Hindawi

Wireless Communications and Mobile Computing

Volume 2022, Article ID 5089078, 10 pages
https://doi.org/10.1155/2022/5089078

Research Article
A Novel CNN-Inception-V4-Based Hybrid Approach for
Classification of Breast Cancer in Mammogram Images

Muhammad Saquib Nazir,1 Usman Ghani Khan ,1 Aqsa Mohiyuddin,2

Mana Saleh Al Reshan,3 Asadullah Shaikh ,3 Muhammad Rizwan,4
and Monika Davidekova5
1
Department of Computer Science, University of Engineering and Technology, Lahore 54890, Pakistan
2
Department of Computer Science, Kinnaird College for Women, Lahore 54890, Pakistan
3
College of Computer Science and Information Systems, Najran University, Najran 61441, Saudi Arabia
4
Secure Cyber Systems Research Group, WMG, University of Warwick, Coventry CV4 7AL, UK
5
Information Systems Department Faculty of Management, Comenius University in Bratislava, Odbojárov 10,
82005 Bratislava 25, Slovakia

Correspondence should be addressed to Usman Ghani Khan; [email protected]

Received 30 April 2022; Revised 13 June 2022; Accepted 18 June 2022; Published 6 July 2022

Academic Editor: Kuruva Lakshmanna

Copyright © 2022 Muhammad Saquib Nazir et al. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work
is properly cited.

Breast cancer is the most frequent disease in women, with one in every 19 women at risk. Breast cancer is the fifth leading cause of
cancer death in women around the world. The most effective and efficient technique of controlling cancer development is early
identification. Mammography helps in the early detection of cancer, which saves lives. Many studies conducted various tests to
categorize the tumor and obtained positive findings. However, there are certain limits. Mass categorization in mammography is
still a problem, although it is critical in aiding radiologists in establishing correct diagnoses. The purpose of this study is to
develop a unique hybrid technique to identify breast cancer mass pictures as benign or malignant. The combination of two
networks helps accelerate the categorization process. This study proposes a novel-based hybrid approach, CNN-Inception-V4,
based on the fusing of these two networks. Mass images are used in this research from the CBIS-DDSM dataset. 450 images
are taken for benign, and 450 images are used for malignant. The images are first cleaned by removing pectoral muscles, labels,
and white borders. Then, CLAHE is used to these images to improve their quality in order to produce promising classification
results. Following preprocessing, our model classifies cancer in mammography pictures as benign or malignant abnormalities.
Our proposed model’s accuracy is 99.2%, with sensitivity of 99.8%, specificity of 96.3%, and F1-score of 97%. We also
compared our proposed model to CNN, Inception-V4, and ResNet-50. Our proposed model outperforms existing classification
models, according to the results.

1. Introduction oped countries, the proportion of cancer cases is growing. This

is a major issue in healthcare. As a result, early identification of
Breast cancer becomes one of the most serious diseases and breast cancer can lower the risk of patient death [4].
leading causes of mortality in women worldwide [1]. Breast Mammography is indeed the good diagnostic imaging
cancer develops when aberrant cells in the breast proliferate technique for detecting breast problems early on [5]. Cancer
excessively [2]. Breast cancer is classified into two types: benign mortality rates could rise by 50% to 15 million by 2030,
and malignant. Benign is treatable, but it can become hazardous according to World Health Organization (WHO) predic-
over time. Malignant is perilous [3]. It has the potential to tions. Developing countries account for over 70% of new
spread to other organs of the body and cause death. In devel- cancer cases. X-ray scans of each breast are acquired from
2 Wireless Communications and Mobile Computing

two perspectives during a mammography screening. Human (i) To propose such a unique strategy for properly clas-
readers, such as an expert radiologist, assess and examine sifying breast cancer
screening mammograms [6]. Dual reading was discovered
(ii) To propose a hybrid technique that can speed up
to enhance evaluation results and has been implemented in
the process of breast cancer classification along with
many countries. Repeated readings with up to ten readers
good results
can boost diagnostic performance even further. Beyond the
dual reading, there is much of room for improvement in (iii) To perform techniques such as image preprocess-
mammography grading. ing, data augmentation, and hyperparameters (e.g.,
Based on the invasive phase of the breast anomaly, the batch size, learning rate), which are appealing
irregularities may be benign or malignant. Detecting lumps choices for reducing overfitting and enhancing
in breast tissue is more difficult than detecting microcalcifi- model generalization
cation. Many industrialised countries are adopting the smart
cities concept in the modern era [7]. All of that in a smart (iv) To compare our model to existing state-of-the-art
city is controlled by a computer. Every industry in the world networks
is moving toward the computer. Humans want computers to The other sections of the research are organized in the
rule everything. In a smart city environment, decision- following manner: related work is done in Section 2, the pro-
making is a significant difficulty [8]. This present era is posed method and dataset used are described in Section 3,
sometimes referred to as the data world. Because consumers the classification results are evaluated in Section 4, and the
contribute data on a daily basis. This information is quite conclusions are presented in the last section.
beneficial in terms of making the environment smarter.
Data-driven decision AI (artificial intelligence) [9, 10]
enables computers to make decisions. Data must be available 2. Related Work
to the computer in order for it to make conclusions.
Healthcare is a major global challenge [11–15]. The use Many researchers used various strategies to reach promising
of computers, smart systems, and intelligent devices is criti- findings. The authors in [28] employed computer vision
cal in healthcare. The diagnosis of sickness at an early stage techniques to help a radiologist evaluate mammography
can save lives [16, 17]. New breakthroughs have the poten- images and diagnose a malignancy in an acceptable amount
tial to lessen the risk to human life [18]. of time. They extracted various characteristics from the
Computer-Aided Diagnosis (CAD) [19] systems are used to mammogram’s selected area, which the surgeon subjectively
assist radiologists in the detection and diagnosis of cancer. CAD labeled. These characteristics are supplied into a classifica-
systems have been shown in studies to be useful; yet, accurate tion engine, which is then used to train and construct the
breast cancer detection remains a difficult issue [20]. Because suggested architecture classification models. The classifica-
of the intricacy of mammography and the large number of tests tion procedure by either SVM or NB with its optimum
performed by radiologist, incorrect diagnoses occur. The CAD parameters resulted in 100% accuracy for the classification
system attempts to reduce incorrect diagnoses, which in turn process. The authors in [29] proposed Fisher linear discrim-
reduces unneeded biopsies, patient anxiety, increased health- ination analysis (FLDA) and quadratic discrimination
care costs, and superfluous review by radiologists. Moving assessment (QDA) inside random projection (RP) filtration
beyond traditional machine learning (ML) [21, 22], current to evaluate our performance and efficiency with SVM classi-
efforts have focused on the application of deep learning (DL) fier (SVM), K-nearest neighbour (KNN), and multilayer
approaches [22–24]. Through statistical analysis, they are able perceptron (MLP) in IoT-based platforms. The modelling
to more correctly diagnose suspected lesions. DL models are results demonstrate not only increased accuracy but also
intended for huge and diverse datasets (for example, the Ima- decreased duration. In our chosen dataset simulated for a
geNet dataset [25]), whereas mammography contains tiny data- relation to health and safety, runtime is reduced by 20 times
sets that are publically available. As a result, when learning from when compared to SVM, KNN, and MLP in QDA and
the start with minimal training samples, the capability and com- FLDA. At the IoT business, RP filtering in the preprocessing
plexity of such networks might have a considerable negative step of feature extraction, which fulfils the model’s runtime,
impact on simulation results. is a point of view.
Many researchers played an important role in the classifi- The authors in [30] proposed Blockchain-Integrated
cation of mass images of breast cancer as benign and malig- Remote Database (BIRD) paradigm that connects block-
nant, but still, mass classification in mammography remains chain to a Relational Database Management System
a significant issue, yet it is essential in assisting radiologists (RDBMS) that employs the Remote Database Access
in making accurate diagnoses [26]. There is also a need for (RDA) protocol and a Cloud Server. This report gives a
such a system that can speed up the process of classification comprehensive history of blockchain and industry genera-
of tumors with good results, and this is possible with a unique tions. Finally, this research proposes a next-generation
hybrid technique. So, using the CBIS-DDSM (Curated Breast blockchain architecture in which a smart government linked
Imaging Subset of DDSM) dataset [27], this research provides to Industry 4.0 monitors price increases and fraud.
a novel hybrid approach CNN-Inception-V4 for classifying The authors in [31] proposed synchronization technique
breast cancer as benign or malignant. The paper’s main contri- based on AI to address smart grid timing difficulties. Back-
butions are based on the following points. propagation neural networks are used as an AI technology
Wireless Communications and Mobile Computing 3

that utilizes time approximations and error corrections for The authors in [34] presented a model for breast cancer
exact results. In order to link the external time server, the categorization, using the CAD technique. MIAS and DDSM
unique AIFS system considers radio communication fea- datasets are used in this study. Preprocessing, categorization,
tures. The suggested AIFS scheme’s performance is tested gathering, and operation grouping are all used in this sys-
using a MATLAB-based simulation technique. The simula- tem. The CAD system includes a CNN model with eight
tion results suggest that the proposed approach outperforms coevolutionary, 4 max pooling, and two totally connected
the present system. layers. The obtained findings are compared to AlexNet and
The authors in [27] proposed a modified network of VGG16, demonstrating that the suggested CNN [35] beat
YOLOv5 to detect and categorize breast cancers. The process both of these models in terms of accuracy and AUC. The
was carried out with an 8-batch size, an activation function suggested model achieved accuracies of 92.54%, 96.47%,
of 0.01, a velocity of 0.843, and an interval value of 300. To and 95% for MIAS, DDSM, and the self-collected dataset,
evaluate the suggested effectiveness of the algorithm, it is respectively.
contrasted to YOLOv3 and a faster RCNN, with 93.50% The authors in [36] developed an intelligent system that
MCC value and 96.50% accuracy. The suggested model beats included picture normalization, microcalcification detection,
YOLOv3 and quicker RCNN. The results show that the pro- and mammogram categorization. The best classification
posed prediction model detects and diagnoses breast tumors accuracy of the system in distinguishing between benign
while also addressing previous researchers’ limitations by and malignant breasts can reach 0.8124 on the training set
decreasing the FPR and FNR and raising the MCC value. and 0.7237 on the test set, with an area under curve
The authors in [32] proposed an efficient, lightweight (AUC) of receiver operating characteristic of 0.8042. The
encryption method to provide a safe picture encryption sensitivity of malignant breast prediction is 0.8891 on the
approach for the healthcare business. To protect medical training dataset and 0.7778 on the testing dataset. The high-
pictures, the proposed lightweight encryption solution uses est accuracy was 0.7511 on the microcalcification lesion
two permutation algorithms. In terms of security and execu- degree, with an AUC of 0.7627 on the testing dataset.
tion time, the suggested approach is investigated, assessed, The authors in [37] proposed a strategy to categorize
and then compared to commonly encrypted ones. Several breast invasive cancer using AGAN for data augmentation
test photos were utilized to evaluate the proposed algo- and CNN for tumor categorization. In many visual recogni-
rithm’s performance. Several studies demonstrate that the tion applications, deep CNN models and other predictive
suggested algorithm for picture cryptosystems is more effi- analysis and ML algorithms have outperformed the state of
cient than existing techniques. the art. This method produced 89.17% accurate results but
The authors in [33] present a study of deep learning- a 19.41% true alarm rate of classification.
based load forecasting strategies from 2015 to 2020. This The authors in [38] investigated the achievement of
study addresses research in the load forecasting process that three types of ML methods: (1) NB, (2) NN, and (3) SVM.
uses deep learning techniques, Distributed Deep Learning SVM delivers the most satisfactory results in correctly char-
(DDL) techniques, backpropagation- (BP-) based works, acterising the condition of the breast as “Good” or “Malig-
and non-BP-based works. Following the survey, it was con- nant,” with a reliability of 99.4% in training and 98.76% in
cluded that data aggregation dependence would be advanta- testing, according to the outcomes of ten experiments. In
geous for lowering computing time in load forecasting. As a the study, the SVM classifier outperformed the NN and
result, a conceptual model of DDL for smart grids has been NB models, suggesting that SVM is an appropriate choice
provided, with the HSIC (Hilbert-Schmidt Independence for assessing the state of the breast at a preliminary phase.
Criterion) bottleneck approach integrated to give more The authors in [39] proposed a method to identify breast
accuracy. tumor utilizing two deep learning model networks, Res-Net-
The authors in [32] offer an experimental analysis of 50 and VGG-16 using the IRMA dataset and then compar-
cryptographic algorithms in order to categorize encryption ing their outcomes. They attained an accuracy of 94% for
methods as symmetric or asymmetric. It provides a thor- VGG-16 and 91.7% for Res-Net50, as well as a sensitivity
ough examination of AES, DES, 3DES, RSA, and Blowfish of 94% for VGG-16 and 94% for Res-Net50. Their restric-
in terms of time complexity, size, encryption, and decryption tion is that for VGG16 and Res-Net50, their accuracy value
performance. The simulation technique was used for the lowers to 89% and 88%, respectively.
evaluation, and the speed of encryption and decryption of The authors in [40] proposed a hierarchical twitter sen-
the chosen encryption algorithms was examined. The exper- timent model (HTSM) to show people’s opinions in short
iments ran the same encryption using the same plaintext five texts. The suggested model’s quality is confirmed by apply-
times for each encryption and decryption, and the average ing it to a popular product and a widely discussed issue.
time is compared. Each encryption algorithm’s key size is The findings reveal that the suggested model beats state-of-
the maximum number of bytes that the cypher may accept. the-art approaches for successfully evaluating people’s opin-
The average time taken by the three devices to compute ions in brief text.
the method is utilized for comparison. In the simulation The authors in [41] provide a novel CNN structure for
for the experimental test, a set of plaintexts—password- classifying histopathological cancer pictures that integrates
sized text and paragraph-sized text—achieves goal fair out- spectral characteristics acquired via a multiresolution wave-
comes when compared to the existing methods in real-time let technique with CNN spatial features. Furthermore, batch
deep learning networks for IoT applications. normalization is employed after each layer in the
4 Wireless Communications and Mobile Computing

convolution network to solve CNN’s poor convergence term “fuzzy” refers to the employment of a clustering
problem, and CNN’s deep layers are trained with spectral– algorithm to segment a breast tumor. The Bi-LSTM
spatial data. Using the BreaKHis dataset and the Breast Can- method generates EHO to diagnose cancer with perfor-
cer Classification Challenge 2015 dataset, the suggested mance characteristics. In this study, the MIAS dataset is
structure is evaluated on malignant histology pictures of employed. The results are then contrasted to CNN and
the breast for both binary and multiclass tissue classifica- DCNN, as well as Bi-LSTM, with EHOBi-LSTM beating
tions. Experiment findings suggest that combining spec- everyone else, although FPR and FNR levels need to be
tral–spatial data enhances CNN network classification decreased. The authors in [51] suggested that system is a
accuracy and needs fewer training parameters than well- combination stages: pretreatment and development of a
known models. On both datasets, the suggested structure Convolution Neural Network (CNN). The preparation
achieves an average accuracy of 97.58% and 97.45% with process sets the mammograms data for the CNN building
7.6 million training parameters. phase, which encompasses format unity, denoising, image
The authors in [42] describe how the VGG16 frame- enhancement, ROI retrieval, augmentation, and image
work of DL methodology was used to categorize breast scaling. In the CNN development phase, the proposed
cancer in mammography. The VGG16 classifier is con- CNN model is built from the ground up in addition to
structed and tested using 322 photographs from the MIAS learning features and identify tumor progression in mam-
dataset. It beats the models AlexNet, EfficientNet, and mograms data.
GoogleNet. The categorization of mammograms will According to the discussion in related work, several
improve mammography screening effectiveness and act as researchers were critical in classifying mass pictures of breast
a valuable mechanism for surgeons. The authors in [43] cancer as benign or malignant. Many studies suggested var-
promptly made available breast cancer diagnoses at a pre- ious CNN models with varying parameters and obtained
liminary phase of malignancy. The authors in [44] results ranging from 90 to 98%. However, there are other
employed quicker R-CNN to identify abnormalities in constraints that must be addressed. The major gap in mam-
mammograms to categorize masses into benign and mography is mass classification using such a model that can
malignant. categorize it effectively and quickly. It is possible to do this
The authors in [45] utilize image fragmentation and lim- using a hybrid strategy. This study uses the CBIS-DDSM
ited preprocessing methods to classify cancer. In semantic dataset to develop a unique hybrid approach CNN-
segmentation, binarization is used to compare the edge Inception-V4 for identifying breast cancer masses as benign
detection strategy to the RCNN methodology. The authors or malignant.
in [46] employed a standard diversity CNN and then incor-
porated two improvement techniques: (i) batch normaliza-
tion (BN) and (ii) drop-outs (DO). They employed rank- 3. Proposed Methodology
based stochastic pooling. As a consequence, BDR-CNN
was developed, which is a hybrid of CNN, BN, DO, and In this research, we developed a novel hybrid strategy for
RSP. This BDR-CNN was hybridised with a GCN to create classifying breast cancer as benign or malignant by combin-
our BDR-CNN-GCN framework, which was then deployed ing CNN and Inception-V4 networks as showing in Figure 1.
as a data augmentation strategy for breast mammography The CBIS-DDSM dataset, which contains mammography
assessment. The authors in [47] explained HRD, a combina- images, is used in this study. 450 images are utilized for
tion model of a radial basis function network with DT. This benign, and 450 images are used for malignant. The images
approach was evaluated versus three prominent algorithms: are first cleaned by removing pectoral muscles, labels, and
kNN, NB, and SVM. The observations reveal that their rec- white borders. Then, CLAHE is used to these images to
ommended technique achieved a high level of precision. The improve their quality in order to produce promising classifi-
authors in [48] created a collection with chaotic tags or cor- cation results.
rupted data and contrasted AlexNet against GoogleNet (GN)
for cancer identification. The proportion of chaotic training
pictures was changed to control the equilibrium of their sys- sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
tem’ memory and cognitive. C = Cmin + 2a2 ln , ð1Þ
The authors in [49] developed a unique hybrid optimal 1 − Pimage ðiÞ
feature selection (HOFS) approach to figure out the rele-
vant characteristics to achieve maximum efficiency for this
classification. To train the neural network, a number of whereC min is the lowest limit of the data point and a2 is a
carefully chosen characteristics are used. In this proposed distribution adjustment factor that is determined for each
technique, the searchable relevant information from the input picture as showing in Equation (1) [52]. CNN is a
minimammographic data capture organization (MIAS) form of neural network model that helps us to extract more
collection was included. The authors in [50] develop a accurate representations of picture material. Unlike tradi-
framework for assessing cancer in mammograms. Chroma tional image recognition, which requires you to define the
noise, fragmentation, and clustering are among the phases image characteristics, CNN processes the input pixel data
in the suggested model. The image enhancement is used to from the picture, trains the model, and then extracts the fea-
improve the image quality from digital mammograms. The ture representation for better categorization.
Wireless Communications and Mobile Computing 5

Pre-processing
BN+ BN+
Removal of PM & labels ReLU Dropout ReLU
Feature map +Stride Feature map 25% +Stride
32 × 32 28 × 28 14 × 14
Feature map
Feature map
10 × 10
5×5

3 × 3 Conv
149 × 149 × 32
(32 2V)

CLAHE
Raw 3 × 3 Conv
data Removal of borders 147 × 147 × 32
(32)

Input Convolution Ma×pool Convolution Ma×pool

5×5 5×5 3×3 2×2 3 × 3 Conv
147 × 147 × 64
(64)

3 × 3 Conv 3 × 3 Ma×pool
(2V) (96 2V)

73 × 73 × 160 Filter concat.

1 × 3 Conv
1 × 1 Conv 1 × 1 Conv

(256)

1 × 1 Conv

7 × 1 Conv

7 × 1 Conv
3 × 1 Conv

1 × 3 Conv

1 × 7 Conv

1 × 7 Conv

1 × 1 Conv

1 × 1 Conv
(64) (64)

3 × 3 Conv

3 × 3 Conv
(384)

(256)

(224)
(512)

(448)

(224)

(192)

(192)

(64)
(96)

(96)
7 × 1 Conv
(64)
3 × 1 Conv
(256)

3 × 1 Conv
1 × 7 Conv

1 × 7 Conv

7 × 1 Conv

1 × 1 Conv

3 × 3 Conv

1 × 1 Conv
(256)
(64)

(256)

(224)

(192)
Filter concat.

(96)

(64)
Filter concat.
Filter concat.

Filter concat.

Filter concat.

Filter concat.
Final output

1 × 1 Conv
Softmax

3 × 3 Conv 3 × 3 Conv

(384)
1 × 3 Conv
(96V) (96V)

(256)

1 × 1 Conv

1 × 1 Conv
(384)

(64)
71 × 71 × 192 Filter concat.
1 × 1 Conv
(256)

Ma×pool 3 × 3 Conv
(2V) (192V)
1 × 1 Conv

1 × 1 Conv

1 × 1 Conv
AvgPool

AvgPool

AvgPool
(256)

(128)

(96)
35 × 35 × 384 Filter concat.

Dropout ReLU Dropout FL+

50% 25% ReLU

Figure 1: The proposed network with all the steps including preprocessing, network fusion, and then Softmax classification (CNN-
Inception-V4 hybrid approach).

Images are cleaned and enhanced before being fed into effectively.
the hybrid model using filters 32 × 32. 8 9 8 9
> x0 > > z0 0 0 0 0 0 >
>
> > > >
> >
>
>
> > > >8 9
0 >
>
> x1 >
>
>
>
>
>
>
> z1 z0 0 0 0 0> >> y0 >
>
>> >
>
> >
> >
> >
>>
> >
>
Z = 〠 Z ðiÞðZ − ZiÞ, ð2Þ >
> >
> >
> >
>>
> >
>
> x2
> > > >
> z2 z1 z0 0 0 0> >>
> y1 >
>
i=0 >
> >
> >
> >
>>
> >
>
>
< x3 >
= >
<0 >> >
z2 z1 z0 0 0 =< y2 =
= : ð6Þ
where Z is the sequence such that Zð0Þ = 1 as mentioned in >
> x4 >
> >
> 0 0 z2 z1 z0 0> >>
> y3 >
>
>
> >
> >
> >
>>
> > >
>
> > > >
> >
>>
> >
>
Equation (2). Equations (3)–(5) [53] show the impulse >
> x5 >
> >
> 0 0 0 z2 z1 z0 >
>>
> y4 >
>
>
> >
> >
> >
>>
> >
>
response. >
> >
> >
> >
>>
: >
;
>
> x6 >
> >
> 0 0 0 0 z2 >
>
z1 > y5
>
> > >
> >
: >
;
>
>
:
>
>
;
Y = Oðxðt ÞÞ, ð3Þ x7 0 0 0 0 0 z2

X is an 8-dimensional vector, Z is a 3-dimensional vector,

and y is a 6-dimensional vector. Then, we make an 8 × 6
! matrix that encodes the values of Z so that X = T ðZÞ y as
0
Y = O 〠 Z ðiÞðZ − ZiÞ , ð4Þ shown in Equation (6). To lower data size and processing time,
i=0 CNN employs max pooling to replace output with a max sum-
mary. This helps you to identify the elements that have the
greatest influence while reducing the danger of overfitting.
0
Max pooling takes two hyperparameters into account: stride
Y = 〠 Z ðiÞOðZ − ZiÞ: ð5Þ and size.
i=0
6
Axðt Þ = 〠 AyðT ÞAz ðt − T Þ, ð7Þ
The information is sent to the following stage, which i=1
extracts features from Convolution 5 × 5 (feature map 28 ×
28) using batch normalization [54]+ReLU and stride. A con- 0
volution runs the window over pictures and computes the Axðt Þ = 〠 AyðT ÞAz ðt − T Þ, ð8Þ
i=0
input and filtration system dot product pixel values. This
enables convolution to highlight the essential characteristics. Axðt Þ = ðAy ∗ Az Þðt Þ: ð9Þ
The data is then transferred to max pool 5 × 5 (feature map
14 × 14) with a 25% dropout rate to extract features more Even in this discrete example, the cumulative distribution
6 Wireless Communications and Mobile Computing

Table 1: Testing and training images of benign and malignant. Inception-v4 is a convolutional neural network structure
that improves on prior generations of the Inception family
Training Testing by simplifying the design and employing additional incep-
Benign 380 70 tion modules than Inception-v3. Inception-v4, which
Malignant 380 70 evolved from GoogLeNet/Inception-v1, has a more consis-
tent streamlined design and more inception modules than
Inception-v3. This network has numerous blocks, including
Model accuracy the Stem block, Inception-A block, Inception-B block, and
Inception-C block, as well as reduction blocks.
0.95 The first block is the stem of the Inception V4 network,
0.90 in which input is transmitted to a convolution of 32 2 V fil-
ters with 149 × 149 × 32 size, then to 3 × 3 convolution with
0.85 147 × 147 × 32 size, and finally to 147 × 147 × 64 with con-
Accuracy

0.80 volution of 64 kernels.

0.75
l l
0.70
yAxy = 〠 〠 J uf biðx,uÞ, jðy, f Þ : ð10Þ
0.65 u=1 f =1

0 10 20 30 40 50
Epoch Another image matrix A is the convolution of the image
b with the kernel J as shown in Equation (10). It then splits
Train
into two branches. The first is the max pooling branch,
Test
which has 96 2 V, and the second is the 3 × 3 convolution
Figure 2: Accuracy graph of training and testing of CNN- branch, which has 2 V. The output of these two branches is
Inception-V4. 73 × 73 × 160 in size. Then, it splits into two branches once
more. The first is 1 × 1 convolution (64), while the second
is1 × 1convolution (64). The first branch advances to various
convolution layers, including 7 × 1 convolution (64), 7 × 1
Model loss
convolution (64), and 3 × 3 convolution (96 V). The second
3.0
branch goes to 3 × 3 convolution (96 V). The output is
concatenated once again with 71 × 71 × 192 dimensions.
2.5 Following that, it separates into two branches once more.
The first is max pool with a voltage of 2 V, and the second
2.0
is 3 × 3 convolution with a voltage of 192 V, and the output
Loss

1.5 from these two branches concatenates to a 35 × 35 × 384

size. The following block is Inception-A, in which the output
1.0 from the stem is sent into the Inception-A block through the
concatenation function. Then, it proceeds to four branches.
0.5
The first is 1 × 1 convolution (64), the second is 1 × 1 convo-
lution (96), and the fourth is AvgPooling. The input from
0 10 20 30 40 50 the first branch is routed to 3 × 3 convolution with 96 and
Epoch
then to 3 × 3 convolution (96). The input from the second
branch is routed to 3 × 3 convolution (96), while the input
Train loss
from the fourth branch is sent to 1 × 1 convolution (96).
Test loss
Then, using the filter concatenation function, all inputs from
Figure 3: Loss graph of training and testing of CNN-Inception-V4. subbranches are concatenated. The input from this block is
sent to the Inception-B block, which also contains branches
and subbranches as shown in Equations (11)–(14). Input
passes to 1 × 7 convolution (192) from the first branch 1 ×
functions ofX, Y, and Z (also known as expected values) are 1 convolution (192), then to 7 × 1 convolution (224), then
shown in Equations (7)–(9) [52]. The stride determines the to 1 × 7 convolution (224), and and finally to convolution
skipping of value pools, whereas the length determines the (256). From the second branch, the input is sent to 1 × 1
amount of the value pools in each skip. The collected features convolution (192), then to 7 × 1 convolution (224), and
are then fed into a convolution of 3 × 3 with 10 × 10 feature finally to convolution (256). 1 × 1 convolution is the third
map with batch normalization+ReLU and stride. The results branch (384).
of the convolution layer was then routed to a max pool layer
of 2 × 2 with a 5 × 5 feature map. Inception-V4 is the result
of the merger of various networks. T ðA = GÞ = T ðI + J = GÞ, ð11Þ
Wireless Communications and Mobile Computing 7

100.00
90.00
80.00
70.00
60.00
(%)

50.00
40.00
30.00
20.00
10.00
0.00
CNN-InceptionV4 CNN InceptionV4 ResNet-50

Accuracy Specificity
Sensitivity F1-Score

Figure 4: Comparison between our proposed model, CNN, Inception-V4, and ResNet-50 on the basis of accuracy, sensitivity, specificity,
and F1-score.

Predicted result: Benign Predicted result: Malignant Predicted result: Benign Predicted result: Benign
Actual result: Benign Actual result: Malignant Actual result: Benign Actual result: Benign
0 0 0 0
25 25 25 25
50 50 50 50
75 75 75 75
100 100 100 100
125 125 125 125
150 150 150 150
175 175 175 175
200 200 200 200

0 50 100 150 200 0 50 100 150 200 0 50 100 150 200 0 50 100 150 200

Predicted result: Benign Predicted result: Malignant Predicted result: Benign Predicted result: Benign
Actual result: Benign Actual result: Malignant Actual result: Benign Actual result: Benign
0 0 0 0
25 25 25 25
50 50 50 50
75 75 75 75
100 100 100 100
125 125 125 125
150 150 150 150
175 175 175 175
200 200 200 200

0 50 100 150 200 0 50 100 150 200 0 50 100 150 200 0 50 100 150 200

Figure 5: The results showing the actual and predicted values of benign and malignant.

6 The input from this block is sent to the Inception-B

T ðA = GÞ = 〠 T ðI + J = G ∣ J = uÞT ð J = uÞ, ð12Þ block, which also contains branches and subbranches as
u=1 showing in Equations (11)–(14). Input passes to 1 × 7 con-
volution (192) from the first branch 1 × 1 convolution
6
(192), then to 7 × 1 convolution (224), then to 1 × 7 convo-
T ðA = GÞ = 〠 T ðI + u = GÞT ð J = uÞ, ð13Þ
lution (224), and finally to convolution (256). From the sec-
u=1
ond branch, the input is sent to 1 × 1 convolution (192),
6 then to 7 × 1 convolution (224), and finally to convolution
T ðA = GÞ = 〠 T ð J = uÞT ðI = G − uÞ: ð14Þ (256). 1 × 1 convolution is the third branch (384). Input is
u=1 sent to 1 × 1 convolution from the fourth branch of
8 Wireless Communications and Mobile Computing

AvgPooling (128). Then, using the filter concatenation func- The F1-score is a mix of the model’s accuracy and
tion, all inputs from subbranches are concatenated. The fol- recall as shown in the following equation:
lowing block is Inception-C. The input is routed through
multiple branches and subbranches in this block. From the
first branch, 1 × 1 convolution (384), input is passed to 1 × 2TP
F1‐score = : ð18Þ
3 convolution (448), then to 3 × 1 convolution (512), then 2TP + FN + FP
to 1 × 3 convolution (256), and finally to 3 × 1 convolution
(256). 1 × 1 convolution (384), 3 × 1 convolution (256), 1 ×
3 convolution (256), and finally to convolution (256). Input Our suggested model’s accuracy is 99.2%, with a sensi-
is sent to 1 × 1 convolution from the third branch of Avg- tivity of 99.8%, specificity of 96.3%, and F1-score of 97%.
Pooling (256). Then, using the filter concatenation function, We compared our proposed model with CNN, Incep-
all inputs from subbranches are concatenated. After going tion-V4, and ResNet-50. Our proposed model outperforms
through CNN and the Inception model, the input enters existing classification models, according to the findings as
the classification phase, which has a 25% dropout rate and shown in Figure 2. The training and testing accuracy
a ReLU activation function. and loss of CNN-Inception-V4 are shown in Figures 3
After passing through classification neurons, it employs and 4, respectively. The x-axis shows the number of
Softmax for binary classification and then predicts benign epochs and the y-axis shows the accuracy for the model.
and aggressive breast cancer. This suggested network was Training results are shown in blue color, and testing
trained and tested using 50 epochs and 32 batch sizes, yield- results are shown in orange shade.
ing encouraging findings that are reported in the next Our proposed model 99.2% accurately classifies breast
section. cancer into benign and malignant. The results showing
actual and predicted values of benign and malignant are
4. Results and Discussion shown in Figure 5.

In this study, we used the CBIS-DDSM dataset to classify

5. Conclusion and Future Work
breast cancer as benign or malignant. And, as described in
Proposed Methodology, we employed a CNN-Inception-V4 Many researchers played an important role in the classifica-
hybrid methodology for this purpose. 450 images are tion of mass images of breast cancer as benign and malig-
employed for benign, and 450 images are used for nant, but still, mass classification in mammography
malignant. remains a significant issue, yet it is essential in assisting radi-
Testing and training images of benign and malignant are ologists in making accurate diagnoses. There is also a need
shown in Table 1. for such a system that can speed up the process of classifica-
50 epochs and 32 batch size for training and testing tion of tumors with good results, and this is possible with a
are used. We utilized accuracy, sensitivity [55], specificity, unique hybrid technique. So, using the CBIS-DDSM
and F-score [55] as evaluation metrics. Accuracy is the (Curated Breast Imaging Subset of DDSM) dataset, this
overall prediction of model as shown in the following research provides a novel hybrid approach CNN-
equation: Inception-V4 for classifying breast cancer as benign or
malignant. The CBIS-DDSM dataset, which comprises
mammography images, is employed. 450 images are used
TP + TN
Accuracy = : ð15Þ for benign, and 450 images are used for malignant. The
TP + TN + FP + FN images are first cleaned by removing pectoral muscles, labels,
and white borders. Then, CLAHE is used to these images to
improve their quality in order to produce promising classifi-
Sensitivity is defined as the proportion of accurately cation results. Following preprocessing, our model diagnoses
identified real positives as shown in the following equa- cancer in mammography pictures as benign or malignant
tion: abnormalities (MGI). To lower data size and processing
time, CNN employs max pooling to replace output with a
max summary. This helps you to identify the elements that
TP have the greatest influence while reducing the danger of
Sensitivity = : ð16Þ
TP + FN overfitting. 50 epochs and 32 batch size for training and test-
ing are used. We utilized accuracy, sensitivity, specificity,
and F-score as evaluation metrics. Our suggested model’s
Specificity is defined as the proportion of accurately accuracy is 99.2%, with a sensitivity of 99.8%, specificity of
identified real negatives as shown in the following equa- 96.3%, and F1-score of 97%. We also compared our pro-
tion: posed model to CNN, Inception-V4, and ResNet-50. Our
proposed model outperforms existing classification models,
according to the outcomes. In our research, we employed
TN mass images. As a result, calcification images can be used
Specificity = : ð17Þ
TN + FP for further study in the future.
Wireless Communications and Mobile Computing 9

Data Availability [12] F. Akram, D. Liu, P. Zhao, N. Kryvinska, S. Abbas, and

M. Rizwan, “Trustworthy intrusion detection in e-healthcare
The data used to support the findings of the study are systems,” Frontiers in public health, vol. 9, 2021.
included within the article. [13] G. Thippa Reddy, A. Srivatsava, K. Lakshmanna, R. Kaluri,
S. Karnam, and G. Nagaraja, “Risk prediction to examine
health status with real and synthetic datasets,” Biomedical
Conflicts of Interest and Pharmacology Journal, vol. 10, no. 4, pp. 1897–1903, 2017.
[14] T. M. Ali, A. Nawaz, A. U. Rehman et al., “A sequential
The authors declare that they have no conflicts of interest. machine learning-cum-attention mechanism for effective seg-
mentation of brain tumor. front,” Oncol, vol. 12, p. 873268,
2022.
Acknowledgments [15] S. U. Rehman, A. R. Javed, M. U. Khan, M. N. Awan,
A. Farukh, and A. Hussien, “Personalisedcomfort: a persona-
This research was supported by the Information Systems lised thermal comfort model to predict thermal sensation votes
Department, Faculty of Management Comenius University for smart building residents,” Enterprise Information Systems,
in Bratislava, Odbojárov 10, 82005 Bratislava 25, Slovakia. pp. 1–23, 2020.
[16] M. U. Sarwar and A. R. Javed, “Collaborative health care plan
through crowdsource data using ambient application,” in 2019
References 22nd international multitopic conference (INMIC), pp. 1–6,
IEEE, 2019.
[1] L. Dorling, S. Carvalho, J. Allen et al., “Breast cancer risk Genes
[17] A. Priyanka, M. Parimala, K. Sudheer, R. Kaluri,
- Association analysis in more than 113,000 women,” The New
K. Lakshmanna, and M. P. K. Reddy, “Big data based on
England journal of medicine, vol. 384, no. 5, pp. 428–439, 2021.
healthcare analysis using IoT devices,” in IOP Conference
[2] S. Abbas, Z. Jalil, A. R. Javed et al., “BCD-WERT: a novel Series: Materials Science and Engineering, vol. 263, IOP Pub-
approach for breast cancer detection using whale optimization lishing, p. 042059, 2017.
based efficient features and extremely randomized tree algo-
[18] M. K. Hasan, S. Islam, I. Memon et al., “A novel resource ori-
rithm,” PeerJ Computer Science, vol. 7, p. e390, 2021.
ented DMA framework for Internet of Medical Things devices
[3] W. Gao, S. Zhang, J. Guo et al., “Investigation of synthetic in 5G network,” IEEE Transactions on Industrial Informatics,
relaxometry and diffusion measures in the differentiation of p. 1, 2022.
benign and malignant breast lesions as compared to BI-
[19] J. Shi, A. Vakanski, M. Xian, J. Ding, and C. Ning, “EMT-NET:
RADS,” Journal of Magnetic Resonance Imaging, vol. 53,
efficient multitask network for computer-aided diagnosis of
no. 4, pp. 1118–1127, 2021.
breast cancer,” 2022, http://arxiv.org/abs/2201.04795.
[4] K. Kerlikowske, S. Chen, M. K. Golmakani et al., “Cumulative
advanced breast cancer risk prediction model developed in a [20] M. Mehmood, M. Rizwan, S. Abbas et al., “Machine learning
screening mammography population,” Journal of the National assisted cervical cancer detection,” Frontiers in public health,
Cancer Institute, vol. 114, no. 5, pp. 676–685, 2022. vol. 9, 2021.
[5] J. T. Schousboe, B. L. Sprague, L. Abraham et al., “Cost-effec- [21] A. R. Javed, L. G. Fahad, A. A. Farhan et al., “Automated cog-
tiveness of screening mammography beyond age 75 years,” nitive health assessment in smart homes using machine learn-
Annals of Internal Medicine, vol. 175, no. 1, pp. 11–19, 2022. ing,” Sustainable Cities and Society, vol. 65, p. 102572, 2021.
[6] S. E. Hickman, R. Woitek, E. P. V. Le et al., “Machine learning [22] M. Rizwan, A. Shabbir, A. R. Javed et al., “Risk monitoring
for workflow applications in screening mammography: sys- strategy for confidentiality of healthcare information,” Com-
tematic review and meta-analysis,” Radiology, vol. 302, no. 1, puters and Electrical Engineering, vol. 100, p. 107833, 2022.
pp. 88–104, 2022. [23] S. Bhattacharya, S. R. K. Somayaji, T. R. Gadekallu, M. Alazab,
[7] D. A. Wood, S. Kafiabadi, A. Al Busaidi et al., “Deep learning and P. K. R. Maddikunta, “A review on deep learning for
to automate the labelling of head MRI datasets for computer future smart cities,” Internet Technology Letters, vol. 5, no. 1,
vision applications,” European Radiology, vol. 32, no. 1, p. e187, 2022.
pp. 725–736, 2022. [24] A. R. Javed, M. U. Sarwar, S. Khan, C. Iwendi, M. Mittal, and
[8] A. R. Javed, R. Faheem, M. Asim, T. Baker, and M. O. Beg, “A N. Kumar, “Analyzing the effectiveness and contribution of
smartphone sensors-based personalized human activity recog- each axis of tri-axial accelerometer sensor for accurate activity
nition system for sustainable smart cities,” Sustainable Cities recognition,” Sensors, vol. 20, no. 8, p. 2216, 2020.
and Society, vol. 71, p. 102970, 2021. [25] N. Tomasev, I. Bica, B. McWilliams et al., “Pushing the limits
[9] G. Kantor, V. Niarchos, and C. Papageorgakis, “Solving con- of self-supervised ResNets: can we outperform supervised
formal field theories with artificial intelligence,” 2021, http:// learning without labels on ImageNet?,” 2022, http://arxiv
arxiv.org/abs/2108.08859. .org/abs/2201.05119.
[10] A. R. Javed, M. U. Sarwar, H. U. Khan, Y. D. Al-Otaibi, and W. S. [26] S. Safdar, M. Rizwan, T. R. Gadekallu et al., “Bio-imaging-
Alnumay, “Pp-spa: privacy preserved smartphone-based personal based machine learning algorithm for breast cancer detection,”
assistant to improve routine life functioning of cognitive impaired Diagnostics, vol. 12, no. 5, p. 1134, 2022.
individuals,” Neural Processing Letters, pp. 1–18, 2021. [27] A. Mohiyuddin, A. Basharat, U. Ghani et al., “Breast tumor
[11] A. R. Javed, M. U. Sarwar, M. O. Beg, M. Asim, T. Baker, and detection and classification in mammogram images using
H. Tawfik, “A collaborative healthcare framework for shared modified YOLOv5 network,” Computational and Mathemati-
healthcare plan with ambient intelligence,” Human-centric Com- cal Methods in Medicine, vol. 2022, Article ID 1359019, 16
puting and Information Sciences, vol. 10, no. 1, pp. 1–21, 2020. pages, 2022.
10 Wireless Communications and Mobile Computing

[28] M. M. Alshammari, A. Almuhanna, and J. Alhiyafi, “Mammog- [43] D. Singh and A. K. Singh, “Role of image thermography in
raphy image-based diagnosis of breast cancer using machine early breast cancer detection-past, present and future,” Com-
learning: a pilot study,” Sensors, vol. 22, no. 1, p. 203, 2022. puter methods and programs in biomedicine, vol. 183,
[29] M. K. Hasan, T. M. Ghazal, A. Alkhalifah et al., “Fischer linear p. 105074, 2020.
discrimination and quadratic discrimination analysis–based [44] S. Hadush, Y. Girmay, A. Sinamo, and G. Hagos, “Breast can-
data mining technique for internet of things framework for cer detection using convolutional neural networks,” 2020,
healthcare,” Frontiers in Public Health, vol. 9, 2021. http://arxiv.org/abs/2003.07911.
[30] M. K. Hasan, A. Md, S. Rayhan Kabir et al., “Evolution of [45] M. Pavithra, R. Rajmohan, T. A. Kumar, and R. Ramya, “Pre-
industry and blockchain era: monitoring price hike and cor- diction and classification of breast cancer using discriminative
ruption using BIoT for smart government and Industry 4.0,” learning models and techniques,” in Machine Vision Inspec-
IEEE Transactions on Industrial Informatics, 2022. tion Systems, vol. Volume 2, pp. 241–262, Machine Learning-
[31] M. K. Hasan, M. M. Ahmed, A. H. A. Hashim, A. Razzaque, Based Approaches, 2021.
S. Islam, and B. Pandey, “A novel artificial intelligence based [46] Y.-D. Zhang, S. C. Satapathy, D. S. Guttery, J. M. Gorriz, and
timing synchronization scheme for smart grid applications,” S.-H. Wang, “Improved breast cancer classification through
Wireless Personal Communications, vol. 114, no. 2, pp. 1067– combining graph convolutional network and convolutional
1084, 2020. neural network,” Information Processing & Management,
[32] M. K. Hasan, S. Islam, R. Sulaiman et al., “Lightweight encryp- vol. 58, no. 2, p. 102439, 2021.
tion technique to enhance medical image security on internet [47] S. Annamalai, R. Udendhran, and M. Balamurgan, “Hybrid-
of medical things applications,” IEEE Access, vol. 9, ized neural network and decision tree based classifier for prog-
pp. 47731–47742, 2021. nostic decision making in breast cancers,” Soft Computing,
[33] M. Akhtaruzzaman, M. K. Hasan, S. R. Kabir, S. N. H. S. vol. 24, no. 11, pp. 7947–7953, 2020.
Abdullah, M. J. Sadeq, and E. Hossain, “HSIC bottleneck based [48] R. K. Samala, H.-P. Chan, L. M. Hadjiiski, M. A. Helvie, and
distributed deep learning model for load forecasting in smart C. D. Richter, “Generalization error analysis for deep convolu-
grid with a comprehensive survey,” IEEE Access, vol. 8, tional neural network with transfer learning in breast cancer
pp. 222977–223008, 2020. diagnosis,” Physics in Medicine & Biology, vol. 65, no. 10,
[34] V. S. Gnanasekaran, S. Joypaul, P. M. Sundaram, and D. D. p. 105002, 2020.
Chairman, “Deep learning algorithm for breast masses classifi- [49] J. J. Patel and S. K. Hadia, “An enhancement of mammogram
cation in mammograms,” IET Image Processing, vol. 14, no. 12, images for breast cancer classification using artificial neural
pp. 2860–2868, 2020. networks,” IAES International Journal of Artificial Intelligence,
[35] E. H. Houssein, M. Hassaballah, I. E. Ibrahim, D. S. AbdElmi- vol. 10, no. 2, p. 332, 2021.
naam, and Y. M. Wazery, “An automatic arrhythmia classifica- [50] M. M. Amalmary and A. Prakash, Elephant herding optimiza-
tion model based on improved marine predators algorithm tion with bi-directional long shortterm memory network
and convolutions neural networks,” Expert Systems with Appli- (EHOBI-LSTM) for breast cancer detection in digital
cations, vol. 187, p. 115936, 2022. mammograms.
[36] Q. Lin, W.-M. Tan, S.-Y. Zhu et al., “Deep learning-based micro- [51] M. F. EnasEl, “Houby and Nisreen IR Yassin. “Malignant and
calcification detection and classification of mammography for nonmalignant classification of breast lesions in mammograms
diagnosis of breast cancer,” 2022, Available at SSRN 4001825. using convolutional neural networks”,” Biomedical Signal Pro-
[37] B. Swiderski, L. Gielata, P. Olszewski, S. Osowski, and cessing and Control, vol. 70, p. 102954, 2021.
M. Kołodziej, “Deep neural system for supporting tumor rec- [52] A. M. Reza, “Realization of the contrast limited adaptive histo-
ognition of mammograms using modified GAN,” Expert Sys- gram equalization (CLAHE) for real-time image enhance-
tems with Applications, vol. 164, p. 113968, 2021. ment,” Journal of VLSI signal processing systems for signal,
[38] S. M. Samuri, T. V. Nova, B. Rahmatullah, S. L. Wang, and image and video technology, vol. 38, no. 1, pp. 35–44, 2004.
Z. T. Al-Qaysi, “Classification model for breast cancer mam- [53] S. SimaSahu, A. K. Singh, S. P. Ghrera, and M. Elhoseny, “An
mograms,” IIUM Engineering Journal, vol. 23, no. 1, pp. 187– approach for de-noising and contrast enhancement of retinal
199, 2022. fundus image using CLAHE,” Optics & Laser Technology,
[39] N. S. Ismail and C. Sovuthy, “Breast cancer detection based vol. 110, pp. 87–98, 2019.
on deep learning technique,” in 2019 International UNIMAS [54] O. Junghun, H. Kim, S. Baik, C. Hong, and K. M. Lee, “Batch
STEM 12th Engineering Conference (EnCon), pp. 89–92, normalization tells you which filter is important,” in Proceed-
IEEE, 2019. ings of the IEEE/CVF Winter Conference on Applications of
[40] A. F. Ibrahim, M. Hassaballah, A. A. Ali, Y. Nam, and I. A. Computer Vision, pp. 2645–2654, 2022.
Ibrahim, “COVID19 outbreak: a hierarchical framework for [55] A. Kovacs, P. Palasti, D. Vereb, B. Bozsik, A. Palkó, and Z. T.
user sentiment analysis,” Computers, Materials, & Continua, Kincses, “The sensitivity and specificity of chest CT in the
pp. 2507–2524, 2022. diagnosis of COVID-19,” European Radiology, vol. 31, no. 5,
[41] H. K. Mewada, A. V. Patel, M. Hassaballah, M. H. Alkinani, pp. 2819–2824, 2021.
and K. Mahant, “Spectral–spatial features integrated convolu-
tion neural network for breast cancer classification,” Sensors,
vol. 20, no. 17, p. 4747, 2020.
[42] A. Pillai, A. Nizam, M. Joshee, A. Pinto, and S. Chavan, “Breast
cancer detection in mammograms using deep learning,” in
Applied Information Processing Systems, pp. 121–127,
Springer, 2022.