A new approach for histological classification of breast cancer using deep hybrid heterogenous ensemble
| DOI | https://doi.org/10.1108/DTA-05-2022-0210 |
| Published date | 21 April 2023 |
| Date | 21 April 2023 |
| Pages | 245-278 |
| Subject Matter | Library & information science,Librarianship/library management,Library technology,Information behaviour & retrieval,Metadata,Information & knowledge management,Information & communications technology,Internet |
| Author | Hasnae Zerouaoui,Ali Idri,Omar El Alaoui |
A new approach for histological
classification of breast cancer using
deep hybrid heterogenous ensemble
Hasnae Zerouaoui
MSDA, Mohammed VI Polytechnic University, Ben Guerir, Morocco
Ali Idri
Modeling, Simulation and Data Analysis, Mohammed VI Polytechnic University,
Ben Guerir, Morocco and
Software Project Management Research Team, ENSIAS, Mohammed V
University, Rabat, Morocco, and
Omar El Alaoui
UM5R ENSIAS, Rabat, Morocco
Abstract
Purpose –Hundreds of thousands of deaths each year in the world are caused by breast cancer (BC).
An early-stage diagnosis of this disease can positively reduce the morbidity and mortality rate by
helping to select the most appropriate treatment options, especially by using histological BC images for
the diagnosis.
Design/methodology/approach –The present study proposes and evaluates a novel approach which
consists of 24 deep hybrid heterogenous ensembles that combine the strength of seven deep learning
techniques (DenseN et 201, Inception V3, V GG16, VGG19, Incepti on-ResNet-V3, Mob ileNet V2 and
ResNet 50) for feature extraction and four well-known classifiers (multi-layer perceptron, support
vector machines, K-nearest neighbors and decision tree) by means of hard and weighted voting
combination methods for histological classification of BC medical image. Furthermore, the best deep
hybrid heterogenou s ensembles were compa red to the deep stacked ens embles to determine th e best
strategy to design the deep ensemble methods. The empirical evaluations used four classification
performance criteria (accuracy, sensitivity, precision and F1-score), fivefold cross-validati on, Scott–
Knott (SK) statistical test and Borda count voting method. All empirical evaluations were assessed
using four performance measures, including accuracy, precision, recall and F1-score, and were over the
histological BreakHis public dataset with four magnification factors (40×, 100×, 200× and 400×). SK
statistical test and Borda count were also used to cluster the designed techniques and rank the
techniques belongi ng to the best SK cluster , respectively.
Findings –Results showed that the deep hybrid heterogenous ensembles outperformed both their singles
and the deep stacked ensembles and reached the accuracy values of 96.3, 95.6, 96.3 and 94 per cent across the
four magnification factors 40×, 100×, 200× and 400×, respectively.
Originality/value –The proposed deep hybrid heterogenous ensembles can be applied for the BC diagnosis
to assist pathologists in reducing the missed diagnoses and proposing adequate treatments for the patients.
Keywords Breast cancer, Computer-aided diagnosis, Digital pathology, Deep convolutional neural
networks, Image processing, Histological images, Ensemble methods
Paper type Research paper
This work was conducted under the research project “Machine Learning based Breast Cancer
Diagnosis and Treatment”, 2020–2023. The authors would like to thank the Moroccan Ministry of
Higher Education and Scientific Research, Digital Development Agency (ADD), CNRST and UM6P for
their support.
Funding: This study was funded by Mohammed VI polytechnic university at Ben Guerir Morocco.
Conflicts of interest/competing interests: The authors report no conflicts of interest.
ThecurrentissueandfulltextarchiveofthisjournalisavailableonEmeraldInsightat:
https://www.emerald.com/insight/2514-9288.htm
245
Received 20 May 2022
Revised 24 June 2022
29 August 2022
Accepted 14 September 2022
Data Technologies and
Applications
Vol. 57 No. 2, 2023
pp. 245-278
© Emerald Publishing Limited
2514-9288
DOI 10.1108/DTA-05-2022-0210
Histological
classification
of BC using
DHHtE
1. Introduction
Breast cancer (BC) is one of the most eminent diseases in women worldwide. In 2020,
2.3 million cases of BC were reported, which makes this cancer the most common cancer in
women and a critical public health problem in both middle-income and developed countries
(Ferlay et al., 2021). Thus, an early diagnosis of the BC disease is a decisive factor to prevent
its progression and reduce the morbidity rate for women (Hameed et al., 2020). BC occurs
from the growth of abnormal breast cells that may invade healthy tissues (Sung et al., 2021).
The analysis of clinical radiology images such as ultrasound imaging, magnetic resonance
imaging (MRI) and mammography are initially used by radiologists for screening the BC
(Hameed et al., 2020); however, histological images remain the best and the most widely
used pathological method for a more precise BC diagnosis to effectively determine the
cancerous areas and therefore propose the effective treatment for the patients (Yan et al.,
2020). The diagnosis of histological images is usually conducted by manual qualitative
analysis by pathologists, which can face many challenges such as (1) lack of pathologists in
small hospitals, (2) dependency on the pathologist’s professional experience and knowledge
on the diagnosis, (3) complexity of the histological BC subtype diagnosis and (4) marginal
errors due to the large number of diagnoses (Fernandes et al., 2019;Gandomkar et al., 2018;
Stoffel et al., 2018). To this end, and facing the emergence of producing whole slide imaging
and digital histological slides, many studies investigated the use of new computer-aided
approaches and frameworks for a BC histological binary classification (Gandomkar et al.,
2018;Hameed et al., 2020;Yan et al., 2018,2020) in order to help the pathologists to confirm
or refute their diagnoses.
Nowadays, machine learning (ML) classifiers have been successfully used in many
application domains especially in the BC classification using medical images and image
processing (Zerouaoui and Idri, 2021). ML techniques have helped to increase the survival
rate by offering new automatic approaches that facilitate the BC diagnosis, ameliorate the
accuracy and reduce the effort and time of the diagnosis (Zerouaoui and Idri, 2021).
Moreover, deep learning (DL) techniques proved their strengths for feature extraction
(FE) due to their notable progress in computer vision when using medical images
(Zerouaoui and Idri, 2021), especially the deep convolutional neural network architectures
for histological BC diagnosis (Beevi et al., 2016;Berisha et al., 2019;Carvalho et al., 2020;
Gandomkar et al., 2018;Saikia et al., 2019;Valkonen et al., 2017;Wang et al., 2021;Zerouaoui
and Idri, 2022;Ghiasi and Zendehboudi, 2021;Kadam et al., 2019;Senan et al., 2021;
Boumaraf et al., 2021;Houssein et al., 2021). Despite the encouraging classification results
provided by ML and DL models, the use of one classifier does not always provide the best
outperforming model and the higher accuracy in all circumstances since (1) it highly
depends on the type of the problem dealt with and (2) each single classifier has its
advantages and weaknesses (Hosni et al., 2019). To deal with this limitation, researchers
investigated the ensemble learning approach (Ahmed and El Sadig, 2019;Idri et al., 2020;
Kassani et al., 2020;Nakach et al., 2022;El Ouassif et al., 2021) which consists of combining
single learners that are accurate and diverse in order to consolidate their advantages and
overcome their weaknesses using a combination rule such as simple majority voting or
weighted voting (Kuncheva, 2003). In the medical field and more precisely for BC diagnosis
classification, many studies proposed the use of ensemble learning to improve the
classification performances. For instance, the study by Hameed et al. (2020) proposed two
homogenous ensemble methods by combining end-to-end architectures of DL models (fine-
tuned VGG16 and VGG19 and fully trained VGG16 and VGG19) by means of average
predicted probabilities for histological image classification; the results proved the power of
the ensemble models compared to their singles. In the study by Ahmed and El Sadig (2019),
the authors proposed and evaluated a heterogenous ensemble method for BC detection
DTA
57,2
246
using mammograms by combining five classical architectures using multi-layer perceptron
(MLP), support vector machines (SVMs), decision tree (DT), K-nearest neighbors (KNNs)
and Naive Bayes (NB) for classification and traditional techniques for FE using the
weighted voting combination method; results showed that the proposed ensemble
technique outperformed their singles and improved the performances. However, some
limitations have been detected in the studies (Ahmed and El Sadig, 2019;Hameed et al.,
2020;Idri et al., 2020;Kassani et al., 2020;El Ouassif et al., 2021;Xiao et al., 2017): (1) the
design of heterogenous or homogenous ensemble using only one combination method, (2)
the use of end-to-end or classical architectures for the design of the ensemble methods and
(3) except for the studies (Idri et al., 2020;El Ouassif et al., 2021), it is observable that there is
a lack of use of statistical test to evaluate the performances of the proposed ensemble
approaches.
In a previous work (El Alaoui et al., 2022), deep stacked ensembles were developed using
seven pretrained DL models: VGG16, VGG19 (Simonyan and Zisserman, 2015), ResNet 50
(He et al., 2016), Inception V3 (Szegedy et al., 2016), Inception-ResNet-V2 (Szegedy et al.,
2017), Xception (Chollet, 2017) and MobileNet (Sandler et al., 2018); then a logistic regression
was used as a meta learner that learns how to best combine the predictions of the DL
models. The results show that the proposed deep stacking ensemble reports an overall
accuracy of 93.8, 93.0, 93.3 and 91.8 per cent over the four magnification factors (MF) values
of the BreakHis dataset: 40×, 100×, 200× and 400×, respectively. In order to compare the
results of the study (El Alaoui et al., 2022), and to elevate the burdens of the previous related
works, this paper develops and evaluates the performances of 24 deep hybrid heterogenous
ensembles (DHHtEs) using DL models (DenseNet 201 (Huang et al., 2017), Inception V3,
VGG16, VGG19, Inception-ResNet-V3, MobileNet V2 and ResNet 50) for FE and ML models
(MLP, SVM, DT and KNN) for classification over the BreakHis histological images dataset.
The choice of the members of base learners for the DHHtEs is based on the finding of the
previous studies (Zerouaoui et al., 2021;Zerouaoui and Idri, 2022) which designed 28 hybrid
architectures using seven DL techniques for FE including DenseNet 201, Inception V3,
VGG16, VGG19, Inception-ResNet-V3, MobileNet V2 and ResNet 50 and four ML classifiers
(MLP, SVM, DT and KNN). Results showed that for all the four MF values 40×, 100×, 200×
and 400× of the BreakHis dataset, the hybrid architecture MLP for classification and
DenseNet 201 for feature extraction (MDEN) constructed using DenseNet 201 for FE and
MLP for classification outperformed the others. Furthermore, to design the proposed new
approach of DHHtE, we selected the best-performing hybrid architectures of the study
(Zerouaoui and Idri, 2022) for each classifier, combining top 2, top 3 and top 4 by means of
hard and weighted voting. We therefore obtain 24 ensembles (3 ensembles with hard voting
for each MF + 3 ensembles with weighted voting for each MF) × 4 MF values. To the best of
our knowledge, this study is the first to propose, design and evaluate DHHtEs using deep
hybrid architectures as base learners, built using DL techniques as feature extractors and
four ML classifiers, tested on the histological BreakHis dataset for a binary BC diagnosis
classification. To this end, the present study discusses three research questions (RQs):
RQ1. What is the overall performance of the DHHtE designed?
RQ2. Does the DHHtE methods outperform their singles?
RQ3. Among the two combination methods (hard voting and weighted voting), the
number and the type of singles, which of them provides a better accuracy for the
DHHtEs?
RQ4. Does the best DHHtEs outperform the deep stacked ensembles?
The main contributions of this empirical study are the following:
Histological
classification
of BC using
DHHtE
247
To continue reading
Request your trial