Jemal, F. Bray, and J. Ferlay, “Global Cancer Statistics:
to large variation in size shape require multi-dimension 2011,” CA Cancer J Clin, vol. 49, no. 2, p. 1,33-64, 1999.
features extracting strategies to effectively distinguish the [3] J. Dheeba, N. Albert Singh, and S. Tamil Selvi, “Computer-
aided detection of breast cancer on mammograms: A swarm
cancer mammograms. Thus, to cope with this problem, we
intelligence optimized wavelet neural network approach,” J.
extracted different features based on texture, morphology, Biomed. Inform., vol. 49, pp. 45–52, 2014.
SIFT and EFDs. To distinguish the normal mammograms from [4] C. E. Desantis, S. A. Fedewa, A. G. Sauer, J. L. Kramer, R.
that of breast cancer, the robust machine learning classification A. Smith, and A. Jemal, “Breast Cancer Statistics , 2015 :
techniques such as SVM, Bayes approach and Decision tree are Convergence of Incidence Rates Between Black and White
developed in Matlab version 2015. The 10-fold cross validation Women,” A cancer J. Clin., vol. 66, no. 1, pp. 31–42, 2016.
was used to train and test the image database. The performance [5] R. L. Siegel, K. D. Miller, and A. Jemal, “Cancer statistics,”
was measured based on specificity, sensitivity, PPV, NPV, CA Cancer J Clin, vol. 66, no. 1, pp. 7–30, 2016.
FPR and AUC. The result revealed that different features [6] S. Wirén, C. Häggström, H. Ulmer, J. Manjer, T. Bjørge, G.
extracting strategies and classifiers effectively provide the Nagel, D. Johansen, G. Hallmans, A. Engeland, H. Concin,
highest performance to detect the breast cancer mammograms. H. Jonsson, R. Selmer, S. Tretli, T. Stocks, and P. Stattin,
“Pooled cohort study on height and risk of cancer and cancer

330

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 19,2022 at 04:07:42 UTC from IEEE Xplore. Restrictions apply.
death,” Cancer Causes Control, vol. 25, no. 2, pp. 151–159, 2001.
2014. [21] S. Rathore, M. Hussain, and A. Khan, “Automated colon
[7] A. A. Ardakani, A. Gharbali, and A. Mohammadi, cancer detection using hybrid of novel geometric features
“Classification of breast tumors using sonographic texture and some traditional features,” Comput. Biol. Med., vol. 65,
analysis,” J. Ultrasound Med., vol. 34, no. 2, pp. 225–231, pp. 279–296, 2015.
2015. [22] “1,2 * 1:,” no. October, pp. 1–11, 2011.
[8] B. L. Sprague, E. F. Conant, T. Onega, M. P. Garcia, E. F. [23] S. Rathore, M. Hussain, M. Aksam Iftikhar, and A. Jalil,
Beaber, S. D. Herschorn, C. D. Lehman, A. N. A. Tosteson, “Ensemble classification of colon biopsy images based on
R. Lacson, M. D. Schnall, D. Kontos, J. S. Haas, D. L. information rich hybrid features,” Comput. Biol. Med., vol.
Weaver, and W. E. Barlow, “Variation in Mammographic 47, no. 1, pp. 76–92, 2014.
Breast Density Assessments among Radiologists in Clinical [24] S. Rathore, M. Hussain, and A. Khan, “Automated colon
Practice: A Multicenter Observational Study,” Ann. Intern. cancer detection using hybrid of novel geometric features
Med., vol. 165, no. 7, pp. 457–464, 2016. and some traditional features,” Comput. Biol. Med., vol. 65,
[9] P. E. Freer, “Mammographic breast density: impact on no. March, pp. 279–296, 2015.
breast cancer risk and implications for screening.,” [25] S. Rathore, A. Iftikhar, A. Ali, M. Hussain, and A. Jalil,
Radiographics, vol. 35, no. 2, pp. 302–15, 2015. “Capture largest included circles: An approach for counting
[10] T. M. Kolb, J. Lichy, and J. H. Newhouse, “Comparison of red blood cells,” Commun. Comput. Inf. Sci., vol. 281 CCIS,
the Performance of Screening Mammography, Physical pp. 373–384, 2012.
Examination, and Breast US and Evaluation of Factors that [26] L. Hussain, A. Ahmed, S. Saeed, S. Rathore, and I. Ahmed,
Influence Them: An Analysis of 27,825 Patient “Prostate cancer detection using machine learning
Evaluations,” Radiology, vol. 225, no. 1, pp. 165–175, 2002. techniques by employing combination of features extracting
[11] H. D. Cheng, X. J. Shi, R. Min, L. M. Hu, X. P. Cai, and H. strategies,” Cancer Biomarkers, vol. 21, no. 2, pp. 393–413,
N. Du, “Approaches for automated detection and 2018.
classification of masses in mammograms,” Pattern [27] Y. Asim, B. Raza, A. Kamran, and M. Saima, “A multi-
Recognit., vol. 39, no. 4, pp. 646–668, 2006. modal , multi-atlas-based approach for Alzheimer detection
[12] P. Skaane and K. Engedal, “Analysis of sonographic via machine learning,” Int. J. Imaging Syst. Technol., pp. 1–
features in the differentiation of fibroadenoma and invasive 11, 2018.
ductal carcinoma. AJR,” Am. J. Roentgenol., vol. 170, no. 1, [28] L. Hussain, W. Aziz, S. Saeed, S. A. Shah, M. S. A.
pp. 109–114, 1998. Nadeem, A. Awan, A. Abbas, A. Majid, S. Zaki, and H.
[13] K. Doi, D. Ph, and K. Rossmann, “Computer- Aided Kazmi, “Complexity analysis of EEG motor movement with
Diagnosis : Potential Usefulness in : Diagnostic Radiology eye open and close subjects using multiscale permutation
and Telemedicine,” in Research, Practice, and entropy ( MPE ) technique .,” Biomed. Res., vol. 28, no. 16,
Opportunities., Proceedings of the National Forum, 1996, pp. 7104–7111, 2017.
pp. 9–13. [29] L. Hussain, S. Saeed, I. A. Awan, and A. Idris, “Multiscaled
[14] A. N. Esgiar, R. N. Naguib, B. S. Sharif, M. K. Bennett, and Complexity Analysis of EEG Epileptic Seizure Using
A. Murray, “Microscopic image analysis for quantitative Entropy-Based Techniques,” Arch. Neurosci., vol. 5, no. 1,
measurement and feature identification of normal and pp. 1–11, 2018.
cancerous colonic mucosa.,” IEEE Trans. Inf. Technol. [30] L. Hussain, W. Aziz, S. Saeed, S. A. Shah, M. S. A.
Biomed., vol. 2, no. 3, pp. 197–203, 1998. Nadeem, I. A. Awan, A. Abbas, A. Majid, and S. Z. H.
[15] A. N. Esgiar, R. N. G. Naguib, B. S. Sharif, M. K. Bennett, Kazmi, “Quantifying the dynamics of
and A. Murray, “Fractal analysis in the detection of colonic electroencephalographic (EEG) signals to distinguish
cancer images.,” IEEE Trans. Inf. Technol. Biomed., vol. 6, alcoholic and non-alcoholic subjects using an MSE based K-
no. 1, pp. 54–8, 2002. d tree algorithm,” Biomed. Eng. / Biomed. Tech., vol. 0, no.
[16] A. N. Karahaliou, I. S. Boniatis, S. G. Skiadopoulos, F. N. 0, 2017.
Sakellaropoulos, N. S. Arikidis, L. E. A. Likaki, G. S. [31] L. Hussain, W. Aziz, J. S. Alowibdi, N. Habib, M. Rafique,
Panayiotakis, and L. I. Costaridou, “Breast cancer diagnosis: S. Saeed, and S. Z. H. Kazmi, “Symbolic time series
Analyzing texture of tissue surrounding analysis of electroencephalographic (EEG) epileptic seizure
microcalcifications,” IEEE Trans. Inf. Technol. Biomed., and brain dynamics with eye-open and eye-closed subjects
vol. 12, no. 6, pp. 731–738, 2008. during resting states,” J. Physiol. Anthropol., vol. 36, no. 1,
[17] S. Rathore, M. Hussain, A. Ali, and A. Khan, “A recent p. 21, 2017.
survey on colon cancer detection techniques.,” IEEE/ACM [32] L. Hussain, “Detecting epileptic seizure with different
Trans. Comput. Biol. Bioinform., vol. 10, no. 3, pp. 545–63, feature extracting strategies using robust machine learning
2013. classification techniques by applying advance parameter
[18] M. A. Kupinski and M. L. Giger, “Automated seeded lesion optimization approach,” Cogn. Neurodyn., vol. 12, no. 3, pp.
segmentation on digital mammograms,” IEEE Trans. Med. 271–294, 2018.
Imaging, vol. 17, no. 4, pp. 510–517, 1998. [33] M. Heath, B. Kevin, K. Daniel, K. P., M. Richard, and C.
[19] B. Sahiner, H. P. Chan, N. Petrick, M. A. Helvie, and L. M. Kyong, “Current status of the digital database for screening
Hadjiiski, “Improvement of mammographic mass mammography,” in In Digital mammography, 1998, pp.
characterization using spiculation measures and 457–460.
morphological features,” Med. Phys., vol. 28, no. 7, pp. [34] Y. Zhang, N. Tomuro, J. Furst, and D. S. Raicu, “Using BI-
1455–1465, 2001. RADS descriptors and ensemble learning for classifying
[20] L. Zheng and A. K. Chan, “An artificial intelligent masses in mammograms,” Lect. Notes Comput. Sci.
algorithm for tumor detection in screening mammogram,” (including Subser. Lect. Notes Artif. Intell. Lect. Notes
IEEE Trans. Med. Imaging, vol. 20, no. 7, pp. 559–567, Bioinformatics), vol. 5853 LNCS, pp. 69–76, 2010.

331

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on May 19,2022 at 04:07:42 UTC from IEEE Xplore. Restrictions apply.

You might also like