Original Article Open Access
DOI: 10.19187/abc.20207122-28
Accurate Detection of Breast Cancer Metastasis Using a Hybrid
Model of Artificial Intelligence Algorithm
a b c a b
Jafar Abdollahi , Atlas Keshandehghan , Mahsa Gardaneh , Yasin Panahi* , Mossa Gardaneh
a
Deputy of Research and Technology, Ardabil University of Medical Sciences, Ardabil, Iran
b
Department of Stem Cells and Regenerative Medicine, National Institute of Genetic Engineering and Biotechnology,
Tehran, Iran
c
Department of Biology, York University, Toronto, Canada
ARTICLE INFO ABSTRACT
Received: Background: Breast cancer (BC) is a prevalent disease and a major cause of
15 December 2019 mortality among women worldwide. A substantial number of BC patients
Revised:
experience metastasis which in turn leads to treatment failure and death. The
08 January 2020
Accepted: survival rate has been significantly increased due to more rapid detection and
18 January 2020 substantial improvements in adjuvant therapies including newer
chemotherapeutic and targeted agents, and better radiotherapy techniques.
Methods: In this study, we cross-compared the application of advanced
artificial intelligence algorithms such as Logistic Regression, K-Nearest
Neighbors, Discrete Cosine Transform, Random Forest Classifier, Support Vector
Machines, Multilayer Perceptron, and Ensemble to diagnose BC metastasis. We
further combined MLP with genetic algorithm (GA) as a hybrid method of
intelligent analysis. The core data we used for comparison belonged to the images
of both benign and malignant tumors collected from Wisconsin Breast Cancer
dataset from the UCI repository.
Results: The application of several different algorithms to the collection of BC
data indicated that these algorithms have comparable accuracy rate in detecting
Key words: and predicting cancer. However, our hybrid algorithm showed superior accuracy,
Breast cancer,
metastasis,
sensitivity and specificity compared to the individual algorithms. Two methods of
machine learning, comparison (Cross-Validation and Holdout) were applied to this study which
specificity selection, produced consistent results.
classification algorithms, Conclusion: Our findings indicate that our MLP-GA hybrid algorithm can speed
hybrid algorithm.
up diagnosis with higher accuracy rate than the individual patterns of algorithm.
Introduction expected to be diagnosed with invasive BC with
Breast cancer (BC) is one of the most prevalent 62,930 women within situ BC in the United States.
1
cancer types among human populations with one out The mortality rate due to BC is estimated to be
of eight women being reportedly affected during her 42,260 deaths (41,760 women and 500 men). Sixty-
lifetime. The increase of BC outbreak has been two percent of BC patients are diagnosed with a
significant in the last fifty years causing the highest tumor localized in the breast whose 5-year survival
number of deaths among all cancer types. In 2019 2
rate is 99%. However, the rate is reduced to 85% for
alone, approximately 268,600 women were those who have the cancer metastasized to the
regional lymph nodes and it will be dramatically
*Address for correspondence: dropped to 27% if the cancer spreads to distant
Yasin Panahi, PhD
Address: Pajoohesh Blv, Tehran-Karaj HWY, Tehran, 14965/161, organs.3,4
Iran Although most cancers originate from a mutated
Tel: +98 21 44580344 single cell, tumors grow and gain heterogeneous cell
Fax: +98 21 44580395
Email: [Link]@[Link], populations, so they have no common genotypic/
panahi_yasin@[Link] phenotypic profile. This heterogeneity is translated
22 Abdollahi, et al. Arch Breast Cancer 2020; Vol. 7, No. 1: 22-28
Artificial intelligence for detection of BC Metastasis
to some cancer cell populations with invasive and converts them into equivalent frequency coefficients.
5
metastasizing capacity causing tumor malignancy. Random Forest Classifier (RFC) is another algorithm
Metastasis is a biological hallmark of malignant based on an ensemble learning method that creates a
6
tumors and the main cause of cancer mortality. set of decision trees from randomly-selected subset
Therefore, early detection can have extremely of training set to aggregate votes and decide the final
profound impact on overall patient outcome and class of the test object. Support Vector Machines
survival. (SVM) is a supervised learning method that
BC is a systemic disease that may be treated via examines the data and sorts it into one of the two
7
surgery and chemo- and radiotherapy. Pathologically, categories. Multilayer Perceptron (MLP), on the
a single factor alone cannot be attributed to BC, scince other hand, is a feed-forward artificial neural
there are several factors that are proposed as BC risk network that generates a set of outputs from a set of
factors including age, personal history, breast inputs. Finally, Ensemble techniques are machine
pathology, family history, and genetic predisposition.8 learning techniques where more than one learner are
Also, interactions between environmental and genetic constructed for a given task, and Multilayer
factors can induce susceptible genes such as BRCA1 Perceptron/Genetic Algorithm (MLP-GA) can more
and BRCA2 to undergo mutation,9 which in turn leads accurately detect tumor cells. Given that the
to tumorigenesis. individual use of AI algorithms may not lead to rapid
Current diagnostic methods of tumor metastasis and accurate detection of tumor cells, we attempted
are developed as preclinical and clinical assays. to cross-compare a number of these index algorithms
Preclinical assays are used in ex-vivo samples and in in our study. We also combined MLP with genetic
in-vivo mice and comprise lab tests/histopathology, algorithm (GA) to create a hybrid pattern for boosted
non-invasive blood tests, small animal imaging, efficiency of detection.
molecular genetic imaging, and circulating tumor
cells. Approaches used to detect metastatic lesions at Methods
clinical levels include biomarker tracking, imaging In this study, we used data of digitalized
procedures, and circulating tumor cells.10 In fact, a pathologic images from Wisconsin Breast Cancer
diagnostic test is considered complete when it can dataset [([Link]
achieve 100% sensitivity and specificity. 11None of Breast+Cancer+Wisconsin+(Diagnostic)] that
these diagnostic methods meet this criterion and included 357 benign and 212 malignant BC samples.
have not yet been successfully translatable from We primarily reviewed the reports of the period
bench to bedside. 2006-2019 on diagnosis and classification of images
Some of these methods have been adopted for collected from cancer tissues based on various
cancer diagnosis for more than 40 years. However, algorithms. Our review indicated that application of
the biopsy-based histopathology is one of the most an ML algorithm alone has not been precise and
widely used methods to detect metastatic cells and successful in detection and prediction of diseases.
classify cancer. Pathologists use histopathological Next, we used 7 standardized and widely used
images to accurately identify tumor cells. With this algorithms to classify BC on our samples for
technique, they examine the microscopic structure of comparison. K-Nearest Neighbors (KNNs) that uses
12
the patient's tissue. Of note, this is the gold standard the probability of density estimation based on
for the detection and diagnosis of micrometastases in nuclear methods, and Decision Tree (DT) employed
13
tissues suspected to have tumor cells. However, for pattern extraction were the first two algorithms
analysis of slides made from biopsies is a laborious we applied. Our third algorithm was Forest Tree (FT)
task for a pathologist and requires sufficient time, and we used a collection of FTs to classify cancer
remarkable skill and prudence. It is not an error-free patients. We entered a set of data to each FT so the
task and might leave small metastases un-detected.14 algorithm could start learning. For prediction, we
In comparison, computer-aided histopathological used a new set of data so the FTs could predict the
analysis can play an important role in the detection outcome.
and prediction of BC tumor cells.15 In this case, the Logistic Regression (LR) was the fourth
use of Artificial Intelligence (AI) algorithms is an algorithm employed in order to segregate benign
attempt to improve speed and accuracy of diagnosis. tumors from malignant ones. For this, we applied 30
Several AI algorithms have been developed and parameters as potential risk factors and compared the
applied to cancer detection. Logistic Regression results using Receiver Operating Characteristics
(LR) is a statistical and mathematical procedure that (ROP) in terms of the level of sensitivity and
explores data sets with one or more free variants that specificity. In MLP that simulates human brain
hold a result, whereas K-Nearest Neighbors (KNN) performance16, we used Gradient-based Multilayer
is a supervised machine learning method used to Perceptron neural network as an MLP type widely
solve classification and regression problems. applied to BC diagnosis. The next algorithm we used
Decision Tree Classifier (DTC) works by separating was SVC(Support Vector Classifier) the kernel of
images into parts of differing importance and which was used for the classification of polished data
Abdollahi, et al. Arch Breast Cancer 2020; Vol. 7, No. 1: 22-28 23
Artificial intelligence for detection of BC Metastasis
and enhanced precision of benign/malignant tumor applied in this study. It shows how our hybrid system
diagnosis. After the basic algorithms, group learning is designed. We entered the data into our machine
algorithm was applied to develop a model of data learning algorithms in 3 steps. In step one, basic
classification with enhanced precision and algorithm was applied individually but in step 2 they
performance compared to the basic algorithms. were then used in groups. In step 3, we applied our
Finally, in order to boost neural network hybrid system comprising of MLP and genetic
algorithms such as MLP, we applied genetic algorithm (MLP-GA). The algorithms were
algorithms (GAs) to add a new hybrid pattern of ultimately evaluated for their performance in
MLG-GA to individual algorithms. Indeed, we used predicting the samples.
GA to optimize initial weights in a neural network
and determine optimal quantities of the parameters in Comparison of algorithms for their accuracy of
both learning and detection processes of the system. diagnosis
We adopted two different strategies called Cross- Table 1 compares the performance of different
Validation and Holdout to our comparison studies. algorithms using the Holdout strategy. Data achieved
We applied 569 samples in each experiment and through KNN indicated that Gaussian Nuclear
devoted 60% for learning, 20% for evaluation, and Density Estimation based on Euclidean distance
20% for the main tests. For this, we applied various detects cancer with 93% accuracy.
criteria that included sensitivity, accuracy and Data collected from DT and FT showed,
specificity, True/False Positive Rates, Positive/ respectively, 93% and 97% accuracy. Logistic
Negative Predictive Values, methods of data Regression predicted cancer with 96% accuracy,
segregation, etc. 94% sensitivity and 95% specificity whereas MLP
produced 95% accuracy. The accuracy figures for
Results SVM and group learning were found, respectively,
Algorithms: an overview 96% and 93%. Figure 2 compares the accuracy rates
Figure 1 shows a flow chart of the algorithms we based on the Holdout approach.
Figure 1. A schematic flow chart of the algorithms.
Table 1. Comparison of accuracy rates.
Algorithm Accuracy Standard deviation Precision Recall F1-score
LR 0.96 0.00987 0.96 0.97 0.96
SVM 0.96 0.00987 0.96 0.97 0.96
RFC 0.97 0.00970 0.97 0.97 0.97
DTC 0.93 0.01101 0.94 0.94 0.94
MLP 0.95 0.00943 0.95 0.95 0.95
KNN 0.93 0.00966 0.94 0.94 0.94
Ensemble 0.93 0.00966 0.94 0.94 0.94
MLP-GA 0.98 0.00543 0.98 0.98 0.98
.Diagnosis of BC was compared in rates among various algorithms with Holdout approach.
24 Abdollahi, et al. Arch Breast Cancer 2020; Vol. 7, No. 1: 22-28
Artificial intelligence for detection of BC Metastasis
Holdout
100
ACC
80
Accuracy (%)
Precision
Recall
60
F1-score
40
20
0
M
LR
RF
LP
e
C
En NN
A
bl
DT
SV
-G
M
m
LP
ce
M
Figure 2. Comparison of accuracy rates. Graph shows the percentage of accuracy for each algorithm using Holdout comparison
approach. Each column represents at least two different comparison experiments carried out independently.
Comparison between these algorithms by the Table 2. Comparison of accuracy rates.
Cross-Validation strategy generated data that was Algorithm Accuracy SD
nearly identical to those of the Holdout approach. LR 0.93 0.01839
This showed that the MLP-GA hybrid has the highest KNN 0.93 0.01282
accuracy rate among all those algorithms used in our DTC 0.92 0.01121
study (Table 2 and Figure 3). RFC 0.95 0.01204
We employed boxplot in order to demonstrate the SVM 0.93 0.01966
accuracy rate of our algorithms in classifying cancer MLP 0.96 0.00966
data using cross-Validation. Figure 4 shows the Ensemble 0.94 0.01615
outcome of the boxplot analysis in the distribution of MLP-GA 0.997 0.00015
accuracy rates for the validation of the algorithms. A Diagnosis of BC was compared in rates among various algorithms
with Cross-Validation approach. LR, Logistic Regression;
boxplot was drawn for each figure of diagnosis KNN, K-Nearest Neighbors; DTC, Decision Tree Classifier;
algorithms. The results indicate that all algorithms RFC, Random Forest Classifier; SVM, Support Vector Machines;
have a range of the least distance between 86 and 100 MLP, Multilayer Perceptron; MLP-GA, Multilayer Perceptron/
with an average of 94. Genetic Algorithm; SD, Standard Deviation.
Cross-validation
100
80
Accuracy (%)
60
40
20
0
M
LR
FC
LP
e
N
A
bl
KN
DT
SV
-G
M
R
m
LP
se
M
En
Figure 3. Comparison of accuracy rates. The graph shows the accuracy percentage for each algorithm using cross-validation
comparison approach. Each column represents at least two different comparison experiments carried out independently.
Abdollahi, et al. Arch Breast Cancer 2020; Vol. 7, No. 1: 22-28 25
Artificial intelligence for detection of BC Metastasis
We next applied randomized search (RS) and grid compares the performance of various algorithms
search (GS) methods in order to optimize estimating using CM.
parameters and improve our results as well as find The data indicate the high performance of our
those parameters that affect the learning process in recommended algorithm MLP-GA so that its
parallel. Figure 5 shows the result of comparison and accuracy, sensitivity and specificity have reached,
evaluation for each algorithm using RS and GS. respectively, 98%, 98% and 94%. They also indicate
These findings indicate that our hybrid algorithm that the algorithm MLP_GA has a superior
MLP_GA has a superior performance in optimization performance and is capable of classifying all
and search for parameters effective in rapid diagnosis datasets.
of cancer with an accuracy rate of 98%.
In this study, we used various algorithms for Hybrid algorithm MLP-GA with the highest level
classifying BC datasets. In order to select the best of accuracy and lowest deviation
algorithm, we applied the timing criterion and its We showed that our hybrid system has an
complexity as an element for comparing multiple accuracy rate of 99.7% which was the highest among
algorithms for problem solving. Table 3 shows the all algorithms we studied (Table 2 and Figure 3).
results of comparing execution time complexity This system also showed a deviation rate of 0.00015
within various algorithms. As shown in Table 3, the disease prediction, the lowest rate compared to those
MLP-GA algorithm has higher performance in terms of other algorithms we studied.
of time complexity and in classifying datasets. Its
execution time reached 0.04 with an accuracy rate of Discussion
98.3%. Breast cancer is a malignant tumor with different
We, therefore, used Confusion Matrix (CM) to phenotypes. We applied AI-based effective algorithms
show the overall performance of algorithms so we to breast tissue images and addressed the question of
could evaluate their function or malfunction in the whether AI can help detect cancerous tissues and BC
classification and diagnosis of datasets. Table 4 metastasis as a result. We particularly found that a
Figure 4. Comparison of accuracy rates among algorithms in the distribution of accuracy rates in validating classification data.
Figure 5. Comparison of accuracy rate of algorithms for optimization and search for parameters effective in cancer detection
26 Abdollahi, et al. Arch Breast Cancer 2020; Vol. 7, No. 1: 22-28
Artificial intelligence for detection of BC Metastasis
Table 3. Comparison of timing complexity in the classification of datasets.
Algorithm Accuracy SD Precision Recall F1-score
KNN 0.120921 96 0.93 0.95 0.96 0.95
DT 0.099941 96 0.93 0.92 0.96 0.92
RF 0.442738 97 0.92 0.92 0.97 0.93
RG 0.101936 93 0.95 0.89 0.93 0.89
SVM 0.117928 95 0.93 0.92 0.95 0.92
MLP 0.124924 93 0.96 0.88 0.93 0.88
Ensemble 0.112928 93 0.94 0.89 0.93 0.89
MLP-GA 0.046721 98 0.997 0.98 0.983 0.983
Table 4. Comparison of performance among various algorithms using Confusion Matrix.
Algorithm LR SVM RFC DTC MLP KNN Ensemble MLP-GA
Accuracy l0.96 0.96 0.97 0.93 0.95 0.93 0.93 0.98
Sensitivity 0.94 0.95 0.96 0.93 0.95 0.93 0.91 0.98
Specificity 0.95 0.95 0.97 0.93 0.95 0.89 0.89 0.94
PPV 0.90 0.90 0.88 0.89 0.91 0.90 0.86 0.98
NPV 0.90 0.89 0.90 0.90 0.90 0.89 0.86 0.98
TPR 0.92 0.90 0.96 0.90 0.89 0.90 0.88 0.98
FPR 0 0 0 0 0 0 0 0
combined algorithmic procedure of MLP-GA can algorithm to detect BC occurrence and recurrence.
more efficiently minimize errors of diagnosis. They designed a multi-layer perceptron using a GA
Of the AI algorithms that we examined, ANN routine and found that this combination is superior to
identifies patterns and processes that originate from individual approaches in providing accurate
the biological neural system and operates similar to classification of BC samples.27
the brain. ANN is comprised of a vast number of Our hybrid system, in line with the systems
extraordinary interconnected processing elements reported above, recommends precise therapeutic
called neurons that act in concert to solve a problem. procedures for complex diseases, reduces errors in
16
It has been applied to BC detection and risk clinics and improves enrolment in clinical trials.
18, 19
assessment. K-Nearest neighbors algorithm is a Since AI technology is rapidly growing in medicine,
non-parametric method used for classification and it will practically address data sharing and privacy,
regression. This approach has also been applied to the transparency of algorithms, standardization of
BC detection. 20, 21 Random Forest or Random data and cooperation in various operating systems.
Decision Forest is considered a supervised learning Our main objective in the near future is designing
algorithm that builds a forest by accident. Accidental intelligent models for BC metastasis and image
Forest makes several Decision Trees and then links classification assisted by early diagnosis of cancer
them to produce more accurate and reliable using AI particularly MLP.
predictions. It has been applied to BC detection.22 In conclusion, our hybrid algorithm MLP-GA
All the algorithms we used to detect BC worked was found superior to our other algorithms in
with reasonable accuracy and speed. However, they detecting BC incidence more reliably and accurately.
differ from one another in their composition and
functionality. In order to benefit from their strength Conflict of Interests
and enhance the efficiency of diagnosis, we created None.
the MLP-GA hybrid algorithm that demonstrated
maximum precision with minimum errors. Similar References
attempts have been already made to detect heart 1. American Cancer Society: [Link] [Link]/
23-25
disease and Coronary Artery Disease. In an content/dam/cancer-org/research/cancer-facts-and-
elegant study, Mobadersany et al. applied their statistics/breast-cancer-facts-and-figures/breast-
convolutional neural networks to combine histology [Link].
images and genomic biomarkers into a single unified 2. Cancer Facts & Figures 2018. Atlanta: American
framework and showed that this hybridized system is Cancer Society; 2018.
superior to current procedures in predicting the 3. Siegel RL, Miller KD, Jemal A. Cancer statistics,
overall survival of glioma patients.26 Belciug and 2018. CA: A Cancer J Clin. 2018; 68(1):7-30.
Gorunescu applied a hybrid neural network/genetic 4. Breast Cancer Facts & Figures 2017-2018.
Abdollahi, et al. Arch Breast Cancer 2020; Vol. 7, No. 1: 22-28 27
Artificial intelligence for detection of BC Metastasis
Atlanta: American Cancer Society, Inc.; 2017. 2010;116(14):3310–3321.
5. Meacham C, Morrison S. Tumour heterogeneity 20. Altman N S. An introduction to kernel and
and cancer cell plasticity. Nature. 2013. nearest-neighbor nonparametric regression. The
501:328–337. Am Stat. 1992; 46 (3): 175–185.
6. National Cancer Institute, 2019, [Link] 21. Sarkar M, Leong TY. Application of K-nearest
[Link]/types/metastatic-cancer. neighbor's algorithm on breast cancer diagnosis
7. Sarıtaş AG, Yalav O, Kekeç Y, Sakman G, et al. problem. Proc AMIA Symp. 2000;759–763.
Effect of neoadjuvant chemotherapy on estrogen 22. Zhang H, Wang M. Search for the smallest
receptor, progesterone receptor, Cerb-B2, random forest. Stat Interface. 2009; 2(3):381.
vascular endothelial growth factor and Ki-67 in [Link] X, Wang X, Su Q, et al. A Hybrid
patients with locally advanced breast cancer. Classification System for Heart Disease
Cukurova Med J. 2019; 44(1):226-231. Diagnosis Based on the RFRS Method. Comput
8. Shah R, Rosso K, Nathanson SD. Pathogenesis, Math Methods Med. 2017; 2017:8272091.
prevention, diagnosis and treatment of breast 24. Verma L, Srivastava S, & Negi PC. A Hybrid
cancer. World J Clin Oncol. 2014;5(3):283–298. Data Mining Model to Predict Coronary Artery
9. Nathanson, KN, Wooster R, & Weber BL. Breast Disease Cases Using Non-Invasive Clinical
cancer genetics: what we know and what we Data. J Med Syst. 2016, 40, 178 (2016).
need. Nature medicine, 2001; 7(5):552. 25. Kolukisa B, Hacilar H, Goy G, Kus M, Bakir-
10. Menezes ME, Das SK, Minn I, Emdad L, Wang Gungor B, Aral A, Gungor VC. Evaluation of
XY, Sarkar D, Pomper MG, Fisher PB. Detecting classification algorithms, linear discriminant
Tumor Metastases: The Road to Therapy Starts analysis and a new hybrid feature selection
Here. Adv Cancer Res. 2016; 132: 1-44. methodology for the diagnosis of coronary artery
11. Mordente, A., Meucci, E., Martorana, G. E., & disease. 2018 IEEE International Conference on
Silvestrini, A. Cancer Biomarkers Discovery and Big Data: 2232-2238.
Validation: State of the Art, Problems and Future 26. Mobadersany P, Yousefi S, Amgad M, et al.
Perspectives. Adv Exp Med Biol. 2015; 867:9-26. Predicting cancer outcomes from histology and
[Link] R, Srivastava R, & Srivastava S. genomics using convolutional networks. Proc Natl
Detection and Classification of Cancer from Acad Sci U S A. 2018;115(13):E2970–E2979.
Microscopic Biopsy Images Using Clinically [Link] S, Gorunescu F. A hybrid neural
Significant and Biologically Interpretable network/genetic algorithm applied to breast
Features. J Med Eng. 2015:14. cancer detection and recurrence. Expert
13. Bird B, Romeo M, Laver N, & Diem M. Spectral Systems. 2013; 30(3):243-254.
detection of micro-metastases in lymph node
histo-pathology. J Biophotonics. 2009; 2(1-
2):37-46.
14. Spanhol FA, Oliveira LS, Petitjean C, Heutte L.
A Dataset for Breast Cancer Histopathological
Image Classification. IEEE Trans Biomed Eng.
2016; 63:1455-1462.
15. Aswathy M, & Jagannath M. Detection of breast
cancer on digital histopathology images: Present
status and future possibilities. Informatics in
Medicine Unlocked. 2017; 8:74-79.
16. Abdollahi J, Moghaddam BN, & Parvar ME.
Improving diabetes diagnosis in smart health
using genetic-based Ensemble learning
algorithm. Approach to IoT Infrastructure.
Future Gen Distrib Systems J. 2019; 1:23-30.
17. Saritas I. Prediction of Breast Cancer Using
Artificial Neural Networks. J Med Syst (2012)
36: 2901.
[Link] M, Taghdir M, Rezaianzadeh A,
Rahimikazerooni S. Assessing Breast Cancer
Risk with an Artificial Neural Network. Asian
Pac J Cancer Prev. 2018; 19(4):1017-1019.
19. Ayer T, Alagoz O, Chhatwal J, Shavlik JW, Kahn
CE Jr, Burnside ES. Breast cancer risk
estimation with artificial neural networks
revisited: discrimination and calibration. Cancer.
28 Abdollahi, et al. Arch Breast Cancer 2020; Vol. 7, No. 1: 22-28