Journal Pre-proof

BreastNet: A novel convolutional neural network model through

histopathological images for the diagnosis of breast cancer

Mesut Toğaçar, Kutsal Baran Özkurt, Burhan Ergen, Zafer Cömert

PII: S0378-4371(19)31999-5
DOI: https://doi.org/10.1016/j.physa.2019.123592
Reference: PHYSA 123592

To appear in: Physica A

Received date : 8 May 2019

Revised date : 4 September 2019

Please cite this article as: M. Toğaçar, K.B. Özkurt, B. Ergen et al., BreastNet: A novel
convolutional neural network model through histopathological images for the diagnosis of breast
cancer, Physica A (2019), doi: https://doi.org/10.1016/j.physa.2019.123592.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the
addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive
version of record. This version will undergo additional copyediting, typesetting and review before it
is published in its final form, but we are providing this version to give early visibility of the article.
Please note that, during the production process, errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.

© 2019 Published by Elsevier B.V.

 Journal Pre-proof

*Highlights (for review)

Highlights

 The classification was performed by using breast tumor histopathological images.

of
 We come up with a novel deep learning model (BreastNet) developed based on CNN.

 The BreastNet model includes attention module, hypercolumn technique, residual block.

p ro
 Other layers of the proposed model consist of dense blocks.

 The BreastNet model achieved 98.80% classification success using BreakHis data.

e-
Pr
al
urn
Jo
 Journal Pre-proof

*Manuscript
Click here to view linked References

BreastNet: A Novel Convolutional Neural Network Model

through Histopathological Images for The Diagnosis of Breast
Cancer

of
Mesut TOĞAÇAR *

Department of Computer Technology, Fırat University

Elazığ, Turkey

p ro
mtogacar@firat.edu.tr

ORCID
0000-0002-8264-3899

*
Corresponding author
e-
Kutsal Baran ÖZKURT

Buca İnci Özer Tırnaklı Science High School

İzmir, Turkey
kutsal_baran@hotmail.com
Pr
ORCID
0000-0002-4697-5802

Burhan ERGEN

Department of Computer Engineering, Faculty of Engineering, Fırat University

Elazığ, Turkey
al

bergen@firat.edu.tr

ORCID
0000-0003-3244-2615
urn

Zafer CÖMERT

Department of Software Engineering, Faculty of Engineering, Samsun University

Samsun, Turkey
zcomert@samsun.edu.tr

ORCID
Jo

0000-0001-5256-7648

1
 Journal Pre-proof

Abstract

Breast cancer is one of the most commonly diagnosed cancer types in the woman and automatically
classifying breast cancer histopathological images is an important task in computer-assisted pathology

of
analysis. Statistics indicate that the breast cancer rate is about 12% in all cancer cases in the world.
Also, approximately 25% of women have breast cancer. Therefore, rapid and accurate analysis of
breast cancer images is extremely important for diagnosis. Recently, deep learning models have been
used in preference for this purpose. In short, the most important reason why we use a deep learning

p ro
model for the diagnosis of breast cancer is can give faster and more accurate results than existing
machine learning based methods. In this study, we come up with a novel deep learning model
developed based on a convolutional neural network. The success of the classification was increased by
using the proposed model named as BreastNet. The general structure of the BreastNet model is a
residual architecture built on attention modules. Each image data is processed by the augmentation
techniques before applying it as input to the model. With augmentation techniques, each image is
e-
processed one by one and transferred to BreastNet. There is no increase in the number of data. The
features of each image are changed using some augmentation techniques, such as flip, shift, brightness
change and rotation. Then, each image that comes to the model performs the selection and processing
Pr
of important key regions of the image via through attention modules. Also, a more stable and accurate
classification of the data is performed by using the hypercolumn technique in the model. Other parts of
the BreastNet model consist of convolutional, pooling, residual and dense blocks. As a result, 98.80%
classification success was achieved with the proposed model. The success rate of the proposed model
was better than the success rates of AlexNet, VGG-16 and VGG-19 models performed on the same
dataset. In addition, the results obtained in this study yielded better results than the other studies that
al

use the current BreakHis dataset.

Keywords: Biomedical image processing, attention module, diagnosis system, breast cancer,
hypercolumn, deep learning.
urn
Jo

2
 Journal Pre-proof

1. Introduction
Cancer is a collection of diseases in which cells in the body come together to form lumps called
malignant tumor. These cells grow in an uncontrollable way, spread into surrounding tissues and
crowd out the normal cells [1]. From past to present, cancer is one of the most important disease that

of
threaten human health [2]. In the study conducted in 2018, it has been estimated that 18.1 million
cancer cases will be added to the existing cancer cases in the world and approximately 9.6 million of
these cancer cases will result in death [2,3]. There are more than 100 different types of cancer in

p ro
medical science [4]. Nowadays, breast cancer is one of the most common type of cancer that is
frequently mentioned among women and there are many studies on breast cancer in the scientific
sense [5–7]. Breast cancer is the most commonly diagnosed cancer among women in 140 of 184
countries worldwide [8]. It has been estimated that approximately 20% of breast cancers over the
world arise from modifiable risk factors, including alcohol use, excess body weight, and physical
inactivity [9]. Therefore, early and accurate diagnosis of breast cancer is extremely important.
e-
In the biomedical field, the examination and diagnosis of the breast cancer histopathological
images by field experts is a sensitive and labor-intensive process requiring time and high qualification.
The diagnosis process can be supported by utilizing existing technological tools and softwares. Thus,
Pr
the cost and diagnosis effort can be significantly reduced. For this purpose, numerous studies have
been conducted based on computational approaches. Support vector machines (SVMs) equipped with
a feature selection algorithm has been introduced to detect breast cancer. 99.51% classification success
was achieved in the study [1]. An expert system based on the association rules (AR) and neural
network (NN) have been suggested for breast cancer diagnosis. AR ensured a reduction in the number
of features whereas NN was employed for the intelligent classification. The accuracy of the expert
al

system was 95.6% [10]. The fuzzy systems and evolutionary algorithms have been combined for the
same purpose. The model provided a few simple rules for the experts’ interpretation [11]. Epithelial
(EP) and stromal (ST) that are two types of tissues in histological images have been segmented
urn

automatically using the deep convolutional neural network (DCNN) and the model provided
satisfactory results. The breast cancer histopathological images have been classified using ImageNet
pre-trained DCNN model named as AlexNet. Various experiments were conducted on a public dataset.
As a result, the model yielded 90% accuracy with deep fusion rules [12]. A novel breast cancer
algorithm, convolutional neural network improvement for breast cancer classification (CNNI-BCC)
has been suggested for the diagnosis task. The model was performed on the mammographic breast
Jo

cancer medical images and reached 90.50% accuracy [13]. CNNI-BCC model has been offered to
support medical experts in breast cancer diagnosis in a timely manner. CNNI-BCC model has a
capable of classifying the incoming breast cancer medical images according to malignant, benign, and

3
 Journal Pre-proof

healthy. The model achieved 90.50% classification accuracy [13]. An intelligent diagnosis approach
for breast cancer diagnosis has been suggested, which utilize information gain directed simulated
annealing genetic algorithm wrapper (IGSAGAW) for feature selection to reveal the top optimal
feature performed the cost-sensitive support vector machine (CSSVM) learning algorithm. The model

of
achieved 95.80% classification success [14]. In another study, incremental boosting convolutional
networks have been introduced for providing an efficient diagnosis model of breast cancer from
histopathological microscopic images. The model yielded an accuracy of 96.4% and 99.5% for four

p ro
and two classes classification tasks, respectively [15].

It is clearly seen that numerous models have proposed to improve the efficiency of breast cancer
diagnosis process. In this regard, one of the common models preferred for the diagnosis of breast
cancer is the deep learning approach [16–18]. In this study, a novel deep learning model is proposed
for breast cancer diagnosis. In the model that we call BreastNet, attention modules and hypercolumn
technique were used together. In addition, more efficient qualifications have been obtained by
e-
applying the augmentation technique to the input images. The input data was set to 224×224 pixels.
The general architecture of BreastNet consists of convolutional, dense and residual blocks. The CNN
models such as AlexNet, VGG -16, VGG -19 were also used to compare the performance of the
Pr
BreastNet model. The features obtained from the last fully connected layer (FC8) of the CNN models
were provided as input to the Softmax activation function. Moreover, the classification performances
of the models are examined separately. The experiments are carried out an open-access BreakHis
dataset composed of histopathological images [12].

The rest of this study is organized as follows: In Section 2, the used publicly available dataset is
described. In Section 3, the proposed novel CNN model, pre-trained deep CNN models, machine-
al

learning algorithms and the steps of the experiments are presented. The experimental results are
reported in Section 4. The discussion is given in Sections 5. Lastly, concluding remarks are presented
in Section 6.
urn

2. Description of the BreakHis Dataset

BreakHis is a dataset that includes totally 7909 images and eight sub-classes of breast cancers.
The source data comes from 82 anonymous patients of Pathological Anatomy and Cytopathology Lab
in Brazil. A normal and cancerous sample recording belongs to the dataset is shown in Fig. 1. The
BreakHis dataset is divided into two main groups: benign tumors and malignant tumors. Of these
patients, 24 have benign breast cancer and 58 have malignant breast cancer [12]. Histologically benign
Jo

is a term referring to a lesion that does not match any criteria of malignancy – e.g., marked cellular
atypia, mitosis, disruption of basement membranes, metastasize, etc. Normally, benign tumors are

4
 Journal Pre-proof

relatively “innocents”, presents a slow-growing and remains localized. A malignant tumor is a

synonym for cancer: lesion can invade and destroy adjacent structures and spread to distant sites to
cause death [19,20].

of
e- p ro
Figure 1. Histopathological images of benign breast tumor and malignant breast tumor: (a) benign tumor with the magnification rate 40X.
(b) benign tumor with the magnification rate 100X. (c) benign tumor with the magnification rate 200X. (d) benign tumor with the
magnification rate 400X. (e) malignant tumor with the magnification rate 40X. (f) malignant tumor with the magnification rate 100X. (g)
malignant tumor with the magnification rate 200X. (h) malignant tumor with the magnification rate 400X.
Pr
Table 1. The BreakHis dataset provides 7909 histopathological images collected from 82 anonymous patients. The samples were divided
into benign and malignant tumors and eight sub-classes tumors that consist of four magnification rates: 40X, 100X, 200X, and 400X.

Magnification Rate
Class Sub-class Total
40X 100X 200X 400X
A 114 113 111 106 444
F 253 260 264 237 1014
al

Benign
TA 109 121 108 115 453
PT 149 150 140 130 569
DC 864 903 896 788 3451
urn

LC 156 170 163 137 626

Malignant
MC 205 222 196 169 792
PC 145 142 135 138 560
Total 1995 2081 2013 1820 7909

BreakHis is separated into benign and malignant tumors that consist of four zooming rates: 40X,
Jo

100X, 200X, and 400X. The benign and malignant cells are divided into subclasses according to cell
diameter and shape by pathologists. Therefore, the dataset currently contains four histopathological

5
 Journal Pre-proof

distinct types of benign breast tumors. These, adenosis (A), fibroadenoma (F), phyllodes tumor (PT),
and tubular adenoma (TA).

The names of the malignant tumor classes are ductal carcinoma (DC), lobular carcinoma (LC),

of
papillary carcinoma (PC), and mucinous carcinoma (MC) [12]. The statistical information of the
disease types is given in Table 1. Also, each image consists of three-channel RGB and each channel
occurs at a depth of eight bits. Each image has 700×460 pixel resolution and the image format is PNG.
All data given in Table 1 were used in this study.

p ro
3. Models and Methods
3.1. Deep Learning Models

In this section, the AlexNet [21], VGG-16, and VGG-19 [22] models are considered and described
briefly. These three models have already announced their name in ImageNet competitions. AlexNet is
e-
one of the most important CNN architectures. This architecture consists of convolution, pooling and
FC layers. The input size of this model is 227×227 pixels. Filters form the output of the related layer
by applying convolution to the inputs got from the previous layer. Filters that are hovered over the
image in the convolution layer are typically selected as 3×3 or 5×5 pixels. The convolutional layer is
Pr
based on the process of circulating a particular filter over the input image. This convolution process
results in an activation map. The activation map consists of local discriminative features [23]. The
pooling layers are used in AlexNet architecture. A pooling layer has a structure that maintains the
image features, reduces image size and computation costs [24]. This structure reduces the number of
parameters in the model operation and does this by preserving the information obtained from the
image [25].
al

VGG-16 consists of convolutional, pooling and FC layers. It contains a total of 21 layers [22].
The most important feature of this architecture is having an increasing network structure. The input
urn

size of the model is 224×224 pixels. The filter size in the convolutional layer is 3×3. In this
architecture, the final layers have FC layers used for feature extraction. In this architecture, Softmax is
used as the activation function of the last layer [26,27].

The VGG-19 consists of 24 layers: convolution layers, pool layers, and FC layers. The number of
steps is two and the maximum pooling layer is used. The VGG-19 contains about 138 million
computational parameters. Compared to AlexNet, the VGG-19 is a deeper CNN architecture with
Jo

more layers. To reduce the number of parameters in such deep networks, it uses small filters in all
convolutional layers. The size of the selected filter in this architecture is 3×3 [28].

6
 Journal Pre-proof

As mentioned in this study, the last fully connected layer, FC8 activation layer, has been used for
deep feature extraction. Softmax function is used as the output function of the last layer. The
architectural designs of AlexNet, VGG-16 and VGG-19 models are illustrated in Fig. 2.

of
e- p ro
Figure 2. The schematic diagram of (a) VGG-16, (b) VGG-19 and (c) AlexNet models.

3.2. Optimization Methods

The main purpose of optimization methods is to update the weights at every batch to find global
Pr
minima.

In the Stochastic Gradient Descent (SGD) method [16,29], the weights update is performed for
each training set. Because of this reason, it tries to achieve the goal as early as possible time. The
formulation of SGD optimization is shown in Eq. (1). Here, is the weight vector to be updated, α is
the learning coefficient and is the cost function.
al

= –α (1)

Stochastic Gradient Descent with Warm Restarts (SGDR) is a variant of learning rate scheduler,
which gradually decreases the learning rate (LR) in defined cycles while model is training. SGDR uses
urn

cosine annealing, which decreases the learning rate in the form of half a cosine curve. Herein is
the learning rate at time step (incremented each mini-batch), and define the range of
desired learning rates, represents the number of epochs since the last restart (this value is
calculated at every iteration and thus can take on fractional values), and defines the number of
epochs in a cycle [30]. The main formula of the SGDR method is shown in Eq. (2). The arguments
used in the SGDR method are:
Jo

 Minimum LR - Maximum LR: The lower and upper bound of the learning rate range for the
experiment.
7
 Journal Pre-proof

 Steps per epoch: Number of mini-batches in the dataset. Calculated as (epoch size/batch size).
 LR Decay: Reduce the maximum LR after the completion of each cycle. To reduce the
maximum LR by 20% after each cycle, this value was set to 0.8.
 Cycle Length: Initial number of epochs in a cycle.

of
 Mult. Factor: Scale epochs to restart after each full cycle completion.

(2)

p ro
The RMSProp method [29,31] is adapted to the average of the slope weights and maintains
learning rates per parameter. This method works well in online and non-stationary situations and
performs the parameter update using momentum on the scaled slope.

The ADAM method [32] is one of the methods that update the learning coefficient in each
e-
batch. It adopts parameter-learning rates based on the average first moment in RMSProp. It also uses
the average of the second moments of the slopes. This method is designed with the advantages of the
RMSProp method. In other words, this is one of the optimization methods that update the learning
coefficient in each batch. The most important feature of the ADAM method is that it adjusts the
Pr
learning rate of the weight parameters by estimating the first and second gradient moments in the
model network. The ADAM uses the exponential moving averages that are evaluated on a valid mini-
batch event and calculated on the gradient. The past gradient ( ) and past square gradient ( )
averages are calculated according to Eq. (3) and Eq. (4). is the variable used to calculate hyper
parameters and usually takes values between 0.9 and 0.999 [33].
al

(3)
(4)
urn

3.3. Machine Learning Method

The Softmax activation function is often used in the final layer of a neural network-based classifier.
Such networks are commonly trained under a log loss or cross-entropy regime, giving a non-linear
variant of multinomial logistic regression. The Softmax method is generally used in classification
processes where the classification label can take more value [34]. Softmax was used as the activation
function in the last layer of the BreastNet model like in the present CNN models used in this study
Jo

[35,36].

8
 Journal Pre-proof

A multi-layer perceptron (MLP) is a forward-class artificial neural network. An MLP consists of

at least three-node layers: an input layer, a hidden layer, and an output layer. The MLP employs a
supervised learning technique called training for backpropagation. Multiple layers and non-linear
activation distinguish MLP from a linear sensor and MLP separates data that cannot be separated

of
linearly [37].

3.4. The proposed CNN Model: BreastNet

CNNs are one of the most convenient architectures for the image classification task. CNNs use filters

p ro
methods to extract the most efficient features from the pixels of an image [38] consist of three bases:

Convolutional layers include convolution filters that are applied to the input image. Mathematical
operations are applied to the feature map to obtain a single value in each region where the filter passes.
After each convolutional layer typically ReLU activation function applied to prevent vanishing
gradient problem [39,40].
e-
Pool layers summarize the presence of features in patches of the feature map. On the other hand,
these layers are utilized to reduce the size of the previous layer output. The working principle of these
layers is based on the sliding window. Applies a statistical function on values in a certain window-size
Pr
known as kernel. In general, there are three types of pooling methods. These are max pooling, mean
pooling, sum pooling. Maximum pooling is often preferred for CNN-based tasks. It extracts sub-
regions of the feature map (e.g., 3x3 sub-region), takes their maximum value and deletes all other
values [41].

The dense layer is known as a fully connected layer. Dense layers perform classification on the
features extracted from the convolutional and pooling layers. The task of a dense layer, all nodes in the
al

layer is connected to each node in the previous layer [38,42].

CNN is composed of blocks that perform feature extraction. The final dense layer in CNN
contains a single artificial neuron node for each target class with a Softmax activation function to
urn

generate probabilities between 0–1 for each artificial neuron. The sum of these estimated Softmax
probabilities are equal to one [43].

BreastNet architecture is an end-to-end model. That is, each image data entering the proposed
model performs probability distributions between classes without another type of feature encoders or
decoders. BreastNet model consists of the following modules in general: convolutional block attention
Jo

module (CBAM), dense block, residual block and hypercolumn technique. CBAM is an attention
layer. It provides identification of important key places in histopathological images. In other words, it
selects the important area of the image and allows to model to focus there. Thus gives higher accuracy

9
 Journal Pre-proof

for the classification process and saves time. The CBAM includes both the channel attention module
and the spatial attention module. The Residual block makes the gradients smoother. It was used in the
BreastNet model to prevent overfitting and underfitting [44,45]. The Hypercolumn technique
examines BreakHis images on various scales and contributes to the understanding of the disease. In

of
other words, it provides more stable results in the classification stage and improves classification
performance [46]. The general design and parameter values of the BreastNet model are shown in Fig.
7.

p ro
3.4.1. Convolutional Block and Dense Block
These blocks consist of Conv2D/Dense (FC), batch normalization and ReLU activation layers. The
main purpose of these blocks is to extract relevant features from input tensors (like the image(2D) or
neuron(1D) output). Batch normalization is performed very well when it used just after convolution
and dense layers [47]. As shown in Fig. 3, Convolutional Block and Dense Block were integrated into
our proposed CNN model. The parameter values used in this study are shown in Fig. 3.
e-
Pr
al
urn

Figure 3. The schematic diagram of (a) Convolutional Block, (b) Dense block.

3.4.2. CBAM
CBAM is an effective module for feed-forward CNN model. CBAM examines the feature map
extracted from images at two dimensions: channel and spatial. The aim is to draw attention maps from
the channel and spatial dimensions. After, the attention maps are multiplied by the input feature map
for adaptive feature refinement. CBAM can be used in any CNN architecture [44]. CBAM includes
Jo

channel attention module and spatial attention module.

10
 Journal Pre-proof

Channel attention module refers to the channels of the feature map. This module shows the
regions that would be focused on the features map and contributes to the classification results more
efficient. With this module, feature maps obtained from input data are compressed [48].

of
In order to collect spatial information, both the average and maximum pooling information are
used together [49]. This information is obtained with both average-pool and max-pool features
simultaneously. The design of both modules is shown in Fig. 4. As shown in Fig. 4, the channel
module utilizes both max-pooling outputs and average-pooling outputs. The spatial module utilizes

p ro
similar two outputs that are pooled along the channel axis and forward them to a convolution layer.
The values specified in Fig. 4 are the parameter values used in this study.

With channel module, average and maximum pooled features process together. Two features set
are collected in a shared network. The shared network is composed of MLP with one hidden layer. It
subjects to a classification process with MLP [50].
e-
To compute the spatial attention, we first apply average-pooling and max-pooling operations
along the channel axis and concatenate them to generate an efficient feature descriptor. The extracted
information is then transferred to the convolutional network.
Pr
al
urn
Jo

11
 Journal Pre-proof

Figure 4. The schematic diagram of (a) Channel attention module, (b) Spatial attention module.

As a result, the channel and the spatial module focus on attention. Channel and spatial module
focuses on ‘what’ and ‘where’ respectively. And they contribute to the classification process by
extracting more efficient features [48,49]. The general design of CBAM consisting of the channel and

of
spatial attention modules is shown in Fig. 5.

e- p ro
Figure 5. The schematic diagram of the general design of the CBAM block.

3.4.3. Residual Block

Pr
In traditional neural networks, each layer feeds the next layer. In a network with residual blocks, each
layer feeds the next layer and initial layer connects the layers about two-three hops away with addition
operation [51].

For conventional deep learning networks, they usually have convolution layers then fully
connected layers for classification task like AlexNet, ZFNet, and VGGNet, without any skip/shortcut
al

connection, these structures are called sequential networks. When the sequential network is deeper
(layers are increased), the problem of vanishing/exploding gradients occurs [52]. To reduce the
negative results of this problem we used residual block after CBAM blocks in this study. Used
parameters and the design of the residual block are shown in Fig. 6.
urn
Jo

12
 Journal Pre-proof

of
e- p ro
Pr

Figure 6. The schematic diagram of the residual block.

3.4.4. Hypercolumn Technique

Hypercolumn at a pixel is the vector of activations of all CNN units above that pixel. By this mean,
al

spatial location information can be brought from earlier layers, and have a more accurate prediction
result. Since CNN normally uses the output of the last layer as a feature representation, it does not use
the features of the preceding layers. Therefore, hypercolumn in a pixel with the hypercolumn
technique is the vector of all CNN units over that pixel. In other words, it also includes the features of
urn

previous layers. In this way, allows the network to achieve better results. This technique consists of
UpSampling2D with bilinear interpolation and "Concatenate" layers [46]. The function of bilinear
upsampling2d determines the values of new neighboring pixels by looking at neighboring pixel values.
In addition, by enabling the size of the image to be input size; The effect of attention blocks has been
observed by the model. Thus, the better estimate of the BreastNet architecture has been provided for
this study. The concatenate parameter allows to combine the features of the same size in the third
Jo

dimension (the channels are combined). In addition, the dropout parameter is used to prevent overfit-
underfit problems in this study.

13
 Journal Pre-proof

Histopathological images in this study can be captured with different zoom rates. Instead of
creating models for every different zoom rates, with hypercolumn, the model examines
histopathological images at three different scales (stages) and selects relevant features. Also,
hypercolumn is preferred in BreastNet model architecture for providing high accuracy in this study.

of
The general design of the BreastNet model is shown in Fig. 7. The parameters specified in Fig. 7 are
the values used in the BreakHis dataset.

e- p ro
Pr
al
urn

Figure 7. The general design of the proposed BreastNet model.

4. Results
This study was carried out by using MATLAB (R2018b) and Python software. In all experiments, the
Jo

present models were compiled using the MATLAB whereas the results of the proposed BreastNet
model were obtained using the Python. Also, the BreakHis dataset was divided into two sets as 80%
training and 20% testing, respectively. The present CNN models used in this study were used with the
14
 Journal Pre-proof

transfer learning approach. The parameter values of the present CNN models used in the experiments
are given in Table 2. As inferred from Table 2, using these parameters with default values were
preferred. Also, mini-batch size was adjusted to 32. Mini-batch is the number of data, you update the
weights and biases during backpropagation. Mini-batch value is selected to be less than the total

of
dataset size. Generally, a number that is divisible by the total dataset size is preferred. Mini-batch
contributes to learning processes by balancing the convergence rate of the network as well as accurate
estimation [53]. However, the size of the mini-batch was not increased more since this is costly in

p ro
terms of memory usage [54].

The parameter values used in the BreastNet architecture are shown in Table 3. In the proposed
model, the K-fold cross-validation value was chosen as five. In the last layer of this architecture,
Softmax was selected as the activation function. The epoch value for each cross-validation was set to
100. Data augmentation was performed to extract more efficient attributes from each sample. The
arguments (flip, shift, brightness change, and rotation, etc.) and hyperparameter values used for this
e-
specific purpose are shown in Table 3. ADAM method was chosen for optimization, and SGDR
method was used to improve accuracy and learning speed during training. Also, the mini-batch size of
the BreastNet model was adjusted to 24. The reason for reducing the mini-batch size arises from the
Pr
insufficient in graphics card memory.

The models were compiled with GPU support. The simulation environment was run on 64-bit
Windows 10 operating system. Other hardware details of the computer are that NVIDIA GeForce
GTX 1070 - 8 GB graphics card, Intel® Core™ i7-8700 processor and 16 GB RAM.

Table 2. The values of the parameters used in the CNN architectures.

al

Used CNN Image Mini Learning

Optimization Momentum Decay
Software Architecture Size Batch Rate
AlexNet 227×227
MATLAB VGG-16 224×224 SGD 0.9 1e-6 32 0.0001
urn

VGG-19 224×224

Table 3. The values of the parameters in the proposed BreastNet architecture.

Optimization Data Augmentations LR Scheduler

& Parameters Parameters & Parameters
Model BreastNet ADAM Vertical Flip = 0.5 SGDR
Used Software Python Beta 1=0.9 Horizontal Flip = 0.5 Min LR=1e-6
Jo

Image Size 224×224 Beta 2=0.999 Random Brightness Contrast = 0.3 Max LR=1e-3
Mini Batch 24 Decay=0.0 Shift Scale Rotate = 0.5 Steps Per Epoch=5

15
 Journal Pre-proof

Framework Keras Shift Limit=0.2 LR Decay=0.9

Categorical Scale Limit=0.2 Cycle Length=10
Loss Type
Cross-entropy Rotate Limit=20° Mult. Factor=2.

The differences in the mini-batch values in Table 2 and Table 3 suffered from the following

of
reasons: The mini-batch value in Table 2 was used to compile present CNN models on MATLAB. The
mini-batch value in Table 3 was compiled on the BreastNet architecture in Python. These mini-batch
values were the largest values that can be used with the existing hardware units in the experiments.

p ro
In order to measure the performances of the models, accuracy (Acc), sensitivity (Se), specificity
(Sp), precision (Pr) and f1-score metrics derived from confusion matrix were used and the
formulations of the metrics are described as follows [55]:

(5)
e-
(6)
Pr
(7)

(8)

(9)
al

where True positive (TP) represents the number of malignant breast images classified as cancerous
breast whereas True Negative (TN) represents the number of normal breast images classified as
normal breast. Also, False positive (FP) represents the number of normal breast images incorrectly
classified as cancerous breast while False Negative (FN) represents the number of cancerous breast
urn

images misclassified as normal breast.

The experiment consists of three parts. In the first part, the results of the designed CNN model
with the results obtained by using the present CNN models were obtained and compared. The
comparison results are shown in Table 4. The BreastNet model compared to the present CNN models,
for each data type group (40X, 100X, 200X, 400X) the proposed model achieved superior results. This
Jo

model achieved the best success rate using 200X data with 98.51%. According to the results of the
BreastNet model, the learning accuracy graphs are shown in Fig. 8 and the confusion matrix are given
in Fig. 9. ROC curves are also shown in Fig. 10.

16
 Journal Pre-proof

Table 4. The comparison results of the existing CNN models and the BreastNet model.

Used BreakHis Acc Se Pr F1 - Score

CNN Model

of
Data Type (%) (%) (%) (%)
40X 70.05 63.25 95.72 76.17
100X 80.83 77.17 87.56 82.03
AlexNet
200X 84.22 93.24 73.80 82.39

p ro
400X 84.38 82.35 87.50 84.85
40X 77.81 72.42 89.84 80.19
100X 72.02 97.75 45.08 61.70
VGG -16
200X 81.28 96.07 65.24 77.71
400X 83.24 81.28 86.36 83.74
40X 83.96 91.51 74.87 82.36
100X 81.87 78.34 88.08 82.93
e-
VGG -19
200X 81.82 95.42 66.84 78.61
400X 82.39 80.98 84.66 82.78
40X 97.99 97.68 97.68 97.68
Pr
100X 97.84 97.21 97.80 97.50
BreastNet
200X 98.51 98.70 97.89 98.28
400X 95.88 95.16 95.36 95.26
al
urn
Jo

17
 Journal Pre-proof

of
e- p ro
Pr
al
urn
Jo

18
 Journal Pre-proof

of
e- p ro
Pr

Figure 8. Visual graph of the BreastNet model training and validation accuracy which has the best score in test dataset among the 5 cross-
validation models for all epochs, (a) Training and validation accuracy graph for 40X images. (b) Training and validation accuracy graph for
100X images. (c) Training and validation accuracy graph for 200X images. (d) Training and validation accuracy graph for 400X images.
al
urn
Jo

19
 Journal Pre-proof

of
p ro
Figure 9. The confusion matrix of the BreastNet model which has the best score in test dataset among the 5 cross-validation models, (a) The
confusion matrix of the model trained using 40X images, (b) The confusion matrix of the model trained using 100X images, (c) The
confusion matrix of the model trained using 200X images, (d) The confusion matrix of the model trained using 400X images.
e-
Pr
al
urn

Figure 10. The ROC curves of the BreastNet model, (a) The proposed model was trained using 40X images, (b) The proposed model was
trained using 100X images, (c) The proposed model was trained using 200X images, (d) The proposed model was trained using 400X
images.

In the second part of the experiment, BreakHis data (40X, 100X, 200X, 400X) were combined.
The combined data was obtained classification results using the present CNN (AlexNet, VGG-16,
Jo

VGG -19) models. Then, the combined data was classified by the BreastNet model. It was then
compared with the existing CNN models with the BreastNet model. The classification success of the
BreastNet increased by 0.29% and obtained as 98.80%. The statistical information on the results is
20
 Journal Pre-proof

shown in Table 5. The combined data was classified by the proposed model. The learning accuracy,
confusion matrix and ROC curve of the classification results are shown in Fig. 11.

Table 5. Results of comparison of existing CNN models with the BreastNet using all of the BreakHis data.

of
Used BreakHis Acc Se Pr F1 - Score
CNN Model
Data Type (%) (%) (%) (%)
40X & 100X &
AlexNet 78.49 72.18 92.74 81.18
200X & 400X

p ro
40X & 100X &
VGG -16 83.13 78.90 90.46 84.29
200X & 400X
40X & 100X &
VGG -19 83.27 83.13 83.47 83.30
200X & 400X
40X & 100X &
BreastNet 98.80 98.35 98.84 98.59
200X & 400X
e-
Pr
al
urn
Jo

Figure 11. BreastNet model which has the best score in test dataset among the 5 cross-validation models in whole trained on whole BreakHis
data, (a) Training and validation accuracy of the BreastNet model, (b) The confusion matrix of the BreastNet model, (c) The ROC curve of
the BreastNet model.
21
 Journal Pre-proof

Benign and malignant classes of BreakHis data were divided into four sub-classes. Benign
subclasses are adenosis, fibroadenoma, phyllodes tumor, tubular adenoma. Malignant subclasses are
ductal carcinoma, lobular carcinoma, mucinous carcinoma, papillary carcinoma. In the three-part of
the experiment, the performance of the BreastNet model in quadruple sub-classification success was

of
measured. When Table 6 is examined, the average classification success of subclasses of Benign data
was 97.78%. Likewise, the average classification success of the subclasses of Malignant data was
96.41%. The training & validation graph and the confusion matrix of the classification results are

p ro
shown in Fig. 12.

Table 6. The sub-classification success of the BreastNet model using benign and malignant data.

BreakHis Accuracy Recall Precision Average

Sub Classification
Data Type (%) (%) (%) Acc. (%)
Adenosis
e- 98.59 96.08 97.03
Fibroadenoma 97.18 97.54 95.65
Benign 97.78
Phyllodes tumor 96.98 91.25 90.12
Tubular adenoma 98.39 94.60 98.13
Pr
Ductal 93.46 95.26 94.29
Lobular 94.01 72.30 81.68
Malignant 96.41
Mucinous 98.90 98.71 93.87
Papillary 99.26 97.20 95.41
al
urn
Jo

22
 Journal Pre-proof

of
p ro
Figure 12. Sub-classification using benign and malignant data with the BreastNet model, (a) training and validation accuracy graph of the
BreastNet model in benign data, (b) the confusion matrix of the BreastNet model in benign data, (c) training and validation accuracy graph of
the BreastNet model in malignant data, (d) the confusion matrix of the BreastNet model in malignant data.

5. Discussion
e-
Breast cancer is on the front bench type among hundreds of cancer diseases. The incidence of this
disease is increasing day by day, especially among women. If this disease is not diagnosed in a timely
manner, the rate of death is fairly high. The early diagnosis of this disease is associated with rapid and
Pr
accurate results of the image processing techniques as regarding the computation approaches. In this
scope, the CNN models have a great advantage in terms of giving faster and better results compared to
the conventional machine learning methods [56]. Especially, these models have a powerful ability to
extracting the discriminative local features. Furthermore, the models can also reflect this ability to the
classification process. As a result, numerous studies have been focused on machine learning and deep
al

learning models to early diagnose breast cancer through BreakHis data. With the BreastNet model, we
realized a classification analysis with BreakHis data. The results of our study were compared with the
other studies using this dataset as shown in Table 7.
urn

Table 7. Comparison of the success of studies using the BreakHis data set with the success of the BreastNet model.

Used
Classifier / Acc
Study Year BreakHis # of Classes Model
Function (%)
Data Type
Linear
Benign &
Shallu et al.[57] 2018 All data VGG-16 Regression 92.60
Malignant
(LR)
Benign 95.03
40X
Malignant Deep Manifold Persevering 84.56
Jo

Yangqin Feng Benign Autoencoder 93.15

2018 100X Softmax
et al. [58] Malignant (DMPA) 82.98
Benign 99.36
200X
Malignant 83.57

23
 Journal Pre-proof

Benign 95.27
400X
Malignant 83.45
40X 95.62
100X The Densely-Connected 95.03
Yun Gu et al.
2018 Multi-Magnification Linear Layer
[59] 200X 97.04
(DCMM)

of
400X 96.31
40X Support Vector 88.30
Hamed 100X Machine 88.30
Local Binary Patterns
Erfankhah et al. 2019
200X (LBP) 87.10
[60] (SVM)

p ro
400X 83.40
40X Benign 81.61
Daniel 100X 84.47
& Various
Lichtblau et 2019 AlexNet
200X Malignant Classifiers 86.67
al.[61]
400X 83.15
40X 97.10
Class Structure-based Deep
Zhongyi Han et 100X Convolutional Neural 95.70
2017 Softmax
al.[62] 200X Network 96.50
(CSDCNN)
400X
e- 95.70
40X 97.99
Proposed 100X 97.84
2019 Designed CNN Model Softmax
Model 200X 98.51
400X 95.88
Pr

Shallu et al.[57] used BreakHis dataset. 40X, 100X, 200X, and 400X have applied the
classification by combining the data. They undertook feature extraction using the previously trained
VGG-16 architecture. They used LR as a classifier. The success of the classification was 92.60%. As
presented in Table 7, The success rate of the proposed model was 98.80%. So, our model produced
al

better results compared to the related study.

Yangqin Feng et al. [58] proposed a new feature extractor, called deep manifold preserving
autoencoder, to learn discriminative features from unlabeled data. Then, they integrated the proposed
urn

feature extractor with a Softmax activation function to classify breast cancer histopathology images.
They were identified the features obtained in the deep model with the sampling patches from the
breast histopathology images. They tried to classify the new sampled patches of the data and combine
the classification results of these patches to predict the label of the data. They have achieved success
by separately applying the model to the benign and malignant data.

Yun Gu et al. [59] are proposed to generate the discriminative binary codes by exploiting the
Jo

histopathological images with multiple magnification factors from the DCMM model. DCMM model
with a mutual guidance learning paradigm is conducted based on high-low magnification pairs of
image data and a densely-connected architecture is applied to fully utilize the cross-magnification
24
 Journal Pre-proof

information. They used multiple magnification levels to learn the binary codes for histopathological
BreakHis dataset. The best performance in the classification was achieved by using the 400X dataset
at a rate of 96.31%.

of
Hamed Erfankhah et al. [60] used 3 different datasets. They used the LBP method in BreakHis
and other datasets to extract features and extracted the texture features from a circularly symmetric
pixel neighborhood. Histogram values were extracted to differentiate between homogeneous and
heterogeneous regions of image tissue. The homogeneous regions have low values, while the non-

p ro
homogeneous regions have high values in the heterogeneity images. In their study, in the LBP method,
feature extraction was based on the heterogeneous and homogeneous status of tissue regions. They
used the SVM method as the classifier and achieved the best success at 88.30% in the BreakHis
dataset.

Daniel Lichtblau et al.[61] proposed a combination of six machine learning techniques employed
for the histopathological image classification task. In their work, they used the AlexNet model and the
e-
pre-trained weights and without fine-tuning. In the classification stage, five classifiers (random forests,
nearest neighbors, LR, naive Bayes, and SVM) exhibiting different features were used. Each classifier
was independently trained and applied to the test set. In addition, Fourier trig transform (FTT) with
Pr
principal component analysis (PCA) was used in the model. The success rate was determined as
86.67% by classifying the selected features with various classifiers.

Zhongyi Han et al.[62] proposed CSDCNN model in their study. This model consists of an input,
convolutional, pooling layer that similar the CNN architecture. In addition, the CSDCNN model has
used its own data augmentation technique. Also, they choose Softmax as the activation function in the
al

last layer of the architecture in their study. The success of the model was 97.1%.

We achieved the best results using the proposed model compared to all the studies given in Table
7. The most efficient results were obtained from 40X, 100X and 200X histopathological images.
urn

Especially, the best result in Table 7 was obtained with the proposed CNN model by using the 200X
images with 98.51% classification success was achieved.

6. Conclusion
The current study focused on improving the classification accuracy on the BreakHis data. It was
comprised of the histopathological images separated into two different classes as benign or malignant.
The classification accuracies of the BreastNet model is superior or approximate to the previously
Jo

attempted techniques on the same dataset. This model can be used in all microscopic images at
different magnification rates. The classification was carried out without using any preprocessing

25
 Journal Pre-proof

procedure for the histopathological images. As a result, the BreastNet model in this study and four
different magnification factors (40X, 100X, 200X, 400X) were classified as binary classification. The
best classification result was 98.51%. The binary classification was repeated by combining all data of
the magnification factors. The best classification result was obtained with an increase of 98.80%. The

of
same model was applied to the subclasses of BreakHis data. The classification results of the subclasses
consisting of four categories can be evaluated as promising.

In the future, we will examine the proposed model on a different dataset. The proposed method

p ro
can be generalized to the design of high-performance computer-aided diagnosis systems for other
medical imaging tasks in the future.

Data Availability and Open Source Code

The source codes of the proposed model can be downloaded on
https://github.com/Goodsea/BreastNet.
e-
Funding
There is no funding source for this article.

Ethical approval
Pr
This article does not contain any data, or other information from studies or experimentation, with the
involvement of human or animal subjects.

Conflicts of interest
The authors declare that there is no conflict to interest related to this paper.
al

References

[1] M.F. Akay, Support vector machines combined with feature selection for breast cancer
urn

diagnosis, Expert Syst. Appl. 36 (2009) 3240–3247.

doi:https://doi.org/10.1016/j.eswa.2008.01.009.

[2] P. Fribert, L. Paulová, P. Patáková, M. Rychtera, K. Melzoch, Alternativní metody separace

kapalných biopaliv z média při fermentaci, Chem. List. 107 (2013) 843–847.
doi:10.3322/caac.21492.
Jo

[3] World Health Organization, World Health Statistics, 2018.

[4] Y. Guo, X. Shang, Z. Li, Identification of cancer subtypes by integrating multiple types of
transcriptomics data with deep learning in breast cancer, Neurocomputing. 324 (2019) 20–30.
26
 Journal Pre-proof

doi:https://doi.org/10.1016/j.neucom.2018.03.072.

[5] American Cancer Society, Breast Cancer Facts & Figures 2012-2014, Breast Cancer Facts
Fig. (2013) 1–44. doi:10.1007/s10549-012-2018-4.Mesothelin.

of
[6] F.A. Spanhol, L.S. Oliveira, C. Petitjean, L. Heutte, A Dataset for Breast Cancer
Histopathological Image Classification, IEEE Trans. Biomed. Eng. 63 (2016) 1455–1462.
doi:10.1109/TBME.2015.2496264.

p ro
[7] G.A. Tataroglu, A. Genc, K.A. Kabakci, A. Capar, B.U. Toreyin, H.K. Ekenel, I. Turkmen, A.
Cakir, A deep learning based approach for classification of CerbB2 tumor cells in breast
cancer, 2017 25th Signal Process. Commun. Appl. Conf. SIU 2017. (2017).
doi:10.1109/SIU.2017.7960587.

[8] L.A. Torre, F. Islami, R.L. Siegel, E.M. Ward, A. Jemal, Global cancer in women: Burden and
trends, Cancer Epidemiol. Biomarkers Prev. 26 (2017) 444–457. doi:10.1158/1055-9965.EPI-
e-
16-0858.

[9] G. Danaei, S. Vander Hoorn, A.D. Lopez, C.J.L. Murray, M. Ezzati, Causes of cancer in the
world: Comparative risk assessment of nine behavioural and environmental risk factors,
Pr
Lancet. 366 (2005) 1784–1793. doi:10.1016/S0140-6736(05)67725-2.

[10] M. Karabatak, M.C. Ince, An expert system for detection of breast cancer based on association
rules and neural network, Expert Syst. Appl. 36 (2009) 3465–3469.
doi:https://doi.org/10.1016/j.eswa.2008.02.064.

[11] C.A. Peña-Reyes, M. Sipper, A fuzzy-genetic approach to breast cancer diagnosis, Artif. Intell.
al

Med. 17 (1999) 131–155. doi:https://doi.org/10.1016/S0933-3657(99)00019-6.

[12] F.A. Spanhol, L.S. Oliveira, C. Petitjean, L. Heutte, Breast cancer histopathological image
urn

classification using Convolutional Neural Networks, Proc. Int. Jt. Conf. Neural Networks.
2016-Octob (2016) 2560–2567. doi:10.1109/IJCNN.2016.7727519.

[13] F.F. Ting, Y.J. Tan, K.S. Sim, Convolutional neural network improvement for breast cancer
classification, Expert Syst. Appl. 120 (2019) 103–115.
doi:https://doi.org/10.1016/j.eswa.2018.11.008.

[14] N. Liu, E.-S. Qi, M. Xu, B. Gao, G.-Q. Liu, A novel intelligent classification model for breast
Jo

cancer diagnosis, Inf. Process. Manag. 56 (2019) 609–623.

doi:https://doi.org/10.1016/j.ipm.2018.10.014.

27
 Journal Pre-proof

[15] D.M. Vo, N.-Q. Nguyen, S.-W. Lee, Classification of breast cancer histology images using
incremental boosting convolution networks, Inf. Sci. (Ny). 482 (2019) 123–138.
doi:https://doi.org/10.1016/j.ins.2018.12.089.

of
[16] A.C. Wilson, R. Roelofs, M. Stern, N. Srebro, B. Recht, The Marginal Value of Adaptive
Gradient Methods in Machine Learning arXiv : 1705 . 08292v2 [ stat . ML ] 22 May 2018,
(2017).

p ro
[17] Q. Liao, Y. Ding, Z.L. Jiang, X. Wang, C. Zhang, Q. Zhang, Multi-task deep convolutional
neural network for cancer diagnosis, Neurocomputing. (2018).
doi:https://doi.org/10.1016/j.neucom.2018.06.084.

[18] S. Khan, N. Islam, Z. Jan, I.U. Din, J.J.P.C. Rodrigues, A novel deep learning based framework
for the detection and classification of breast cancer using transfer learning, Pattern Recognit.
Lett. 125 (2019) 1–6. doi:https://doi.org/10.1016/j.patrec.2019.03.022.
e-
[19] K.A. Chun, Case Reports on the Differentiation of Malignant and Benign Intratracheal Lesions
by 18 F-FDG PET/CT, Med. (United States). 94 (2015) e1704.
doi:10.1097/MD.0000000000001704.
Pr
[20] H. Kjellin, H. Johansson, A. Höög, J. Lehtiö, P.J. Jakobsson, M. Kjellman, Differentially
expressed proteins in malignant and benign adrenocortical tumors, PLoS One. 9 (2014).
doi:10.1371/journal.pone.0087951.

[21] A. Krizhevsky, I. Sutskever, G.E. Hinton, Machine Learning and Computer Vision Group
Deep Learning with Tensorflow, (2012). http://cvml.ist.ac.at/courses/DLWT_W17/.
al

[22] K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image
recognition, ArXiv Prepr. ArXiv1409.1556. (2014).
urn

[23] J. Koushik, Understanding Convolutional Neural Networks, (2016).

doi:10.1016/j.jvcir.2016.11.003.

[24] K. O’Shea, R. Nash, An Introduction to Convolutional Neural Networks, (2015).

doi:10.1007/978-3-642-28661-2-5.

[25] N. Passalis, A. Tefas, Learning Bag-of-Features Pooling for Deep Convolutional Neural
Networks, Proc. IEEE Int. Conf. Comput. Vis. 2017-Octob (2017) 5766–5774.
Jo

doi:10.1109/ICCV.2017.614.

[26] M.E. Sertkaya, B. Ergen, M. Togacar, Diagnosis of Eye Retinal Diseases Based on

28
 Journal Pre-proof

Convolutional Neural Networks Using Optical Coherence Images, in: 2019 23rd Int. Conf.
Electron., 2019: pp. 1–5. doi:10.1109/ELECTRONICS.2019.8765579.

[27] M. Toğaçar, B. Ergen, M.E. Sertkaya, Zatürre Hastalığının Derin Öğrenme Modeli ile Tespiti

of
Detection of Pneumonia with Deep Learning Model, 31 (2019) 223–230.

[28] Z. Huang, Nasrullah, J. Wen, S. Song, M. Mateen, Fundus Image Classification Using VGG-19
Architecture with PCA and SVD, Symmetry (Basel). 11 (2018) 1. doi:10.3390/sym11010001.

p ro
[29] S. Ruder, An overview of gradient descent optimization, (2016) 1–14.

[30] I. Loshchilov, F. Hutter, SGDR: Stochastic Gradient Descent with Warm Restarts, (2017) 1–
16.

[31] M.D. Zeiler, ADADELTA: An Adaptive Learning Rate Method, (2012).

doi:http://doi.acm.org.ezproxy.lib.ucf.edu/10.1145/1830483.1830503.
e-
[32] R. Shindjalova, K. Prodanova, V. Svechtarov, Modeling data for tilted implants in grafted with
bio-oss maxillary sinuses using logistic regression, AIP Conf. Proc. 1631 (2014) 58–62.
doi:10.1063/1.4902458.
Pr
[33] A. Wibowo, P.W. Wiryawan, N.I. Nuqoyati, Optimization of neural network for cancer
microRNA biomarkers classification, J. Phys. Conf. Ser. 1217 (2019) 012124.
doi:10.1088/1742-6596/1217/1/012124.

[34] P. Vamplew, R. Dazeley, C. Foale, Neurocomputing Softmax exploration strategies for

multiobjective reinforcement learning, Neurocomputing. 263 (2017) 74–86.
al

doi:10.1016/j.neucom.2016.09.141.

[35] M. Toğaçar, B. Ergen, Deep Learning Approach for Classification of Breast Cancer, in: 2018
Int. Conf. Artif. Intell. Data Process., 2018: pp. 1–5. doi:10.1109/IDAP.2018.8620802.
urn

[36] Z. Cömert, A.F. Kocamaz, Fetal Hypoxia Detection Based on Deep Convolutional Neural
Network with Transfer Learning Approach, in: R. Silhavy (Ed.), Softw. Eng. Algorithms Intell.
Syst., Springer International Publishing, Cham, 2019: pp. 239–248.

[37] B. Choubin, S. Khalighi-Sigaroodi, A. Malekian, Ö. Kişi, Multiple linear regression, multi-

layer perceptron network and adaptive neuro-fuzzy inference system for forecasting
Jo

precipitation based on large-scale climate signals, Hydrol. Sci. J. 61 (2016) 1001–1009.

doi:10.1080/02626667.2014.966721.

[38] T. Wang, C. Wen, H. Wang, F. Gao, T. Jiang, S. Jin, Deep Learning for Wireless Physical
29
 Journal Pre-proof

Layer : Opportunities and Challenges, (2017). doi:10.1109/CC.2017.8233654.

[39] Y. Altuntaş, Z. Cömert, A.F. Kocamaz, Identification of haploid and diploid maize seeds using
convolutional neural networks and a transfer learning approach, Comput. Electron. Agric. 163

of
(2019) 104874. doi:https://doi.org/10.1016/j.compag.2019.104874.

[40] M. G Alaslani, L. A. Elrefaei, Convolutional Neural Network Based Feature Extraction for
IRIS Recognition, Int. J. Comput. Sci. Inf. Technol. 10 (2018) 65–78.

p ro
doi:10.5121/ijcsit.2018.10206.

[41] Z. Cömert, A. Şengür, Ü. Budak, A.F. Kocamaz, Prediction of intrapartum fetal hypoxia
considering feature selection algorithms and machine learning models, Heal. Inf. Sci. Syst. 7
(2019) 17. doi:10.1007/s13755-019-0079-z.

[42] V. Suárez-paniagua, I. Segura-bedmar, Evaluation of pooling operations in convolutional

architectures for drug-drug interaction extraction, 19 (2018).
e-
doi:https://doi.org/10.1186/s12859-018-2195-1.

[43] M. Geist, Soft-max boosting, Mach. Learn. 2958 (2015) 305–332. doi:10.1007/s10994-015-
5491-2.
Pr
[44] S. Woo, J. Park, J.-Y. Lee, I.S. Kweon, CBAM: Convolutional Block Attention Module,
Eccv2018. (n.d.).

[45] S. Xie, R. Girshick, P. Dollár, Z. Tu, K. He, Aggregated Residual Transformations for Deep
Neural Networks, (2016). doi:10.1109/CVPR.2017.634.
al

[46] B. Hariharan, P. Arbeláez, R. Girshick, J. Malik, Object Instance Segmentation and Fine-
Grained Localization Using Hypercolumns, IEEE Trans. Pattern Anal. Mach. Intell. 39 (2017)
627–639. doi:10.1109/TPAMI.2016.2578328.
urn

[47] S. Ioffe, C. Szegedy, Batch Normalization: Accelerating Deep Network Training by Reducing
Internal Covariate Shift, (2015). http://arxiv.org/abs/1502.03167.

[48] J. Park, S. Woo, J.-Y. Lee, I.S. Kweon, BAM: Bottleneck Attention Module, (2018).
http://arxiv.org/abs/1807.06514.

[49] L. Chen, H. Zhang, J. Xiao, L. Nie, J. Shao, T.-S. Chua, SCA-CNN: Spatial and Channel-wise
Jo

Attention in Convolutional Networks for Image Captioning, 2016.

[50] N. Schilling, M. Wistuba, L. Drumond, L. Schmidt-Thieme, Hyperparameter optimization with

factorized multilayer perceptrons, Lect. Notes Comput. Sci. (Including Subser. Lect. Notes
30
 Journal Pre-proof

Artif. Intell. Lect. Notes Bioinformatics). 9285 (2015) 87–103. doi:10.1007/978-3-319-23525-

7_6.

[51] H. Wen, J. Shi, W. Chen, Z. Liu, Deep Residual Network Predicts Cortical Representation and

of
Organization of Visual Features for Rapid Categorization, Sci. Rep. 8 (2018) 3752.
doi:10.1038/s41598-018-22160-9.

[52] A. Nugaliyadde, K.W. Wong, F. Sohel, H. Xie, Language Modeling through Long Term

p ro
Memory Network, (2019). http://arxiv.org/abs/1904.08936.

[53] P. Reeskamp, Is comparative advertising a trade mark issue ?, Eur. Intellect. Prop. Rev. 30
(2008) 130–137. doi:10.1145/2623330.2623612.

[54] Z. Yang, C. Wang, Z. Zhang, J. Li, Mini-batch algorithms with online step size, Knowledge-
Based Syst. 165 (2019) 228–240. doi:https://doi.org/10.1016/j.knosys.2018.11.031.

[55]
e-
D.M.W. Powers, Ailab, Evaluation: From Precision, Recall and F-Measure To Roc,
Informedness, Markedness & Correlation, 2 (2011) 37–63. doi:10.9735/2229-3981.

[56] Z. Cömert, A.F. Kocamaz, V. Subha, Prognostic model based on image-based time-frequency
Pr
features and genetic algorithm for fetal hypoxia assessment, Comput. Biol. Med. (2018).
doi:10.1016/j.compbiomed.2018.06.003.

[57] Shallu, R. Mehra, Breast cancer histology images classification: Training from scratch or
transfer learning?, ICT Express. 4 (2018) 247–254. doi:10.1016/j.icte.2018.10.007.

[58] Y. Feng, L. Zhang, J. Mo, Deep Manifold Preserving Autoencoder for Classifying Breast
al

Cancer Histopathological Images, IEEE/ACM Trans. Comput. Biol. Bioinforma. PP (2018) 1.

doi:10.1109/TCBB.2018.2858763.

[59] G. Yun, Y. Jie, Densely-Connected Multi-Magnification Hashing for Histopathological Image

urn

Retrieval, IEEE J. Biomed. Heal. Informatics. XX (2018) 1–10.

doi:10.1109/JBHI.2018.2882647.

[60] H. Erfankhah, M. Yazdi, M. Babaie, H.R. Tizhoosh, Heterogeneity-Aware Local Binary

Patterns for Retrieval of Histopathology Images, IEEE Access. 7 (2019) 18354–18367.
doi:10.1109/ACCESS.2019.2897281.
Jo

[61] D. Lichtblau, C. Stoean, Cancer diagnosis through a tandem of classifiers for digitized
histopathological slides, PLoS One. 14 (2019) 1–20. doi:10.1371/journal.pone.0209274.

[62] Z. Han, B. Wei, Y. Zheng, Y. Yin, K. Li, S. Li, Breast Cancer Multi-classification from
31
 Journal Pre-proof

Histopathological Images with Structured Deep Learning Model, Sci. Rep. 7 (2017) 1–10.
doi:10.1038/s41598-017-04075-z.

of
e- p ro
Pr
al
urn
Jo

32
 Journal Pre-proof

Credit Author Statement

Credit Author Statement

All authors mentioned in this article were equally involved in all stages of the article.

of
e- p ro
Pr
al
urn
Jo