IRJET - A Conceptual Method for Breast Tumor Classiﬁcation using SHAP Values and Adaboost

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 02 | Feb 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 2985
A Conceptual Method for Breast Tumor Classification using SHAP
Values and Adaboost
Sulthana Rinsy A. P.1, Anish Kumar B.2
1Dept. of Computer Science and Engineering, MEA Engineering College, Kerala, India
2Assistant Professor, Dept. of Computer Science and Engineering, MEA Engineering College, Kerala, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Breast cancer is one of the most tremendous
cancers among women worldwide. It is the most common, and
the leading cause of cancer related deaths among women
between 20 to 59 years of age. This is an attempt aiming to
assist clinicians in improving the accuracy of diagnostic
decisions by classifying the breast cancer as benign and
malignant tumors from ultrasound using computer-aided
diagnosis (CAD) system with human-in-loop. In this method,
feature acquisition is performed on the basis of Breast
Imaging Reporting and Data System (BI-RADS) lexicon and
experience of doctors by an user-participated feature scoring
scheme. The classification is done by combining SHAP value
mining and Adaboost algorithm. SHAP value based mining is
done because it can form meaningful pattern clusters on the
training data. The patterns frequently appearing within the
tumors with an equivalent label are often considered a
possible diagnostic rule. Subsequently, thediagnosticrulesare
utilized to construct component classifiers of the Adaboost
algorithm via a completely unique rules combinationstrategy.
Finally, the Adaboost learning is performed to discover
effective combinations and integrate them into a strong
classifier. The experimental results show that the proposed
method yielded the simplest prediction performance,
indicating an honest potential in clinical applications.
Key Words: BI-RADS lexicon, SHAP Value mining,
Adaboost
1. INTRODUCTION
Breast cancer is the one of the most commoncancer
among women worldwide and about 2.1 million women are
undergoingtreatmenton breastcancerworldwideaccording
to global estimates of cancer 2018 [1]. Global burden of
cancer worldwide using GLOBOCAN 2018estimatedthat the
leading cause of cancer death in women is breast cancer
whereas in men that is lungcancer.Accordingtotheirsurvey
the death rate due to breast cancer is getting in the peak
rapidly. However the early detection and diagnosis can
reduce the rapid growth in mortality and thereby can
improve the survival rate. Breast cancer can occur both in
men and women, but it’s far more common in women.
Finding carcinoma early and getting state-of-the-art cancer
treatment are the foremost important strategies to stop
deaths from carcinoma. Breast cancer that’s found early,
when it’s small and has not spread, is simpler to treat
successfully. Getting regular screening tests is the most
reliable thanks to find carcinoma early. After detecting the
cancer the key challenge faced by clinicians and doctors are
in classifying the cancer into benign and malignant where
machine learning techniques can play a vital role in the
classification of tumor by applying proper classification
algorithms.
Different tests are often wont to search for and
diagnose carcinoma. Number of imaging technologies have
been demonstrated to be of great help to early diagnosis for
breast cancer [2] [3]. Mammography is the most commonly
used screening method for breast cancer in early stage.
However mammography has some limitations, not all
breasts look the same on a mammogram, a woman’s age or
breast density can make cancers more or less difficulttosee.
In general, screening mammograms are less effective in
younger women because they tend to have dense breast
tissue and also radiation from mammography does harm to
the patient’s body and can significantly increase the risk of
breast cancer [4].
Ultrasonography has become a popular alternative
to mammography in clinical practice. Comparatively
speaking, ultrasonography has the advantages of beingnon-
radioactive, non-invasive, low cost and more convenient in
practice [5] [6]. In addition, ultrasonography is not only
more sensitive to dense breast tissues, it also has higher
accuracy in discriminating malignant and benign tumors.
Breast Imaging Reporting and Data System (BI-
RADS) is another helpful tool frequently used in clinical
practice [7]. The system is developed to standardize the
reporting of characteristics descriptions in mammography,
ultrasound or MRI, so as to promote communication among
clinicians. However, there is still a high misdiagnosis rate in
the clinical application due tothesubjectivedependenceand
experience variation among clinicians. Hence, computer
aided diagnosis (CAD) system has important research value
in helping clinicians improve the accuracy of diagnosis of
breast tumors.
2. RELATED WORK
In recent years the research in the area of analysis
and classification in machine learning plays a vital role. For
the tumor detection and analysis there are various methods
used in the recent studies [8], for the breast tumor
classification a large number of CAD approaches have been
proposed in recentyearsDifferentCADsystemusesdifferent
methods for classification such as some of the system uses

SVM [9] [10] and some of the system used robust phase
based texture descriptor[11]andsomeusedfuzzycerebellar
model neural network [12] which classify tumor through a
learning mechanism to imitate the cerebellum of human
being and some uses CNN [13]. A CAD systemusedSVM with
28 features in the ultrasound image and achieved a high
accuracy of 94.3% in the classification of tumor as benign
and malignant [14]. For automatically detecting the tumor a
CAD system used fuzzy SVM and regressionfeatureselection
[15]. For efficient use of the principle component analysis
and image retrieval a 3D systemforbreastnodulesdiagnosis
is proposed in [16]. Some studies used a clustering method
and affinity propagation for identifying breast tumors as
benign and malignant.
Recently almost all systems have been designed to
work in an automated way for both feature extraction and
classification, However in some cases that will lead to
complicating the situation by providingfalsepositiveresults.
Feature extraction and selection are the two important
stages in the classification even in the image processing and
pattern classification on [17] [18] where poor quality image
leads to unstable performance.Thereforethehuman-in-loop
CAD system performs better than that a fully automated
system.
3. PROPOSED METHOD
Before going to the technical detailsoftheproposed
conceptual framework it might be useful knowing about
various types and characteristics of breast cancer. There are
various types of breast carcinoma, according to WHO
classification the breast carcinoma falls into fifteen
categories and when diagnosing with different breasttumor
each shows distinct image characteristics in BUS images. By
using BI-RADS lexicon which covers all these characteristics
clinician can analysethebreasttumor.HoweverBI-RADS has
a limitation that it will only consider a few features of a
particular breast carcinoma and it will not show a distinct
pattern for other features. That is it considersa subsetoffew
features for diagnosing the tumor which resulting BI-RADS
feature will not be all consistent for different types of
carcinoma in practice, in contrast this limitation leading to
reduce the use of BI-RADS features for classification as an
accurate one.
The framework here used for classifying the breast
carcinoma is a CAD method which includes BI- RADS
features along with human- in- the loop strategy for
improving the classification accuracy. It has diagnostic rule
mining and ensemble learning algorithm. There is a lot of
feature extraction and learning algorithm in machine
learning, from that we extracted the proper mining method
and learning algorithm by analyzing and testing different
methods along with a human-in-the-loop. Weproposea CAD
framework involving diagnostic rule mining and ensemble
learning algorithm along with the involvement of human as
an operator. It became differentfroma typical CADsystemin
the case of feature extraction, where an operator also
involved for rating the feature along with BIRADS lexicon.
This framework uses both the prepossessed trained data
along with knowledgeand experienceofdoctor whichmakes
this system dominant over a typical CAD system.
Usually clinicians predict somefindingsonthebasis
of some features not considers all the features, likewisehere
for extracting such subset of features as diagnostic rule we
propose the mining method using the shapely addictive
explanation (SHAP) values. SHAP values can be applied to
the rated feature data from that we can predict the high risk
group malignant and low risk group benign. The SHAP
values can be applied at global and local level, from that we
can predict how much the contribution of each feature on
the final prediction on average. It helps us to analyze and
interpret our model intuitively and creates clusters for
benign and malignant diagnostic rules by extracting the
pattern from each feature.
Our motivation is to use SHAP method sufficiently
for diagnostic rule mining and use an efficient ensemble
learning algorithm. Adaboost algorithm is a relevant and an
appropriate learning algorithm in ensemble learning. This
Adaboost algorithm is implanted to make SHAP based
classifier together into a strong robustclassifier.Thenotable
feature of this method is data mining using SHAP values
along with an operator and also in an automated way and
hence good interpretability of the final model. Those SHAP
based rules used to build the ensemble classifier would help
understand the relevant clinical manifestation related to
breast tumors in ultrasound.
3.1 System design
In this section, we describe the novel SHAP based
ensemble learning method. The workflow of the proposed
method is given in Figure 1, we start by how the SHAP value
can be applied for discovering diagnostic rulesfromthedata
matrix obtained after feature scoring.Afterthata conceptual
novel strategy is applied to build component classifier and
finally the Adaboost algorithm is briefly described.
3.2 Data preparation
Machine learning techniques train and test the data
before implementing, for that first we have to collect and
rate the data using BI-RADS lexicon with the help of a
human-in-loop CAD system.

Figure 1: Workflow of proposed system
3.3 BI-RADS feature scoring scheme
BI-RADS feature scoring was widely used in
mammograms and also in recent ultrasound images.
Generally clinicians guess and classifya breasttumorusinga
BI-RADS descriptors on the basis of clinician or doctor
experience. BI-RADS lexicon is a tool comprising all the
necessary features such as shape, orientation, margin, echo
pattern, posterior acoustic characteristics, calcification etc.
Here in this research we have considered 24 features from
that we have concluded that 12 features are more
convenient where the other 12featuresaremorecommonin
both benign and malignant cases. These 12 important
features show different behavioural characteristics in
patients with different tumors.These12featuresforma new
feature subset for showing malignant and benign results.
Here the rating features can be done using BI-RADS
reference table which rates the value from 0 to 4. Higher the
value more it inclined towards malignant otherwise benign.
The clinician rate the feature after getting the BUS image on
the reference table. For showing malignant benign it shows
the values 1 and 0 respectively. Instead of image processing
and feature extraction method this proposedmethodadopts
a straightforward method, because some image processing
methods show some limitations such as highly sensitive to
the quality of breast ultrasound imageandthetypeofnoises.
BI-RADS feature scoring scheme helps the clinicians for
analysing and interpreting the diagnostic rule in an efficient
way. From the perspective of a doctor such a system would
be more convenient and can be applied in real time medical
application than a fully automated system with complex
image processing and feature extraction tasks.
3.4 Mining of diagnostic rule using SHAP values
For applying SHAP value first fit the data into a
complex model like Xgboost. It is done for getting the SHAP
values for each example in the data and then cluster them to
find the pattern. This can lead to finding the
intervention of diagnostic rules of malignant and benign
cancers. SHAP values can createmeaningful clustersbecause
SHAP values for all features are on the same scale (log odds
for binary Xgboost).
3.5 The classification of different feature spaces
For performing the ensemble learning a similarity
classifier containing both malignant and benign rules needs
to be constructed, for constructing such a weak classifier a
novel combination strategy is used. After separating the
benign and malignant rule anyone from benign rule and
anyone from malignant rule can be matched together into a
weak classifier that works on the basisofsimilarityprinciple
for classifying the new tumor. In this researchforfindingthe
distance measure between different feature spaces, Feature
space dependent normalized distance which is calculated
using the following Equation 1 for evaluating the similarity
of a test instance and diagnostic rule.
Equation 1
Where VT and VR are stands for vectors of test instance and
diagnostic rule respectively. V1 and V0 are the maximum and
minimum vector, the denominator in the Equation 1 plays a
normalization in the feature space of diagnosticrule.Totake
a decision on tumor, FSDNDs of the test instance malignant
and benign rules are measured and for smaller FSDNDmore
similar they are. With the FSDND a classifier can be created
from a pair of benign and malignant rule by SHAP method
using the Equation 2. Let Rm and Rb be the malignant and
benign rules respectively and x be the new tumor instance
then as shown below the predicted class of such a similarity
classifier S(x) is the same as the attribution of the rule with
smaller FSDND.

Equation 2
3.6 Ensemble learning
Ensemble learning is a method which can comprise
multiple models such as classifier to solve a particular
computation problem. Adaboost is one of the important and
relevant algorithms in ensemble learning algorithmsusedin
machine learning. It can improve the accuracy of
classification by combining multiple weak classifiers. Here,
similarity classifiers are treated as weak classifier. In this
method, on various training examples the classifier should
be trained interactively. In each iteration it tries tominimize
the training error by providing an excellent fit for those
examples. The weight of those correctly classified instances
become lower whereas theother becomehigher. Aftertheall
iteration all thecomponentclassifiersarecombinedtogether
to get the final hypothesis.
4. EXPERIMENTS
In this section the effectiveness of tumor
classification based on SHAP values and Adaboostalgorithm
is investigated. First the parameter setup and experiment
procedure are introduced. Then the evaluation metrics of
each experiment is performed. Here 150 malignant tumor
affected patient feature scored data set and 100 benign
tumor affected patient feature scored data set is collected
and used to train and test the system. The operation is
performed in anaconda Spyder python 3.4 version. Afterthe
testing the performance is evaluated in terms of accuracy,
specificity and sensitivity. The performance comparison of
different classifiers are given below in the Table 1.
Evaluation indices are shown in the Table 2.
Table 1: COMPARISON RESULT AMONG DIFFERENT CAD
SYSTEMS
Classifier Accuracy sensitivity specificity
SVM 94.4% 94.3% 94.4%
Fuzzy SVM 94.25% 91.67% 96.08%
Fuzzy cerebellar
model NN
92.31% 93.55% 91.18%
Judgement by
experience
85.62% 93.8% 72.97%
SHAP value +
Adaboost
96.41% 96.72% 95.75%
Table 2: EVALUATION INDICES
Accuracy (TP+TN)/(TP+TN+FP+FN)
Sensitivity (TP)/(TP+FN)
Specificity (TN)/(TN+FP)
5. CONCLUSION
In this paper a conceptual human-in-loop
framework using SHAP values and Adaboost algorithm is
implemented for classifying benign and malignant breast
tumors. It is an innovative method whichadoptsanoperator
based feature scoring scheme rather than fully automated
image processing and feature extraction methods. In
contrast this method introduces the experience of clinicians
during the feature extraction, which is easily acceptable to
doctors in real application and that improvesthe robustness
of our system.
ACKNOWLEDGEMENT
The work was supported by Computer science and
engineering department of MEA Engineering College
Perithalmanna, Malappuram, India.
REFERENCES
[1] F. Bray, J. Ferlay, I. Soerjomataram, R. L. Siegel, L. A.
Torre, and A. Jemal, “Global cancer statistics 2018:
Globocan estimates of incidence and mortality
worldwide for 36 cancers in 185countries,”CA:a cancer
journal for clinicians, vol. 68, no. 6, pp. 394–424, 2018.
[2] Z. Chen, H. Strange, A. Oliver, E. R. Denton, C. Boggis, and
R. Zwiggelaar, “Topological modeling and classification
of mammographic micro-calcification clusters,” IEEE
transactions on biomedical engineering, vol. 62, no. 4,
pp. 1203–1214, 2014.
[3] B. Al-Shargabi and F. Al-Shami, “An experimental study
for breast cancer prediction algorithms,”inProceedings
of the Second International Conference on Data Science,
E-Learning and Information Systems, 2019, pp. 1–6.
[4] T. Yin, F. H. Ali, and C. C. Reyes-Aldasoro, “A robust and
artifact resistant algorithm of ultrawideband imaging
system for breast cancer
detection,”IEEETransactionsonBiomedicalEngineering,v
ol.62,no.6, pp. 1514–1525, 2015.
[5] T. Ungi, G. Gauvin, A. Lasso, C. T. Yeo, P. Pezeshki, T.
Vaughan, K. Carter, J. Rudan, C. J. Engel, and G.
Fichtinger, “Navigated breast tumor excision using
electromagnetically tracked ultrasound and surgical
instruments,” IEEE Transactions on Biomedical
Engineering, vol. 63, no. 3, pp. 600–606, 2015.

[6] M. Costantini, P. Belli, R. Lombardi, G. Franceschini, A.
Mulè, and L. Bonomo, “Characterization of solid breast
masses: use of the sonographic breastimagingreporting
and data system lexicon,” Journal of ultrasound in
medicine, vol. 25, no. 5, pp. 649–659, 2006.
[7] S. Ackerman, “Acr bi-rads for breast ultrasound,”
Ultrasound in Medicine and Biology, vol. 41, no. 4, p.
S95, 2015.
[8] S. Zhou, J. Shi, J. Zhu, Y. Cai, and R. Wang, “Shearlet-based
texture feature extraction for classification of breast
tumor in ultrasound image,” Biomedical Signal
Processing and Control, vol. 8, no. 6, pp. 688–696, 2013.
[9] Q. Huang, F. Yang, L. Liu, and X. Li, “Automatic
segmentation of breast lesions for interaction in
ultrasonic computer-aided diagnosis,” Information
Sciences, vol. 314, pp. 293–310, 2015.
[10] Q. Huang, Y. Chen, L. Liu, D. Tao, and X. Li,“Oncombining
biclustering mining and adaboost for breast tumor
classification,” IEEE Transactions on Knowledge and
Data Engineering, 2019
[11] L. Cai, X. Wang, Y. Wang, Y. Guo, J. Yu, and Y. Wang,
“Robust phasebased texture descriptorforclassification
of breast ultrasound images,” Biomedical engineering
online, vol. 14, no. 1, p. 26, 2015.
[12] C.-M. Lin, Y.-L. Hou, T.-Y. Chen, and K.-H. Chen, “Breast
nodules computer-aided diagnostic systemdesignusing
fuzzy cerebellar model neural networks,” IEEE
Transactions on Fuzzy Systems, vol. 22, no. 3, pp. 693–
699, 2013.
[13] R. Nicole, “of paper with only first word capitalized j.
name stand. abbrev.”
[14] Y.-L. Huang, K.-L. Wang, and D.-R. Chen, “Diagnosis of
breast tumors with ultrasonic texture analysis using
support vector machines,” Neural Computing &
Applications, vol. 15, no. 2, pp. 164–169, 2006.
[15] X. Shi, H.-D. Cheng, L. Hu, W. Ju, and J. Tian, “Detection
and classification of masses in breast ultrasound
images,” Digital signal processing,vol.20,no.3,pp.824–
836, 2010.
[16] Y.-L. Huang, S.-H. Lin, and D.-R. Chen, “Computer-aided
diagnosis applied to 3-d us of solid breast nodules by
using principal component analysis and image
retrieval,” in 2005 IEEE Engineering in Medicine and
Biology 27th Annual Conference. IEEE, 2006, pp. 1802–
1805.
[17] Y.Li,W.Liu,X.Li,Q.Huang,andX.Li,“Ga-
sift:Anewscaleinvariant feature transform for
multispectral image using geometric algebra,”
Information Sciences, vol. 281, pp. 559–572, 2014.
[18] C. Hou, F. Nie, X. Li, D. Yi, and Y. Wu, “Joint embedding
learning and sparse regression: A framework for
unsupervised feature selection,” IEEE Transactions on
Cybernetics, vol. 44, no. 6, pp. 793–804, 2013.

IRJET - A Conceptual Method for Breast Tumor Classiﬁcation using SHAP Values and Adaboost

Recommended

More Related Content

What's hot (19)

Similar to IRJET - A Conceptual Method for Breast Tumor Classiﬁcation using SHAP Values and Adaboost (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET - A Conceptual Method for Breast Tumor Classiﬁcation using SHAP Values and Adaboost