diagnostics

Article
A Modified LeNet CNN for Breast Cancer Diagnosis in
Ultrasound Images
Sathiyabhama Balasubramaniam 1, *, Yuvarajan Velmurugan 1 , Dhayanithi Jaganathan 1
and Seshathiri Dhanasekaran 2, *

1 Computer Science and Engineering, Sona College of Technology, Salem 636005, India;
[email protected] (Y.V.); [email protected] (D.J.)
2 Department of Computer Science, UiT The Arctic University of Norway, 9037 Tromso, Norway
* Correspondence: [email protected] (S.B.); [email protected] (S.D.)

Abstract: Convolutional neural networks (CNNs) have been extensively utilized in medical image
processing to automatically extract meaningful features and classify various medical conditions,
enabling faster and more accurate diagnoses. In this paper, LeNet, a classic CNN architecture,
has been successfully applied to breast cancer data analysis. It demonstrates its ability to extract
discriminative features and classify malignant and benign tumors with high accuracy, thereby
supporting early detection and diagnosis of breast cancer. LeNet with corrected Rectified Linear Unit
(ReLU), a modification of the traditional ReLU activation function, has been found to improve the
performance of LeNet in breast cancer data analysis tasks via addressing the “dying ReLU” problem
and enhancing the discriminative power of the extracted features. This has led to more accurate,
reliable breast cancer detection and diagnosis and improved patient outcomes. Batch normalization
improves the performance and training stability of small and shallow CNN architecture like LeNet.
It helps to mitigate the effects of internal covariate shift, which refers to the change in the distribution
of network activations during training. This classifier will lessen the overfitting problem and reduce
the running time. The designed classifier is evaluated against the benchmarking deep learning
models, proving that this has produced a higher recognition rate. The accuracy of the breast image
recognition rate is 89.91%. This model will achieve better performance in segmentation, feature
extraction, classification, and breast cancer tumor detection.
Citation: Balasubramaniam, S.;
Velmurugan, Y.; Jaganathan, D.; Keywords: deep learning; breast cancer; convolutional neural networks; LeNet; medical image
Dhanasekaran, S. A Modified LeNet
processing; batch normalization
CNN for Breast Cancer Diagnosis in
Ultrasound Images. Diagnostics 2023,
13, 2746. https://doi.org/10.3390/
diagnostics13172746
1. Introduction
Academic Editor: Dechang Chen
Breast cancer is one of the most common causes of death in women worldwide.
Received: 19 April 2023 Early detection saves the life of many women. Healthcare practitioners extensively use
Revised: 6 July 2023 mammograms for screening breast cancer. Breast ultrasound and diagnostic mammogram
Accepted: 11 July 2023 are the two imaging tests that are used to evaluate breast tissue for abnormalities or signs
Published: 24 August 2023 of breast cancer. While both are imaging studies of the breast, they differ in how they use
sound and X-rays to obtain images and the information they provide [1].
A diagnostic mammogram is a low-dose X-ray that provides detailed images of the
breast tissue. It can detect calcifications, masses, and abnormalities indicating cancer or
Copyright: © 2023 by the authors.
other conditions. Diagnostic mammograms are typically recommended for women with
Licensee MDPI, Basel, Switzerland.
a breast lump, nipple discharge, and women at risk of breast cancer [2,3] due to family
This article is an open access article
history or other related diseases.
distributed under the terms and
Hubbard et al. [4] have presented a cohort study report which discusses breast cancer
conditions of the Creative Commons
risk assessment, screening guidelines in average-risk women, and certain controversies
Attribution (CC BY) license (https://
surrounding breast cancer screening. It presents recommendations for using a framework
creativecommons.org/licenses/by/
4.0/).
of shared decision-making to assist women in balancing their values regarding the benefits

Diagnostics 2023, 13, 2746. https://doi.org/10.3390/diagnostics13172746 https://www.mdpi.com/journal/diagnostics

Diagnostics 2023, 13, 2746 2 of 28

and harms of screening at various ages and intervals to make personal screening choices
within a range of reasonable options. It includes recommendations for women at elevated
risk and discusses new technologies like tomosynthesis [4].
A breast ultrasound normally uses a high-pitched sound to create an image of the
breast. It can provide more detailed information about the structure of a breast mass, such
as whether it is a fluid-filled cyst or a solid lump. Ultrasound is often used as a follow-up
test to mammography as a screening [3,5].
Early diagnosis and accurate diagnosis are important for effective treatment. Computer-
aided diagnosis (CAD) is a method that uses machine learning and deep learning algorithms
to help doctors diagnose cancer [6]. CAD can be used at all stages of cancer screening, in-
cluding mammography, ultrasound, and magnetic resonance imaging (MRI). CAD systems
analyze images and provide radiologists with visualization and diagnostic information,
helping them identify abnormalities that might be overlooked. CAD can detect calcifica-
tions, masses, and other breast cancer symptoms. The system then generates a report of the
findings and recommends action. The stages in the table CAD system are preprocessing,
segmentation, feature extraction, and classification. Additionally, the segmentation and
identification of stages are difficult for researchers to diagnose breast cancer accurately.
Therefore, advanced tools and techniques are required to accurately diagnose and classify
breast cancer cases [7].
CAD is used to support radiologists and healthcare practitioners to make accurate
diagnoses. In addition, it helps them to suggest and recommend suitable prescriptions and
treatments for care. Researchers report the effectiveness of CAD in cancer diagnosis and
find it to have better sensitivity and specificity. In addition, CAD can significantly reduce
the time required for radiologists and healthcare practitioners to interpret images [8].
CAD is also being used to predict breast cancer risk. By analyzing factors such as age,
family history, and breast density, CAD can generate a personalized risk assessment that
can help guide screening and treatment decisions. CAD is a valuable tool in breast cancer
analysis that improves breast cancer detection accuracy, reduces unnecessary biopsies, and
treatments, and provides personalized risk assessments [5]. As technology advances, CAD
will be increasingly important in breast cancer diagnosis and management.
By leveraging the ability of deep learning networks to learn intricate representations,
these networks can achieve high accuracy rates in breast cancer detection, aiding in early
diagnosis and reducing false positives and false negatives. Deep learning has succeeded
highly in medical image processing, Computer Vision (CV), natural language processing,
and video/speech recognition. Deep learning networks can assist in predicting patient
outcomes and tailoring treatment plans based on individual characteristics. These networks
can identify different breast cancer subtypes and stages via analyzing large-scale clinical
and test data. This information can guide oncologists in selecting appropriate therapies
and optimizing treatment strategies. Breast cancer screening and diagnosis can be time-
consuming and labor-intensive for radiologists and pathologists. Deep learning networks
can automate various stages of the diagnostic workflow, such as image analysis, lesion
segmentation, and classification, leading to increased efficiency, reduced human error, and
potentially faster turnaround times. Breast cancer analytics using deep learning networks
enhance accuracy, efficiency, and provide personalized care in breast cancer detection,
diagnosis, and treatment, ultimately improving patient outcomes and saving lives [8].
Arevalo, J. et al. [9] have proposed a hybrid representation learning framework for
breast cancer diagnosis in mammography that integrated convolutional neural networks
(CNNs) to learn discriminative features automatically. They have applied a biopsy-proven
benchmarking dataset (344 breast cancer patients) containing 736 film mammography
(mediolateral oblique and craniocaudal) views, representing manually segmented lesions
associated with masses: 426 benign and 310 malignant lesions. The developed method
comprises two main stages: (i) preprocessing to enhance image details and (ii) super-
vised training for learning both the features and the breast imaging lesions classifier in
Diagnostics 2023, 13, 2746 3 of 28

a supervised way instead of designing particular descriptors to explain the content of

mammography images.
Rampun, A. et al. [10] have explored ensemble deep learning for breast mass classifi-
cation in mammograms. The authors have modified the Alex Net to adapt it to the breast
mass classification problem. A model selection is performed to select the best three results
based on the highest validation accuracies. Then, the prediction is made based on the
average probability of the models. The results show that accuracy from individual models
ranges between 75% and 77%. However, the ensemble networks provide 80% accuracy and
area under the curve.
Choosing appropriate algorithms for image enhancement, segmentation, feature ex-
traction, feature selection, and prediction is important for cancer diagnosis. It is an im-
portant research area where many researchers have analyzed these datasets using deep
learning algorithms [11,12]. Several deep learning algorithms can be applied to breast
cancer data analytics. These algorithms leverage the power of neural networks to analyze
complex patterns in the data and make predictions. Convolutional neural networks (CNN)
are commonly used in deep learning algorithms for breast cancer data analytics. The spe-
cific choice of algorithm depends on the nature of the data, the research or clinical question,
and the available computational resources [13]. The features can be selected after feature
selection and then classified into normal, benign, and malignant classes. By leveraging the
hierarchical nature of CNNs, the complex patterns and structures can be analyzed within
medical images, aiding in tasks such as tumor detection, disease classification, and image
segmentation [13].
The primary objectives of breast cancer classification are:
• To accurately distinguish between malignant (cancerous) and benign (non-cancerous)
breast lesions.
• To improve treatment outcomes and survival rates, using the LeNet model to assist in
identifying subtle abnormalities in breast images that may indicate the presence of
cancer at an early stage.
• To reduce false positives and false negatives, the LeNet model can be trained to strike
a balance between sensitivity (detecting true positive cases) and specificity (avoiding
false positives) and enhance the accuracy of breast cancer classification.
• To provide prognostic information, predict the likelihood of disease progression and
patient survival, and to help assess the risk of recurrence and guide decisions regarding
post-treatment surveillance and follow-up care.
• To recommend tailoring treatment approaches to individual patients. By classifying
tumors based on specific molecular markers or genetic profiles, personalized treatment
plans can be developed and therapeutic outcomes can be optimized.
• It is important to note that the modified LeNet can serve as a foundation for breast
cancer classification. The architecture offers increased model depth, enabling the
learning of more complex features, and potentially achieving higher classification
accuracy.

2. Literature Review
Breast cancer is one of the most common diseases among women, and early detection
is a crucial step to save the lives of millions of women. Due to the recent developments in
healthcare, breast cancer patients can be diagnosed at an early stage. More importantly,
conventional methods of analyzing breast cancer images suffer from high false detection
rates. Different breast cancer imaging modalities are used to extract and analyze the key
features affecting the diagnosis and treatment of breast cancer. These imaging modalities
can be divided into subgroups such as mammograms, ultrasound, magnetic resonance
imaging, histopathological images, or any combination [14]. Radiologists or pathologists
analyze images produced by these methods manually, which leads to an increase in the risk
of wrong decisions for cancer detection. Thus, the utilization of new automatic methods to
Diagnostics 2023, 13, 2746 4 of 28

analyze all kinds of breast screening images to assist radiologists in interpreting images is
required.
Artificial intelligence (AI) has recently been widely utilized to automatically improve
the early detection and treatment of different types of cancer, specifically breast cancer,
thereby enhancing the survival chance of patients. Advances in AI algorithms, such as deep
learning, and the availability of datasets obtained from various imaging modalities have
opened an opportunity to surpass the limitations of current breast cancer analysis methods.
In this article, breast cancer imaging modalities, and their strengths and limitations are
discussed. Then, the most recent studies that have employed AI in breast cancer detection
and various breast imaging modalities have been summarized. In addition to this, the
paper presents the datasets on the breast-cancer imaging modalities, which are important in
developing AI-based algorithms and through offering training of deep learning models [14].
Recently, there has been a great demand for the classification of breast cancer using
ultrasound images and mammography [15]. Mammograms, ultrasound scans, and MRI
imaging are specific imaging modalities in breast cancer screening and diagnosis that
radiologists and oncologists majorly use. CAD can further support healthcare practitioners
in detecting and diagnosing breast cancer. Early detection of pectoral muscles becomes
necessary in reducing the ambiguities which could provide movements to the upper limbs
or ribs. Their recognition is essential in enhancing the diagnostic performance of breast
cancer detection.
A deep convolutional neural network is used to classify the 1.2 million high-resolution
images in the ImageNet [16]. On the test data, top-1 and top-5 error rates are 37.5% and
17.0%, respectively. In the neural network, five convolutional layers are followed by max-
pooling layers, and three fully connected layers with a final 1000-way SoftMax. To make
training faster, they used GPU implementation of the convolution operation. To reduce
overfitting, the authors used a “dropout” regularization method that proved very effective.
They are the winners of top-5 test error rate of 15.3% in the ILSVRC-2012 competition,
compared to 26.2% achieved by the second-best entry [16].
The authors [17] have designed a deep feature-based framework for breast mass clas-
sification using CNN and a decision mechanism. The combined intensity information and
deep features have been automatically extracted using the trained CNN. Then, classifiers
are used to predict classes of test images. Their method is applied to the DDSM dataset
and achieved high accuracy and classification performance.
Vidhushavarshini et al. [18] have designed a simple machine learning classifier for
analyzing thyroid datasets. This model used the patients’ history to predict the disease via
creating the knowledge prediction model, easing the medical practitioner’s job with better
accuracy.
Zahoor et al. [19] have designed a CAD-based classification model to prevent the
disease as well as to reduce the risk of breast cancer in women. The CAD system’s accuracy
was improved via reducing the false-positive rates. The modified entropy whale optimiza-
tion algorithm (MEWOA) was used to optimize the features. The fine-tuned MobilenetV2
and NASNet Mobile are applied for simulation. Then, the machine learning classifiers are
applied to classify breast cancer images using the optimized deep features. Three publicly
available datasets are used to extract the features and perform the classification: INbreast,
MIAS, and CBIS-DDSM. Their classification accuracies are 99.7%, 99.8%, and 93.8% and
their performance is reasonably good compared to other approaches [19].
Deep learning networks in breast cancer analytics have emerged as a promising
approach to improve early detection, diagnosis, and personalized treatment. Deep learning
networks excel in analyzing such vast amounts of data, allowing for improved detection
and classification of breast cancer. Deep learning networks can automatically learn complex
patterns and features from breast images without explicit feature engineering. Breast cancer
detection and diagnosis often require the analysis of subtle, and localized abnormalities,
which can be challenging for traditional machine learning methods [6].
Diagnostics 2023, 13, 2746 5 of 28

The researchers have applied appropriate pre-processing steps and have improved
the input images in contrast to improving the target detection accuracy [13]. For example,
Sadad et al. [20] have developed a CAD model to diagnose cancer and to reconstruct the
images using Hilbert Transform (H.T.) for better visualization. Next, they have segmented
the tumor using a marker-controlled basin transformation. In the next step, shape and
texture features were extracted and classified using a hybrid k-nearest neighbors algorithm
(KNN) and cluster decision tree model. Badawi et al. [21] have used fuzzy logic in the
first step and defined the region of interest using a semantic segmentation method. Then,
they have used eight pre-trained models for accurate tumor classification. The authors
have used image processing, machine learning, and deep learning to diagnose cancer [21].
They have used reliable methods to detect breast cancer early to save women’s lives:
microcalcifications and the malignant cells CAD system for breast cancer detection. The
test results show that the CAD system can detect breast cancer initially. Segmentation and
distribution of stages in this model are unsuitable for breast cancer diagnosis.
Zahoor et al. [22] have presented image processing, machine learning, and deep
learning methods to diagnose breast cancer. The authors have highlighted better choices
and reliable methods to diagnose breast cancer in the initial stages to save women’s
lives . Additionally, several CAD methodologies were discussed to detect breast masses,
microcalcifications, and malignant cells [22]. Lévy, D. and Jain [23] have designed a hybrid
model combining CNN and transfer learning to classify breast masses in mammograms
as benign or malignant. The authors added pre-processing and data augmentation to
overcome the inadequate training data issues. They achieved a better performance on the
DDSM dataset [23].
Ting et al. [24] have presented an improved CNN for breast cancer classification to
assist the medical practitioners. This improved CNN model classifies the breast lesions into
benign, malignant, and normal with 90.5% accuracy, 89.47% sensitivity, 90.71% specificity,
and 0.901 receiver operating characteristics [24]. Vidhushavarshini S et al. [25] have pro-
posed a hybrid optimization algorithm-based feature selection design for thyroid disease
classification with rough type-2 fuzzy support vector machine. This work combined the
firefly algorithm (FA) and the butterfly optimization algorithm (BOA) namely hybrid firefly
butterfly optimization-rough type-2 fuzzy support vector machine (HFBO-RT2FSVM) to
select the top-n relevant features. HFBO-RT2FSVM is evaluated with several key metrics
such as specificity, accuracy, and sensitivity. We compare our approach with well-known
benchmark methods such as improved gray wolf optimization linear support vector ma-
chine (IGWO Linear SVM) and mixed-kernel support vector machine (MKSVM) methods.
The HFBO-RT2FSVM technique improved the classification accuracy and thereby precision
in thyroid disease identification. The HFBO-RT2FSVM model attained an accuracy of
99.28%, having specificity and sensitivity of 98% and 99.2%, respectively.
Dense breast tissue is an independent risk factor for breast cancer and it lowers
the sensitivity of screening mammography [26]. Hence, medical practitioners rely on
supplemental screening like ultrasound or MRI to improve breast cancer detection rate.
Supplemental screening is also influenced by a history of breast biopsy and family history
of breast cancer, race, age, socioeconomic status, density category, and physician’s specialty
and region [26].
Feature extraction is a very important step when mammography image analysis is
addressed using learning-based approaches [27]. Automated feature extraction is used to
represent the content of images with CNN, and the classifier is evaluated to learn features
from mammography mass lesions. Empirical results prove that this approach is effective
in identifying the target, and that its classification accuracy ranges from 79.9% to 86% in
terms of area under the ROC curve [27].
The Chan–Vese level set method [28] is used to extract the initial contour of mammo-
grams, and a deep learning CNN (DL-CNN) is used to learn the features of mammary-
specific mass and microcalcification clusters. A fully complex-valued relaxation network is
used in the last stage of DL-CNN network to increase the classification accuracy. Experi-
Diagnostics 2023, 13, 2746 6 of 28

ments are conducted using the standard benchmarking breast cancer datasets MIAS and
the Breast Cancer Digital Repository (BCDR). The results show that the proposed method
has significantly improved the performance over the traditional methods. Performance
measures such as accuracy, sensitivity, specificity, AUC achieved are 99%, 0.9875, 1.0 and
0.9815, respectively, and the effective classification of the mammogram images has been as
normal, benign, or malignant and its subclasses [28].
Chai et al. [29] have provided a review of recent achievements in terms of techniques
and applications in CV methods. The authors have identified the emerging techniques and
have investigated their applications in the scenarios, including recognition, visual tracking,
semantic segmentation, and image restoration. They have also discussed the future research
directions, prospective growth, and impact of these technologies in various domains. The
summarization, knowledge accumulation, and creation would benefit researchers in these
domains and participators in the CV industries [29].
N. Borah et al. [30] have applied automation based on ViT compared to the existing
systems; the effectiveness of the proposed model is in relation to mammogram images
(IN breast database). Real-time performance provides 96.48% accuracy and requires little
training time to analyze medical images. The suggested model is built using a graphical
user interface (GUI), which might help medical professionals make wiser judgments and
identify B.C. more quickly [30].
Many hybrid optimization algorithms are trapped in local optima and have slow
convergence speeds, which reduces the classification accuracy. To resolve these issues, a
hybrid optimization algorithm that combines the grasshopper optimization algorithm and
the crow search algorithm for feature selection and classification of the breast mass with
multilayer perceptron has been developed [31]. The simulation is performed using MAT-
LAB 2019a. The model’s efficacy is compared with other related optimization algorithms,
and the results are better to a reasonable degree compared to the other models in terms
of classification accuracy (97.1%), sensitivity (98%), and specificity (95.4%) for the MIAS
dataset [31].
Manikandan et al. [32] have studied clinical, epidemiology, and end outcome datasets
to distinguish between cancer cases and deaths. This study presents a machine learning-
based approach to classify SEER breast cancer data. In addition to this, SEER breast cancer
data have been selected for analysis using a two-step selection method based on baseline
differences and values. After selecting the features, Ada, XG, gradient, naive Bayes, and
decision tree supervised and ensemble learning techniques are used to classify the breast
cancer dataset. Decision trees have the highest accuracy (98%) for train–test split and
cross-validation [33].
Arikidis, N. et al. [11] have experimented with a two-stage semiautomated segmenta-
tion method of microcalcification (MC) clusters. In the first stage, efficient segmentation
of most of the particles of a MC cluster, and shape refinement of selected individual MCs
is performed in the second stage. The effect of the proposed segmentation method on
MC cluster characterization accuracy was evaluated in a case sample of 162 pleomorphic
MC clusters (72 malignant and 90 benign). Ten MC cluster features, targeted to capture
morphologic properties of individual MCs in a cluster (area, major length, perimeter, com-
pactness, and spread), were extracted and a correlation-based feature selection method
yielded a feature subset to feed in a support vector machine classifier. Classification per-
formance of the MC cluster features was estimated by means of the area under receiver
operating characteristic curve utilizing tenfold cross-validation methodology. Interob-
server and intraobserver segmentation agreements (median and [25%, 75%] quartile range)
were substantial with respect to the distance metrics HDISTcluster (2.3 [1.8, 2.9] and 2.5
pixels) and AMINDISTcluster (0.8 [0.6, 1.0] and 1.0 [0.8, 1.2] pixels), while moderate with
respect to the AOMcluster (0.64 [0.55, 0.71] and 0.59 [0.52, 0.66]). This method outper-
formed (0.80 ± 0.04) statistically significantly (Mann–Whitney U-test, p < 0.05) the B-spline
active rays segmentation method (0.69 ± 0.04), suggesting the significance of the proposed
semiautomated method.
Diagnostics 2023, 13, 2746 7 of 28

Early detection of breast cancer ensures appropriate treatment and survival, and breast
cancer diagnosis is mainly based on histopathological images [33]. Therefore, CAD is used
for the automatic identification and diagnosis of cancer to help doctors diagnose cancer.
The authors use five ConvNets: ResNet, VGG19, VGG16, Xception, and MobileNet. Their
approach supports limiting the hardware needed to complete the large-scale ConvNets
training task. It outperforms other handcrafted features for histopathology images.
Various solutions based on CV have been proposed in the literature [34]. These
solutions have been unsuccessful due to large video sequences that need to be processed in
surveillance systems. The problem arises in the presence of multi-view cameras. Recently,
the development of deep learning-based systems has shown significant success for human
action recognition (HAR), even for multi-view camera systems. The authors [34] have
designed a deep learning-based HAR. This classifier has multiple steps, including feature
mapping, fusion, and selection. Two pre-trained models are considered for the initial
feature mapping step, and the extracted deep features are fused using the Serial-based
Extended approach. Then, the best features are selected using kurtosis-controlled weighted
k-nearest neighborhood. Benchmark datasets, such as KTH, IXMAS, WVU, and Hollywood,
are used to conduct the experiments, and achieved accuracies of 99.3%, 97.4%, 99.8%, and
99.9%, respectively [34].
CV researchers used deep learning techniques on medical images to diagnose COVID-
19 patients. An automated technique was proposed using parallel fusion and optimization
of deep learning models for classifying the COVID-19 data [35]. This technique starts with
a contrast enhancement using a combination of top-hat and Wiener filters. Optimal features
are selected using the entropy-controlled firefly optimization method. The selected features
are classified using machine learning classifiers such as multiclass support vector machine
(MC-SVM). Two pre-trained deep learning models AlexNet and VGG16 are used to classify
the data into COVID-19 and healthy as target classes. Experiments are carried out using
the Radiopaedia database and an accuracy of 98% was achieved [35].
Shervan et al. [36] have proposed a multi-layer perceptron (MLP) neural network
model to analyze cervical cancer datasets. The number of hidden layer neurons is tuned
and ResNet-34, and VGG-19 deep networks are used to feed the MLP. The authors [37]
have modified the layers related to the classification phase in these CNN networks, and
the outputs feed the MLP after passing through a flatten layer. CNNs are trained on
related images using the Adam optimizer to improve performance. Herlev benchmark
cervical dataset is applied to this model and achieved 99.23% and 97.65% accuracy for the
two classes. The results show that this method is more accurate than the other models [36].
Tan et al. [37] have proposed a federated learning (FL) approach to overcome the
conduct of experiments in central learning environments, which may breach patients’ pri-
vacy. To address these difficulties, an FL facility that extracts features from participating
environments rather than a CL facility has been developed. This study’s novel contri-
butions include (i) the application of transfer learning to extract data features from the
region of interest (ROI) in an image, which aims to enable careful pre-processing and
data enhancement for data training purposes; (ii) the use of a synthetic minority over-
sampling technique (SMOTE) to process data, which aims to classify data and improve
diagnostic prediction performance for diseases more uniformly; (iii) the application of
FeAvg-CNN + MobileNet in an FL framework to ensure customer privacy and personal
security; and (iv) the presentation of experimental results from different deep learning,
transfer learning, and FL models with balanced and imbalanced mammography datasets,
which demonstrate that our solution leads to much higher classification performance than
other approaches and is viable for use in AI healthcare applications.
N. S. Patil et al. [38] have designed a multi-classification framework using various
magnification factors, and this dataset is used. Their model examined several transfer
learning models while fine-tuning adaptation. Better results were obtained from the trials
using both the DenseNet201 and DenseNet121 pre-trained models. Compared to other
Diagnostics 2023, 13, 2746 8 of 28

models, the results are improved for the multi-classification of breast cancer histopathology
images [38].
R. Sanyal et al. [39] have proposed a hybrid method for classifying breast histopathol-
ogy images. Its frameworks include several fine-tuned CNN architectures as supervised
feature extractors and eXtreme gradient booster trees (XGBoost) as best classifiers. The
model consists of multiple discriminant representations of patches and XGBoost for robust
classification to optimize patches. The experimental data show that the proposed method
outperforms the most advanced form.
Bagchi, A. et al. [40] have used deep learning to classify breast cancer from histopatho-
logical images. The patch classification model is used where patches are preprocessed
using taint normalization, and amplification techniques. Image patches were divided into
four groups benign, normal, invasive, and in situ, using machine learning-based classi-
fiers and ensemble techniques. There are two main groups classified as cancerous and
non-cancerous, and the other four groups are classified as benign, in situ, and invasive
according to the model. This model used the probabilities for two classes and achieved
a 97.50% classification accuracy. The model achieves classification accuracies of 97.50%
and 98.6% for 4-class and 2-class image classifications, respectively, on the ICIAR BACH
dataset [40].
Guleria HV et al. [15] have proposed a combination of variational auto Encoder and
the Denoising Variational Auto Encoder algorithm for reconstructing the breast histopatho-
logical images. Then, they used CNN and tested this multi-classifier using Kaggle dataset
to predict cancerous or non-cancerous classes. This model has produced 73% accuracy,
which is much better than the traditional CNN [15].
To capture the discriminant features, a deep neural network (AHoNet) was designed
to acquire in-depth regional features of cancer images. On the other hand, AHoNet
uses a channel tracking strategy with low attenuation and local features located between
channels. Then, power matrix normalization and second-order comparison statistics have
been calculated to provide an excellent global representation of breast cancer images. The
BACH breast cancer database has been extensively analyzed by AHoNet. Experimental
results show that AHoNet outperforms a state-of-the-art model in this clinical application,
achieving the best results at the patient level. Classification accuracy of 99.29% and 85% in
the BreakHis and BACH databases, respectively, has been achieved [15].
Kumar, D. et al. [41] have designed a soft voting-based 7-CNN model. The proposed
method uses a VGG 19 transfer learning model (with and without data augmentation),
VGG 16 transfer learning model (without data augmentation), simple CNN with four
convolution layers, simple CNN with five convolution layers (with data augmentation),
and Xception. Another learning model with and without data augmentation is also used. A
hematoxylin–eosin dye has been used in the experiment. The performance of each model
is compared with accuracy, precision, recall, and F1 score. The main criteria is taken as the
basis for the evaluation. Using the H&E dataset, the proposed method achieves an accuracy
of 96.91%.
Shahidi et al. [42] have studied the ResNeXt, Dual Path Networks, SENet, and NASNet
deep learning models for analyzing the breast cancer histopathology image databases. Their
aim is to examine the state structure via applying the following steps: pre-processing, data
augmentation, and transfer learning methods. This study has identified the most accurate
models in terms of the binary, four, and eight classifications of breast cancer histopathology
image databases. The experimental results show that the models like Res-NeXt, Dual Path
Net, SENet, and NASNet have been identified with the most cutting-edge results for the
ImageNet database.
Kathale et al. [43] have proposed a technique for identifying malignant tissue and
classifying healthy and cancerous patients. The input is preprocessed using morphological
operations to separate the tumor area from the mammogram and highlight the size of
the raw mammogram. If the mammogram looks normal, the patient is healthy. Breast
Diagnostics 2023, 13, 2746 9 of 28

cancer patients and healthy individuals are classified using random forest classifiers, and
its accuracy for various patient images is 95% [43].
Farhadi et al., 2019 [44] have designed an efficient deep transfer learning method
to handle the imbalanced data problem in breast cancer data. The classifier focused on
structured data and used several publicly available breast cancer datasets to generate a ‘pre-
trained’ model, and transfer learned concepts to predict malignant tumors. The authors
compared their results with state-of-the-art techniques for addressing the problem of
imbalanced learning to handle different degrees of class imbalance; a series of experiments
are performed on publicly available breast cancer data under simulated class imbalanced
settings. The experimental results proved that the deep transfer learning model is used to
handle imbalanced class problems effectively.
Cutting-edge CNN architectures, such as ResNet and Inception, evaluate the effec-
tiveness of the proposed method. Experiments on various criteria, such as CIFAR-10 and
CIFAR-100, have been conducted. It has been evaluated on real-world data, such as the
Caltech-101 dataset containing 101 items. Finally, after going through the learning process
of the image using convolution and pooling techniques, all features are extracted and put
into the long tube. Hence, the extracted features are ready for classification with the help of
the usual fully connected layers [45].
Pectoral muscle detection is an important task to improve the diagnostic performance
of breast cancer detection. Early detection of pectoral muscles significantly reduces the
ambiguity between the tumor cell and pectoral muscle. It becomes necessary to suppress
the pectoral muscle to achieve appropriate segmentation. For artifacts and pectoral muscle
removal, Otsu’s threshold is used to identify pectoral muscles and connected compo-
nent labelling to recognize and eliminate the connected pixels outside the breast region,
classes 10 and 100, respectively [46].
To help pathologists analyze cancer, CAD is introduced to automatically identify and
diagnose breast cancer. The CAD system is based on deep convolution neural networks
(ConvNets) in this work. Deep learning (DL) establishes an important development in arti-
ficial intelligence (AI) and has been especially successful in image processing. Nevertheless,
the training of ConvNets needs a huge number of images. The transfer learning is utilized
to extract features from a pre-trained network of the ImageNet challenge (ILSVRC) for
further classification and used five famous ConvNets: ResNet, VGG19, VGG16, Xception,
and MobileNet have been used to resolve this issue over feature extraction. This method
supports limiting the hardware necessary to complete the large computation task of training
ConvNets. ConvNets with ML techniques are understood to perform better than the other
hand-crafted features utilized for histopathology images. From the experimental results,
it is found that VGG16 combined with the SVM approach outperformed in automated
histopathological image classification [47].
To enhance the diagnosis of unlabeled data, W. Sun et al. [33] have proposed a semi-
supervised training program based on weighted back data, selection of the feature concept,
and pre-diagnostic data collection using deep CNN model. However, many studies fo-
cused on benign and malignant differentiation, tumor localization, and detected a mass in
dense breast regions and pectoral muscles, which may be difficult due to high intensities.
Experimental results show that the schema only requires a few pieces of data (100 cases) for
training and performs well on unlabeled data (3058 points). This deep CNN-based CAD
model may help improve the radiologists’ analysis and detection of breast cancer [33].
There is a research gap in determining the ideal number of models to utilize in an
ensemble, even though using LeNet CNN with ensembling has improved the accuracy
of picture classification tasks [48]. Most researchers have used a fixed number of models,
ranging from a few to dozens, but are yet to look into the advantages or disadvantages of
changing the ensemble size. In addition, there is a lack of knowledge regarding the most
effective ways to mix various models into an ensemble, particularly when the models have
various architectures or hyperparameters. To increase accuracy rates in picture classification
tasks, further study is required to investigate the ideal ensemble size and the best ways to
Diagnostics 2023, 13, 2746 10 of 28

mix models in an ensemble. The results of this study can demonstrate the benefits of using
LeNet-CNN for clustering and provide useful suggestions on how to tune the clustering
process for different datasets and applications [48].
Min-pooling layer-based LeNet model is designed for Alzheimer’s Disease (AD)
classification, comparing it with 20 commonly used DNN models and related works [48].
The findings suggested that the LeNet model achieved superior performance and has
significantly improved classification accuracy. The LeNet model incorporates a separate
min-Pooling layer alongside the traditional MaxPooling layers, allowing preservation of
important low-intensity features.
Gray wolf optimizer (GWO) with a rough set theory-based CAD system is proposed
for mammogram image analysis [49]. Texture, intensity, and shape-based features are
extracted from mass-segmented mammogram images. A novel dimensionality reduction
algorithm is proposed based on GWO with rough set theory to derive the appropriate
features from the extracted feature set. GWO is a bio-inspired optimization algorithm
simulated based on hunting activities and the social hierarchy of the gray wolves. This
paper uses a hybridization of GWO and rough set (GWORS) methods to find the significant
features from the extracted mammogram images. To evaluate the effectiveness of the
proposed GWORS, particle swarm optimizer (PSO), genetic algorithm (GA), quick reduct
(QR) and relative reduct (RR) are compared. Experimental results have revealed that the
proposed GWORS outperforms the other techniques regarding overall accuracy, f-measures,
and receiver operating characteristic (ROC) curve [49].
Many CAD systems have been established to diagnose the disease and provide better
treatment. There is still a need to improve existing CAD systems via incorporating new
methods and technologies to provide more precise results. Most of the CAD models used
only mammogram datasets for experimentation. CAD models in breast cancer classification
have several limitations that should be considered. Some of these limitations include:
• CAD models may produce false-positive results, leading to unnecessary follow-up
tests or interventions. False positives can cause patient anxiety, additional healthcare
costs, and potential harm from invasive procedures. Striking a balance between
sensitivity and specificity is crucial to reduce false-positive rates in CAD models.
• CAD models may have lower sensitivity in detecting certain breast lesions, such as
small or subtle lesions, non-mass-like lesions, or early-stage cancers. These lesions
may not exhibit clear visual cues or distinctive features that CAD models can reliably
detect, leading to missed diagnoses [5,7].
• CAD models are often developed and trained on specific datasets, which may not
adequately represent the population or exhibit heterogeneity in terms of demograph-
ics, imaging protocols, or lesion types. This limited generalizability can affect the
performance of CAD models when applied to different populations or images.
• CAD models are often considered “black boxes” since medical professionals do not
readily interpret their decision-making process. This lack of transparency and inter-
pretability can create challenges in understanding the features or patterns driving the
model’s predictions, hindering trust and acceptance from clinicians.
• Integrating CAD models into clinical practice can be challenging. Incorporating CAD
systems requires workflow adjustments, radiologist training, and potential integration
with existing healthcare information systems. Adoption barriers, resistance to change,
and logistical constraints may hinder the effective integration of CAD models into
routine clinical workflows.
Addressing these limitations and continuously refining CAD models through the
proposed research maximizes their potential in breast cancer classification and improves
patient outcomes. The proposed research focuses on developing advanced image analysis
techniques and feature extraction methods to improve the discriminative power of CAD
models. Breast cancer datasets often suffer from class imbalance, where the number of
malignant samples is significantly lower than that of benign samples. Imbalanced data can
lead to biased model performance and reduced sensitivity in detecting cancerous cases.
Diagnostics 2023, 13, 2746 11 of 28

Developing techniques to address these research gaps leads to more accurate, inter-
pretable, and clinically useful CAD models to ensure reliable and flexible classification.
1. The proposed modified LeNet, like many other deep learning models, is designed for
performance and aims to classify images accurately.
2. Modified LeNet demonstrates promising performance on breast cancer ultrasound
datasets; their generalizability and applicability to real-world clinical settings, evalu-
ating their impact on clinical decision-making and patient outcome, are better than
any other CAD models.
LeNet [48] is a specific and relatively simple architecture within the broader category
of CNNs. LeNet typically has a shallower architecture compared to more recent CNN
models, which can limit its ability to capture intricate and complex features in breast cancer
images. LeNet is a simple CNN-type architecture; the authors have proposed the following
modifications, which can be considered to improve its classification performance for breast
cancer [48]. To increase the depth, more convolutional and fully connected layers have
been added, enhancing the model’s ability to capture complex patterns in breast cancer
images. This allows for the extraction of higher-level features that may contribute to better
classification performance.
Batch normalization [17] is applied after each convolutional or fully connected layer to
standardize the inputs and reduce internal covariate shifts. This technique helps stabilize
and accelerate the training process, improving classification accuracy. Dropout regular-
ization [22] can reduce overfitting via randomly dropping out a fraction of the neurons
during training. This encourages the network to learn more robust and generalized fea-
tures, enhancing the model’s ability to classify breast cancer images accurately and quickly.
LeNet can be combined with multiple networks called ensemble methods to average or
stack predictions from multiple LeNet models and can improve classification accuracy via
leveraging diverse representations learned via different network instances. Thus, modified
LeNet architectures for breast cancer classification, ultimately improves early detection,
diagnosis, and treatment decisions for patients efficiently.

3. Proposed System
The convolutional layer in a CNN performs an operation where two functions, the
input image and the convolution filter, are combined to generate a third variable. The
essential components of a CNN include a convolutional layer, a pooling layer, a fully
connected layer, and another convolutional layer.

3.1. CNN Model

The image pixel values are passed through the first convolutional layer in CNN
architecture which is presented in Figure 1. Each subsequent convolutional layer aims to
discover relationships between the pixels and their neighboring pixels utilizing kernels
to extract various features from the image [8]. CNN typically consists of a convolutional
layer, which applies a set of learnable filters to the input image. These filters detect various
features such as edges, textures, and patterns. The output of the convolutional layer
is a feature map that highlights the presence of these features in the image. Following
the convolutional layer, a common choice is to use a pooling layer, such as max pooling
or average pooling. This layer reduces the spatial dimensions of the feature map while
preserving the most important features. Pooling helps to downsample the data, reducing
the computational complexity and extracting the most relevant information.
After the pooling layer, additional convolutional and pooling layers can be stacked
to capture increasingly complex features in the image. This hierarchical structure enables
the network to learn abstract representations of the input data. To introduce non-linearity
into the model and enhance its expressive power, activation functions like ReLU (Rectified
Linear Unit) are applied after each convolutional layer. ReLU sets negative values to
zero, preserving positive values, and introducing non-linear transformations. Towards
the end of the network, fully connected layers are typically used. These layers process the
3, x FOR PEER REVIEW 12 of 29

Diagnostics 3.1.
2023, CNN
13, 2746Model 12 of 28

The image pixel values are passed through the first convolutional layer in CNN ar-
chitecture which is presented in Figure 1. Each subsequent convolutional layer aims to
extracted
discover relationships features
between theand make
pixels predictions
and based on them.
their neighboring Fully
pixels connected
utilizing layersto
kernels connect
all the neurons from the previous layer to the neurons in the current layer, enabling the
extract various features from the image [8]. CNN typically consists of a convolutional
network to learn complex relationships between the features and the output labels. Finally,
layer, which applies a set
the output oflayer
learnable
of the filters to the input
CNN depends on theimage.
specificThese
task atfilters
hand. detect var- in a
For example,
ious features suchclassification
as edges, textures, and patterns.
task, a SoftMax layer can The output
be used of thethe
to produce convolutional layer
probabilities for each class,
while in a segmentation task, a convolutional layer with a pixel-wise
is a feature map that highlights the presence of these features in the image. Following the activation function
may
convolutional layer, a be employed.
common It is important
choice is to use that the architecture
a pooling and sequence
layer, such as max of layers in
pooling ora CNN
can vary depending on the specific requirements of the task and the dataset being used.
average pooling. This layer reduces the spatial dimensions of the feature map while
Further, the researchers often experiment with different architectures, hyperparameters,
preserving the mostandimportant features.
regularization Pooling
techniques helps to
to optimize downsample
performance the data,
and achieve reducing
accurate and reliable
the computationalresults.
complexity and isextracting
The result passed onthe most
to the relevantlayers
subsequent information.
and the working of the kernel
presented in Figure 2.

3, x FOR PEER REVIEW 13 of 29

Figure 1. Convolutional neural network.

Figure 1. Convolutional neural network.

After the pooling layer, additional convolutional and pooling layers can be stacked
to capture increasingly complex features in the image. This hierarchical structure enables
the network to learn abstract representations of the input data. To introduce non-linearity
into the model and enhance its expressive power, activation functions like ReLU (Recti-
fied Linear Unit) are applied after each convolutional layer. ReLU sets negative values to
zero, preserving positive values, and introducing non-linear transformations. Towards
the end of the network, fully connected layers are typically used. These layers process the
extracted features and make predictions based on them. Fully connected layers connect
all the neurons from the previous layer to the neurons in the current layer, enabling the
Figure 2. CONV layer.
Figure 2. CONV layer.
network to learn complex relationships between the features and the output labels. Fi-
nally, the output layer Figure
of the2CNN shows depends
theThe
CONV onlayer.
the specific
The number taskofat hand. For example,
parameters in a
Figure 2 shows the CONV layer. number of parameters in a CNNinisadetermined
CNN is determined
classification task,by
a SoftMax layer can
the combination be used
of these layertosizes
produce the probabilities
and depths. Each parameterfor each class, to a
corresponds
by the combination of these layer sizes and depths. Each parameter corresponds to a
weight value
while in a segmentation task, that is learned during
a convolutional layerthewith
training process. The
a pixel-wise number of
activation parameters in
function
weight value that is learned duringlayers the training process. The number of parameters inofthe
may be employed.the It convolutional
is important that the is influenced
architectureby theand sizesequence
of the filtersofand the depth
layers in a CNN the layers.
convolutional layersMore is filters
influenced byfilters
or larger the size
lead toof an
theincreased
filters andnumberthe ofdepth of the The
parameters. layers.
number of
can vary depending on the specific requirements of the task and the dataset being used.
More filters or larger filters lead
parameters in theto an connected
fully increasedlayers number of parameters.
is determined Theofnumber
by the size the layers, ofas each
Further, the researchers often experiment with different architectures, hyperparameters,
parameters in the fully connected layers is determined by the size of the layers, as each more
connection between neurons requires a weight parameter. While it is true that adding
and regularizationconvolutional
techniques to optimize
layers performance
can increase and of achieve accurate
is notand reliable
connection between neurons requires a weightthe number
parameter. parameters,
While it is ittrue the only
that adding factor that
results. The resultaffects
is passed on tocount.
parameter the subsequent
The number of layers and the
parameters in a working
neural networkof thearchitecture
kernel can
more convolutional belayers
influencedcan increase
by variousthe number
factors, of parameters,
including the sizes and it is not the
depths onlyconnected
of fully factor and
presented in Figure 2.
that affects parameter count. The number of parameters in a neural network
convolutional layers. With multiple convolutional layers, CNNs can learn increasingly architecture
can be influenced abstract
by various factors,representations
and complex including the sizes
of the andLower
input. depths of capture
layers fully connected
low-level features,
and convolutional layers. With multiple convolutional layers, CNNs can learn increas-
ingly abstract and complex representations of the input. Lower layers capture low-level
features, while deeper layers capture high-level semantic information. Increasing the
depth or size of these layers can generally lead to more parameters, but other architec-
 Diagnostics 2023, 13, 2746 13 of 28

while deeper layers capture high-level semantic information. Increasing the depth or size
of these layers can generally lead to more parameters, but other architectural choices like
pooling, strides, and padding also impact the total number of parameters. Convolutional
layers apply filters to the input data, enabling them to extract meaningful features, such
as edges, textures, shapes, and to generalize well. These learned features are crucial for
subsequent layers to easily understand and classify the data. Convolutional layers can
detect features regardless of their exact spatial location. This property makes CNNs robust
to translations or shifts in the input data.

3.2. Fully Connected Layer

A type of neural network layer called the fully connected layer has connections
Diagnostics 2023, 13, x FOR PEER REVIEW 14 of 29
between each neuron in the upper layer which is shown in Figure 3. This layer, also known
as a dense layer, is frequently employed in deep learning models to perform tasks including
speech recognition, natural language processing, and image categorization [34].

Figure
Figure 3. Fully
3. Fully connectedlayer.
connected layer.

A completely connected layer has a vector as input and another as output. Each neuron
inAthe
completely connected
layer performs layer
a weighted sumhas
of a
thevector
inputsasandinput and
applies an another
activationas output. Each
function
neuron in the layer performs a weighted sum of the inputs and applies
to produce the output. Backpropagation and gradient descent methods are used in the an activation
function to process
training produce tothe output.the
determine Backpropagation
weights and biasesand gradient
of the descentlayer.
fully connected methods
Everyare used
training iteration involves adjusting these weights and biases to reduce the loss
in the training process to determine the weights and biases of the fully connected layer of function.
Fully linked layers are adaptable and robust, but they also include a lot of parameters,
Every training iteration involves adjusting these weights and biases to reduce the loss of
which, if not regularized appropriately, can result in overfitting. Deep learning models
function. Fully linked layers are adaptable and robust, but they also include a lot of pa-
frequently employ weight decay and dropout strategies to overcome this problem [22].
rameters, which, if not regularized appropriately, can result in overfitting. Deep learning
4. LeNet
models Architecture
frequently employ weight decay and dropout strategies to overcome this prob-
lem [22].LeNet [48] is used for image classification tasks. Compared to other deep learning
classifiers, LeNet has several advantages: parameter efficiency, spatial invariance, hierar-
chical Architecture
4. LeNet feature learning, weight sharing, activation function, and pooling layer. A modified
LeNet CNN for image classification and ensembling combines multiple models to improve
LeNet [48]
accuracy. is used for
The modified image
LeNet classification
architecture includestasks. Compared
additional to other
convolutional deep learning
and pooling
classifiers,
layers toLeNet
increasehas several
model depthadvantages: parameter
and capture more complex efficiency, spatial to
features. In addition invariance,
this, the hier-
classifier
archical provides
feature batch normalization
learning, weight sharing,[17] activation
and continuous outputand
function, to prevent overfitting
pooling layer. A mod-
ifiedand improve
LeNet CNN generalization. The limitationsand
for image classification of current models that
ensembling use LeNet
combines for cancer
multiple models to
diagnosis and classification are: although LeNet can be used effectively for cancer diag-
improve accuracy. The modified LeNet architecture includes additional convolutional
nosis and classification, it also has some drawbacks, which include limited capacity and
and pooling layers to increase model depth and capture more complex features. In addi-
tion to this, the classifier provides batch normalization [17] and continuous output to
prevent overfitting and improve generalization. The limitations of current models that
use LeNet for cancer diagnosis and classification are: although LeNet can be used effec-
Diagnostics 2023, 13, 2746 14 of 28

flexibility, overfitting, pre-processing, and hardware requirements which are significant

challenges [32]. To overcome these drawbacks, the present model aims to develop LeNet
architecture with the following modifications: In the data preparation process, the Breast
Ultrasound (BUS) image datasets are obtained and divided into three groups: training,
validation, and testing [50]. The 60:20:20 data split refers to dividing the available dataset
into three subsets, a training set (60% of the data), a validation set (20% of the data), and a
test set (20% of the data). This split allows for model training on the training set, hyperpa-
rameter tuning on the validation set, and final evaluation on the test set. Hyperparameter
tuning is essential to maximize the model’s performance and to find the best configuration
for a specific task. It requires a careful exploration of the hyperparameter space and an
understanding of the trade-offs among different choices. Through hyperparameter tuning,
models can be fine-tuned to achieve optimal performance on the given dataset. Hyper-
parameter tuning is an important step in optimizing the performance of deep learning
models. The following are the key hyperparameters that can be considered for tuning:
1. Learning rate is 0.01 which determines the step size at each iteration during training
and achieves the best performance.
2. Dropout rate is 40% (0.4) which controls the amount of regularization applied to the
model and prevents overfitting.
3. Batch size is 32 which determines the number of samples processed before the model’s
weights are updated. This will lead to faster convergence and will result in better
generalization which impacts the model’s performance.
4. Number of Hidden Units is three which can significantly impact the model’s capacity
and performance and finding the optimal balance between model complexity and
overfitting.
5. Number of training epochs is 10 which determines how many times the model will
iterate over the entire training dataset which leads to balance between underfitting
and overfitting.
6. Activation functions: The activation function used in the convolutional layers (ReLU)
and the dense layer (SoftMax) can also be experimented with to determine if other
activation functions yield better results for your specific problem.
7. Number of filters in the convolutional layers are 32, 64, and 128 which can be tuned
to determine the complexity and capacity of the model.
8. Kernel size in the convolutional layers is 3, 5, and 4 which can be tuned to capture
different spatial patterns in the input data.
9. Optimizer: The choice of optimizer can also influence the model’s performance. An
optimizer is a function that modifies the attributes of the neural network, such as
weights and learning rate. Common optimizers include Adam, RMSprop, and SGD
with momentum. Each optimizer has different hyperparameters of its own, such as
momentum or decay rates. Thus, it helps in reducing the overall loss and improving
accuracy. Here, the Adam optimizer is used which is best suited for breast cancer
datasets.
This approach helps in assessing the model’s performance on unseen data and avoid-
ing overfitting. Pre-processing images are fed as input into an ML model. In the case
of the LeNet model, it is necessary to pre-process the images into a format that can be
effectively processed by the model. This typically entails resizing the images to a specific
size, normalizing the pixel values, and potentially applying other transformations as per the
model architecture’s requirements. The unique aspect of LeNet that sets it apart from many
other classifier networks is its utilization of weight sharing and the choice of activation
function [51]. LeNet, a classic CNN architecture, has several unique aspects that make it
well-suited for image classification tasks:
Diagnostics 2023, 13, 2746 15 of 28

1. Weight sharing: LeNet utilizes weight sharing in its convolutional layers. The same
set of weights is applied to different spatial locations across the input. This sharing of
parameters enables the network to extract and recognize similar features throughout
the image. It helps reduce the total number of parameters and allows LeNet to
efficiently capture local patterns.
2. Activation function: LeNet employs the sigmoid activation function in its fully con-
nected layers. The sigmoid function squashes the output of each neuron into a range
between 0 and 1. This non-linearity introduces non-linear transformations and allows
the network to model complex relationships between features.
3. Convolution and pooling operations: LeNet incorporates convolutional layers for
spatial feature extraction. Convolutional layers apply filters to the input, capturing
local patterns and creating feature maps. Pooling operations, specifically max pooling
in LeNet, downsample the feature maps via selecting the maximum value in each
pooling region which reduces the spatial dimensions of the features.
4. Architectural simplicity: LeNet architecture consists of alternating convolutional and
pooling layers followed by fully connected layers. Overall, LeNet’s unique features
of weight sharing, sigmoid activation, and the combination of convolutional and
pooling operations made it a pioneering architecture for image classification tasks and
achieving higher performance.

5. Modified LeNet Architecture

The modified LeNet architecture is trained on a large-scale image classification dataset
such as ImageNet, which contains over 1 million images with 1000 classes. The training
process includes data augmentation techniques such as random clipping, horizontal flips,
and dithering to improve model robustness and prevent overfitting. After training the
individual models, ensembling techniques combine their predictions. Specifically, a simple
averaging method is used where the outputs of the unique models are averaged to obtain
the final prediction. The classifier uses more advanced ensembling techniques, such as
stacking and boosting, which can further improve accuracy.
The traditional LeNet-5 architecture is shown in Figure 4a which consists of the
convolutional layer with six filters of size 5 × 5, a stride of 1, no padding, and Sigmoid
as an activation function. A 5 × 5 convolutional layer with stride 1 and no padding has a
receptive field of 5 × 5, which can be too large for specific image classification tasks. A large
receptive field can cause the model to lose some spatial information and fine-grained details
in the input image. The 5 × 5 convolutional layers typically have a higher computational
cost than 3 × 3 convolutional layers, especially when the number of input channels is large.
This can make the training and inference of the model slower and more computationally
expensive.
In this proposed modified LeNet model (presented in Figure 4b), the 5 × 5 layers are
replaced using two 3 × 3 layers for better feature extraction. The intuition behind this
modification is that stacking two 3 × 3 convolutional layers one after the other allows the
model to learn more complex and abstract features, while keeping the same receptive field
as a single 5 × 5 convolutional layer. This is because a 3 × 3 convolutional layer with
stride 2 and no padding has a receptive field of 3 × 3 and stacking two of these layers
in sequence results in a receptive field of 5 × 5. Further, using two 3 × 3 convolutional
layers can result in more significant nonlinearity, improving the model’s ability to learn
complex and non-linear relationships between the input image and the target labels. The
batch normalization is applied after each convolutional layer. It normalizes the activations,
stabilizing and accelerating the training process.
The original pooling layers are replaced with convolutional layers having a stride
of 2. This achieves downsampling while enabling the network to learn more intricate
representations. The sigmoid activation function of the traditional LeNet is replaced with
the Rectified Linear Unit (ReLU) [51] activation function. ReLU mitigates the vanishing
gradient problem and improves training convergence. Dropout regularization is introduced
 The modified LeNet architecture is trained on a large-scale image classification da-
taset such as ImageNet, which contains over 1 million images with 1000 classes. The
training process includes data augmentation techniques such as random clipping, hori-
zontal flips, and dithering to improve model robustness and prevent overfitting. After
Diagnostics 2023, 13, 2746 training the individual models, ensembling techniques combine their predictions.16Spe- of 28

cifically, a simple averaging method is used where the outputs of the unique models are
averaged to obtain the final prediction. The classifier uses more advanced ensembling
techniques,
after specificsuch as stacking
layers which are and boosting,
shown which4b.
in Figure canItfurther
randomlyimprove accuracy.
deactivates a fraction of
inputThe
unitstraditional LeNet-5
during training, architecture
helping preventisoverfitting
shown in and Figure 4a which
enhancing consists of the
generalization. The
convolutional
modified layer with
architecture six filters
is made of size
deeper via5adding
× 5, a stride of 1, noconvolutional
additional padding, andand/or
Sigmoid as
fully
an activation
connected function.
layers A 5 initial
after the × 5 convolutional
layers. Thislayer with the
increases stride 1 and nocapacity
network’s paddingtohas
learna
receptivepatterns.
complex field of 5The
× 5,modified
which can be too large
architecture for specific
involves image multiple
ensembling classification
CNNtasks.
models.A
Ensembling
large receptive combines
field multiple
can cause models to improve
the model the some
to lose overallspatial
performance and robustness
information and fi-
of predictions.
ne-grained Instead
details in theof relying on a single
input image. The 5model, ensembling leverages
× 5 convolutional the wisdom
layers typically have of a
crowds via aggregating
higher computational the
cost predictions
than from multiple
3 × 3 convolutional models
layers, to make
especially whena final prediction.
the number of
Modified LeNetismodels
input channels would
large. This canbe creating
make an ensemble,
the training trained on
and inference of different
the modelsubsets
slowerofand
the
training data or with different
more computationally initialization. The final prediction of the ensemble would be
expensive.
determined through soft voting or averaging of the predictions from the individual models.

(a)

(b)
Figure 4.
Figure 4. (a)
(a) Classical
Classical LeNet
LeNet architecture.
architecture. (b)
(b) Modified
ModifiedLeNet
LeNetarchitecture.
architecture.

5.1. Batch Normalization

In this proposed modified LeNet model (presented in Figure 4b), the 5 × 5 layers are
replaced using two 3 × 3 [17]
Batch normalization layers
is afor better feature
technique that theextraction.
LeNet modelThecan
intuition
use to behind
improvethis
the
modification is that stacking two 3 × 3 convolutional layers one after the
performance of neural networks, including various image classification tasks thatother allows the
include
model cancer
breast to learn more complex
detection and abstract
using ultrasound features,
images. Here while keeping
are some the same
advantages receptive
of using batch
normalization in LeNet for breast cancer detection. Batch normalization can help accelerate
the confluence of the training process via reducing the internal covariate shift, which refers
to the change in the distribution of the input to a layer due to the changing weights in the
previous layers and improve convergence. Batch normalization [17] can reduce internal
variability via normalizing data at each layer and help stabilize gradients during recovery
quickly. The advantages are regularization, improved accuracy, and greater flexibility.
LeNet with batch normalization for breast cancer detection using ultrasound images can
improve convergence, regularization, accuracy, and flexibility, all of which can contribute
to better performance on this task.
Diagnostics 2023, 13, 2746 17 of 28

5.2. Replacing the Pooling Layers

The rationale behind replacing the pooling layers with convolutional layers of stride 2
in the LeNet architecture is to improve information retention, enhance feature extraction,
and potentially enhance the model’s performance [51]. The reasons behind this modification
lie in the potential limitations of pooling operations. Pooling operations can discard some
information from the feature maps, especially if the pooling kernel size is large or the
stride is high. By replacing the pooling layers with convolutional layers of stride 2, more
information can be retained in the feature maps. This can be beneficial for preserving
important details and preventing the loss of valuable information during the downsampling
process.
Convolutional layers can learn more complex and abstract features than pooling layers,
especially when they have more parameters or are deeper. The model can extract more
relevant features from the input images utilizing convolutional layers instead of pooling
layers, ultimately improving the accuracy of the classification task. Another advantage
of this modification is the potential reduction in computational cost. Pooling operations
can be computationally expensive, especially with large kernel sizes or high strides. The
computational burden can be reduced via replacing the pooling layers with convolutional
layers of stride 2, making the model faster to train and evaluate. Overall, replacing the
pooling layers with convolutional layers of stride 2 enhances information retention, feature
extraction, model performance, and computational efficiency within the LeNet architecture.

5.3. ReLU Activation

The activation of each layer in a neural network controls the activation of the preceding
layer and contributes to the nonlinearity of the overall network. The proposed architecture
introduces a mesh consisting of five convolutional stack normalization groups. Replacing
the sigmoid activation function with the ReLU activation function effectively alters the
LeNet architecture for diagnosing cancer using ultrasound images. This modification
brings several benefits, including improved performance, faster computation, increased
learning capacity, enhanced abstraction, and mitigating the vanishing gradient problem.
The vanishing gradient problem arises when the gradient of the activation function becomes
too small, impeding the adjustment of weights during training. ReLUs are less prone to this
issue than sigmoid processing, enabling the network to effectively learn the characteristics
and patterns within ultrasound images [51]. This modification aims to enhance accuracy,
improve comprehension, and increase specificity.

5.4. Dropout
Adding dropouts to the LeNet architecture for breast cancer detection can be useful.
Dropout is a regularization technique that randomly drops out (sets to zero) some of the
activations in the network during training [22,51]. The percentage of dropouts used is
40% (0.4) for all the dropout layers (Dropout(0.4)). Four dropout layers exist: two after
the convolutional layers and two before the final dense layer. Dropout can help prevent
overfitting of the model to the training data via reduction of the co-adaptation of neurons.
This improves the model’s ability to expand to new, unseen data, which is important
for accurate breast cancer detection and performing robust learning [51]. By randomly
dropping out activations during training, the network is forced to rely on a wider range
of features and patterns in the input data. This can improve the model’s robustness to
input data variations, such as differences in patient demographics or imaging conditions.
Dropout can help reduce the model’s sensitivity to the network’s initial conditions (i.e.,
the weights and biases). This increases the reproducibility of the results and reduces the
risk of the model getting stuck at the local minimum during training. By lowering the
co-adaptation of neurons and forcing the network to rely on a broader range of features,
dropout can make the training process more efficient and less prone to becoming stuck in
local minima [33]. Figure 5 shows the various layers in the modified LeNet model.
stics 2023, Diagnostics 13, 2746
2023,REVIEW
13, x FOR PEER 19 of 29 18 of 28

Figure 5. LayersFigure
in the5. Layers in
modified the modified
LeNet model. LeNet model.
Diagnostics 2023, 13, 2746 19 of 28

6. Materials and Methods

6.1. Dataset Description
The “Breast Ultrasound Images Dataset” is available at Kaggle [50]. It contains
971 breast ultrasound images in JPEG format and their corresponding labels (0, 1, and 2)
indicating a resolution of 500 × 500 pixels. The images are categorized into three classes,
which are normal, benign, and malignant.

6.2. Software Details

Google Colab, a cloud-based Jupyter notebook environment was used for conducting
our experiments. Google Colab was connected to Google Drive to access the image sources.
The presence of GPUs in Google Colab, specifically configured with type T4, greatly
expedited the model training process. Deep learning models are built using the Keras 2.6.0
library, which is a user-friendly neural networks API that operates on top of TensorFlow
2.7.0, a robust open-source machine learning framework. The seamless integration of
Keras and TensorFlow with Google Colab allowed the effortlessly import of the necessary
libraries and executed code within a collaborative environment. This integration not only
facilitated the development of the models but also maximized the utilization of GPU
resources. Python 3.10.11 is used for the implementation.

7. Performance Metrics
To demonstrate the performance of the proposed model, breast ultrasound images are
used. Our model was trained for 40 epochs. The accuracy of the breast image recognition
rate is 89.91%. The various performance evaluation metrics are discussed below.

7.1. Classification Accuracy

Classification accuracy is a metric used to evaluate the performance of a classifier
model during the training phase. It measures the proportion of correctly classified instances,
both positive and negative, out of the total instances in the training set. It is usually defined
as the percentage of samples excluded from the training process. Based on this, the training
accuracy is represented in Equation (1):

Accuracy = TP + TN/(TP + TN + FP + FN) (1)

where TP represents true positives; TN represents true negatives; FP represents false

positives; and FN represents false negatives. TP is defined as the cases where the model
correctly predicts the positive class. TN is defined as the cases where the model correctly
predicts the negative class. FP is defined as the cases where the model incorrectly predicts
the positive class. FN is defined as the cases where the model incorrectly predicts the
negative class. The training accuracy metric provides a general overview of how well
the model can classify instances correctly during training. Training accuracy metrics
provide insights into different aspects of the model’s behavior, such as its ability to handle
imbalanced classes or its performance in specific classes.
Furthermore, it is important to split the data into separate training and testing sets
to accurately evaluate the model’s performance on unseen data. This helps in identifying
and mitigating issues such as overfitting, where the model becomes overly specialized
to the training data and fails to generalize well. In conclusion, while training accuracy
is a valuable metric to assess the model’s performance during training, it should not be
considered in isolation. It is crucial to analyze various evaluation metrics, use separate
validation and testing sets, and consider the model’s performance on unseen data to gain a
more comprehensive understanding of its effectiveness.
Diagnostics 2023, 13, 2746 20 of 28

7.2. Validation Loss

Validation loss is a valuable metric for assessing model performance on unseen data.
It is crucial to guide model selection, implement early stopping, and optimize hyperparam-
eters to improve this model’s generalization and predictive capabilities. The validation
loss is calculated via feeding the dataset (a subset of the total dataset not used for training)
through the trained model and computing the loss on this dataset and is presented in
Equation (2):

L = −(1/N) × ∑(y × log(y_pred) + (1 − y) × log(1 − y_pred)) (2)

where N is the number of samples in the validation dataset; y is the true label (either 0 or
1); and y_pred is the predicted probability of the positive class (i.e., the class with label
1). Results are obtained from all N samples in valid data. The total validation loss is the
average of all sample losses is presented in Equation (3):

Val_loss = (1/N) × ∑(L) (3)

7.3. Precision
Precision measures the proportion of correctly predicted positive samples out of all
the samples predicted as positive and is presented in Equation (4):

Precision = TP/(TP + FP) (4)

7.4. Recall
Recall (also known as sensitivity or true positive rate) measures the proportion of
correctly predicted positive samples out of all the actual positive samples and is presented
in Equation (5):
Recall = TP/(TP + FN) (5)

7.5. F1 Score
The F1 score is the harmonic mean of precision and recall, providing a balanced
measure of model performance which is presented in Equation (6):

F1 Score = 2 × (Precision × Recall)/(Precision + Recall) (6)

8. Results and Discussion

The proposed ensembled LeNet CNN outperforms the original LeNet architecture
and performs comparably to other state-of-the-art CNN architectures. According to our
experimental findings, the ensembled LeNet CNN models considerably increased the
dataset’s breast image recognition accuracy, especially when the individual models have
different strengths and weaknesses. Experimental evaluation of an ensembled LeNet CNN
for breast cancer analytics would involve training multiple LeNet models on a breast cancer
dataset and then combining their predictions using some ensemble method. It is worth
noting that the choice of ensemble method and hyperparameters can significantly impact
the performance of the ensembled LeNet CNN. Therefore, different ways and settings
should be tested to find the best combination for the given data and the problem to be
solved. Training/validation accuracy per epoch for all the 13 models is shown in Figure 6.
Experimental evaluation of ensembled LeNet CNN for breast cancer analytics is
performed as follows:
Diagnostics 2023, 13, 2746 21 of 28
Diagnostics 2023, 13, x FOR PEER REVIEW 22 of 29

Figure
Figure 6.
6. Training
Training and
and validation
validation Accuracy
Accuracy per
per epoch
epoch for
for ensemble
ensemble of
of 13
13 LeNet
LeNet Models
Models in
in BUS
BUS
Dataset.
Dataset.

Experimental evaluation of ensembled LeNet CNN for breast cancer analytics is

performed as follows:
Diagnostics 2023, 13, 2746 22 of 28

1. The first step is to preprocess the BUS dataset to ensure it is in the appropriate format
for training and testing models. This will include resizing the image, normalizing the
pixel values, and dividing the data into training, testing sets, and validation sets.
2. Train multiple LeNet models with different initialization or hyperparameter settings
on the training set. This can be achieved using the TensorFlow deep learning frame-
work.
3. Combine the predictions of the individual LeNet models using soft voting.
4. Hyperparameter tuning optimizes the model’s performance via finding optimal
values for different hyperparameters. The number of filters in convolutional layers
affects model complexity. Kernel size captures spatial patterns. Adjusting dropout
rate prevents overfitting. Learning rate in Adam optimizer controls gradient descent
step size. Random search samples hyperparameters randomly within a predefined
range and provides computational efficiency for exploring the hyperparameter space.
5. Evaluate the performance of the LeNet CNN cluster via calculating parameters such
as accuracy, precision, recall, and F1 score. Compare the overall performance of a
LeNet model and other high-end models.

8.1. Modified LeNet’s Performance Scores

From Table 1, it is understood that the model achieved a good balance between
precision and recall, considering both false positives and false negatives. A higher precision
score indicates that the model has a low false positive rate, meaning it correctly identifies
positive instances without wrongly classifying negative instances. A higher recall score
suggests that the model has a low false negative rate, meaning it successfully identifies most
positive instances without missing many. The F1 score suggests that the model’s predictions
have a good balance between correctly identifying positive instances and minimizing false
positives and false negatives. Ensembling, ReLU activation function, and 40% dropout that
are incorporated in LeNet have contributed to these performance scores.
Table 1. Precision, recall and F1 score of modified LeNet.

Method Precision Recall F1 Score

Modified LeNet 0.85 0.92 0.88

The ensemble method combines multiple models to improve overall prediction accu-
racy, while the LeNet architecture is specifically designed for BUS image recognition tasks.
Dropout regularization can help prevent overfitting via randomly dropping out a fraction
of the units during training. The ReLU activation function is commonly used in deep
learning models as it helps introduce non-linearity and can improve the model’s ability to
learn complex patterns. Overall, the modified LeNet model appears to perform well, with
high accuracy, precision, recall, and F1 score values, indicating an effective classification
performance.

8.2. Experimental Evaluation

With ensembled LeNet CNN, breast cancer classification is successfully completed.
540 images have been used for training, 180 images for testing, and 180 images for validation
and these images (40 nos. only) are represented in Table 2 (indexed with Image ID) which
are fed into each LeNet CNN model which predicted the output (0 → normal, 1 → benign,
and 2 → malignant) and finally the majority of the outputs of the thirteen ensemble models
(shown in Table 1) is declared as a result. For example, in Image ID 1, the first LeNet CNN
model predicted 0 which means normal, while the rest of the ensemble models predicted
1, which means benign. Since most models indicated it as 1, the result is also 1. The same
procedure is followed for all 40 imageids. The average runtime taken for each iteration is
the 1750s which is much better than the average runtime of classical LeNet (2120s).
Diagnostics 2023, 13, 2746 23 of 28

Table 2. Predictions (for 40 images) made using the proposed model using soft voting approach.

Image ID Won Image ID Won Image ID Won Image ID Won

1 11 1 21 1 31 1
2 1 12 1 22 1 32 1
3 1 13 1 23 1 33 1
4 1 14 1 24 1 34 1
5 1 15 1 25 1 35 1
6 1 16 1 26 1 36 2
7 1 17 1 27 1 37 1
8 1 18 1 28 2 38 1
9 0 19 0 29 0 39 0
10 1 20 1 30 1 40 1

8.2.1. Modified LeNet Model’s Predictions Using BUS Dataset

During the training process, early stopping was triggered at the 13th iteration, after
27 epochs. At this point, the proposed model demonstrated higher accuracy than the
traditional neural network and LeNet models. This indicates that the proposed model
outperformed the other models significantly, showcasing superior performance and its
effectiveness. The use of modified LeNet in model construction allows for early stopping.
The classifier has made predictions on medical images, where the labels indicate whether
the image shows normal, benign, or malignant tissue. Table 2 shows the soft voting-based
ensemble learning method that combines the predictions of multiple individual classifiers
to make a final prediction. Each individual classifier gives a probability distribution over
the possible classes, and the soft voting classifier takes the average of these probabilities to
make the final prediction.
Early stopping is commonly used in machine learning to prevent overfitting and find
the optimal training point. It involves monitoring a chosen metric, such as validation
loss or accuracy, during training. Training is stopped early to prevent overfitting and
retain the best-performing model when the metric stops improving or consistently starts to
deteriorate. In Table 3 of the ensembled runs, when early stopping (due to hyperparameter
tuning) is triggered during the training process, the model weights from the iterations
that achieved the best performance are typically saved. This saved model corresponds to
the best-performing iteration (e.g., the 13th iteration) and is then utilized for inference or
further evaluation on the test set. Figure 7 shows the predicted images for normal, benign,
and malignant. From the empirical results, it is observed that the changes performed in the
modified LeNet model reduced the computational time and thus enhanced the classification
performance (compared to classical LeNet).
Table 3. Ensembled LeNet iterations to find the best performing model.

Models Accuracy Val_Accuracy Loss Val_Loss

Net1 0.8493 0.7394 1.4750 2.1780
Net2 0.8427 0.85906 1.5121 2.1867
Net3 0.8402 0.7258 1.4423 1.9361
Net4 0.8557 0.7943 1.3765 2.1965
Net5 0.8427 0.7459 1.4236 1.9475
Net6 0.8522 0.7856 1.2977 1.5963
Net7 0.8470 0.7681 1.3175 2.2156
Net8 0.8469 0.7698 1.4092 2.3700
Net9 0.8459 0.7322 1.3739 2.1124
Net10 0.8469 0.7469 0.4733 0.4431
Net11 0.8396 0.7983 1.5013 1.7907
Net12 0.8991 0.7652 1.3770 1.7707
Net13 0.8516 0.7591 1.4053 1.8033
 Net9 0.8459 0.7322 1.3739 2.1124
Net10 0.8469 0.7469 0.4733 0.4431
Net11 0.8396 0.7983 1.5013 1.7907
Diagnostics 2023, 13, 2746 Net12 0.8991 0.7652 1.3770 1.7707
24 of 28

Net13 0.8516 0.7591 1.4053 1.8033

Figure 7. 7.Predictions
Figure Predictionsfor
forBUS
BUS dataset obtainedusing
dataset obtained usingmodified
modifiedLeNet
LeNet showing
showing normal,
normal, malignant,
malignant,
and benign classifications.
and benign classifications.

8.2.2.
8.2.2. ModifiedLeNet
Modified LeNetModel
Model’s Performance Analysis
s Performance Analysis
The classifier’s predictions for each Image ID is presented in Table 2. Values 1, 2 and
The classifier s predictions for each Image ID is presented in Table 2. Values 1, 2 and
3 in the “won” in Table 2 represent normal tissue, malignant tissue, and benign tissue,
3 in the “won” in Table 2 represent normal tissue, malignant tissue, and benign tissue,
respectively. The classifier has apparently predicted that most of the images are either
respectively. The classifier has apparently predicted that most of the images are either
normal or benign, with only a few images (IDs 6, 28, and 36) being predicted as malignant.
normal or benign,
Experimental results withareonly showna few images7.(IDs
in Figure The6,results
28, and are36)
very being
muchpredicted
identical as malig-
to the
nant. Experimental
clinician’s results and results
from are shown
Figure in Figure
7, malignant, 7. Theand
benign, results
normal areclasses
very much
obtained identical
based to
theonclinician s results and from Figure 7, malignant, benign, and
the BUS dataset are clearly depicted. This is possible because of the model’s ability normal classes obtained
to
based
obtainonmuch the BUS
betterdataset are clearly
results from depicted. that
the modifications Thisare is made
possible becauseLeNet.
in modified of the From
model s
this, ittocan
ability be claimed
obtain muchthat betterthe results
proposed modified
from LeNet model “WON”
the modifications that areinmadethe agreement
in modified
LeNet. From this, it can be claimed that the proposed modified LeNet modelperfectly
between the model and clinicians. This means that the model’s predictions align “WON” in
thewith the expert
agreement judgment
between the of the clinician.
model Overall,
and clinicians. themeans
This classifier
thatisthe
useful
modelin classifying
s predictions
breast
align tissues with
perfectly into malignant
the expert(pred:2,
judgment in Figure
of the7),clinician.
benign (pred:1,
Overall,in the
Figure 7), andis
classifier normal
useful in
(pred:0, in Figure 7). When mapped with the trained data, the predicted class is accurately
classifying breast tissues into malignant (pred:2, in Figure 7), benign (pred:1, in Figure 7),
matching.
and normal (pred:0, in Figure 7). When mapped with the trained data, the predicted class
In Table 4, the performance metrics (accuracy, validation accuracy, loss, and validation
is loss)
accurately
of various matching.
models on the BUS dataset are observed [47,48]. The performance metrics
In Table
provide insights 4, the performance
into how well each metrics
model(accuracy,
performs invalidation accuracy, loss,
terms of classification and and
accuracy valida-
tion
theloss)
amount of various
of error ormodels on the
loss during BUS dataset
training are observed
and validation. [47,48].metric
The accuracy The performance
indicates
metrics provide of
the proportion insights
correctly into how well
classified each From
samples. modelTable performs in terms LeNet
4, the modified of classification
model
accuracy
achievesand the amount
the highest accuracy of of
error or loss
0.8991, during
indicating thattraining andclassifies
it correctly validation. The accuracy
(approximately)
89.91%
metric of the samples.
indicates It outperforms
the proportion all other
of correctly modelssamples.
classified in terms ofFromaccuracy.
TableOn themodified
4, the other
hand,model
LeNet the “SCAN”
achievesmodel has theaccuracy
the highest lowest accuracy
of 0.8991, ofindicating
0.8056. The validation
that it correctlyaccuracy
classifies
metric measures89.91%
(approximately) the model’s
of theperformance
samples. Iton unseen validation
outperforms all otherdata. Like accuracy,
models in termstheof ac-
“Modified LeNet” model achieves the highest validation accuracy of 0.7652, indicating
good generalization ability. The “U-Net” model has the lowest validation accuracy of
0.7003.
Diagnostics 2023, 13, 2746 25 of 28

Table 4. Comparison of proposed model with pre-trained models on the BUS dataset.

Models Accuracy Val_Accuracy Training_Loss Val_Loss

AlexNet 0.8725 0.7325 1.6250 2.0125
SegNet 0.8127 0.7426 1.3232 2.3625
U-Net 0.8619 0.7003 1.7426 1.9981
CE-Net 0.8352 0.6982 1.3526 2.2635
SCAN 0.8056 0.7169 1.3667 1.9478
Dense U-Net 0.8102 0.7456 1.1125 1.4698
LeNet 0.8506 0.7223 1.3225 2.2215
Modified LeNet 0.8991 0.7652 1.3770 1.7707

The loss metric represents the error or discrepancy between predicted and actual
values during training. Lower loss values indicate better model performance. The “Dense
U-Net” model has the lowest training loss of 1.1125, indicating a better fit to the training
data. Conversely, the “U-Net” model has the highest loss of 1.7426. Dense U-Net has
the lowest value for validation loss which is 1.4698 and next to it, modified LeNet has
1.7707 suggesting better generalization on unseen validation data. In summary, the “Modi-
fied LeNet” model outperforms other models on the BUS dataset based on the provided
metrics It achieves the highest accuracy and validation accuracy while maintaining rela-
tively low loss values. This suggests that the “Modified LeNet” model might be the most
effective among the models considered for the BUS dataset.

9. Conclusions
Deep learning models, like CNNs, have shown promising results in breast cancer
prediction using medical images. However, a single CNN model may not consistently
achieve the desired accuracy or robustness. Batch learning is a technique that combines
predictions of multiple models to improve performance. Evaluating the ensembled LeNet
CNN for breast cancer prediction combines multiple LeNet models, each trained on the
same dataset with different initialization or hyperparameter settings. The predictions of
the individual models are combined using soft voting techniques. LeNet CNN cluster
includes comparing its performance with a LeNet model and other cutting-edge models
using metrics such as accuracy, precision, recall, and F1 score. The choice of mixing
operation and hyperparameters can affect the performance of the model. The empirical
results show the effectiveness of ensembled CNN models for breast cancer prediction. In
one study, an ensembled CNN achieved an accuracy of 85.57% on a breast cancer dataset,
outperforming a single CNN model and other state-of-the-art models. The thirteen LeNet
CNN models ensembled had an accuracy of 89.91%. The use of ensembling allows for the
combination of multiple models to mitigate the drawbacks of individual models, such as
overfitting and bias. The proposed work shows that ensembling LeNet CNN models can
significantly increase the accuracy of image classification tasks and can help reach more
excellent accuracy rates. In future, a clustering LeNet CNNs is a promising method to
improve the accuracy and robustness of breast cancer prediction using clinical images, and
their effectiveness should be further investigated.

Author Contributions: Conceptualization, S.B. and Y.V.; Data curation, Y.V.; Formal analysis, S.B.
and D.J.; Methodology, S.B., Y.V. and S.D.; Software, S.B. and S.D.; Supervision, S.D.; Validation, S.B.
and D.J.; Visualization, D.J.; Writing—original draft, S.B., Y.V. and D.J.; Writing—review and editing,
S.D. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not Applicable.
Informed Consent Statement: Not Applicable.
Data Availability Statement: Synthetic datasets [49,50] are used to conduct experiments.
Conflicts of Interest: The authors declare no conflict of interest.
Diagnostics 2023, 13, 2746 26 of 28

References
1. American Cancer Society. Breast Cancer Facts & Figures; American Cancer Society, Inc.: Atlanta, GA, USA, 2015.
2. Madani, M.; Behzadi, M.M.; Nabavi, S. The Role of Deep Learning in Advancing Breast Cancer Detection Using Different Imaging
Modalities: A Systematic Review. Cancers 2022, 14, 5334. [CrossRef] [PubMed]
3. Xie, W.; Li, Y.; Ma, Y. Breast mass classification in digital mammography based on extreme learning machine. Neurocomputing
2016, 173, 930–941. [CrossRef]
4. Hubbard, R.A.; Kerlikowske, K.; Flowers, C.I.; Yankaskas, B.C.; Zhu, W.; Miglioretti, D.L. Cumulative probability of false-positive
recall or biopsy recommendation after 10 years of screening mammography: A cohort study. Ann. Intern. Med. 2011, 155, 481–492.
[CrossRef] [PubMed]
5. Heath, M.; Bowyer, K.; Kopans, D.; Kegelmeyer, P., Jr.; Moore, R.; Chang, K.; Munishkumaran, S. Current status of the digital
database for screening mammography. In Digital Mammography; Springer: Berlin/Heidelberg, Germany, 1998; pp. 457–460.
6. Cho, N.; Han, W.; Han, B.K.; Bae, M.S.; Ko, E.S.; Nam, S.J.; Chae, E.Y.; Lee, J.W.; Kim, S.H.; Kang, B.J.; et al. Breast cancer screening
with mammography plus ultrasonography or magnetic resonance imaging in women 50 years or younger at diagnosis and
treated with breast conservation therapy. JAMA Oncol. 2017, 3, 1495–1502. [CrossRef]
7. Tsochatzidis, L.; Zagoris, K.; Arikidis, N.; Karahaliou, A.; Costaridou, L.; Pratikakis, I. Computer-aided diagnosis of mammo-
graphic masses based on a supervised content-based image retrieval approach. Pattern Recognit. 2017, 71, 106–117. [CrossRef]
8. Hamidinekoo, A.; Denton, E.; Rampun, A.; Honnor, K.; Zwiggelaar, R. Deep learning in mammography and breast histology, an
overview and future trends. Med. Image Anal. 2018, 47, 45–67. [CrossRef]
9. Arevalo, J.; González, F.A.; Ramos-Pollán, R.; Oliveira, J.L.; Lopez, M.A.G. Representation learning for mammography mass
lesion classification with convolutional neural networks. Comput. Methods Programs Biomed. 2016, 127, 248–257. [CrossRef]
10. Rampun, A.; Scotney, B.W.; Morrow, P.J.; Wang, H. Breast Mass Classification in Mammograms using Ensemble Convolutional
Neural Networks. In Proceedings of the 20th International Conference on e-Health Networking, Applications and Services
(Healthcom), Ostrava, Czech Republic, 17–20 September 2018; IEEE: Piscataway, NJ, USA, 2018; pp. 1–6.
11. Arikidis, N.; Vassiou, K.; Kazantzi, A.; Skiadopoulos, S.; Karahaliou, A.; Costaridoua, L. A two-stage method for microcalcification
cluster segmentation in mammography by deformable models. Med. Phys. 2015, 42, 5848–5861. [CrossRef]
12. Lee, R.S.; Gimenez, F.; Hoogi, A.; Miyake, K.K.; Gorovoy, M.; Rubin, D. A curated mammography data set for use in computer-
aided detection and diagnosis research. Sci. Data 2017, 4, 170–177. [CrossRef]
13. Rouhi, R.; Jafari, M.; Kasaei, S.; Keshavarzian, P. Benign and malignant breast tumor classification based on region growing and
CNN segmentation. Expert Syst. Appl. 2015, 42, 990–1002. [CrossRef]
14. Huynh, B.Q.; Li, H.; Giger, M.L. Digital mammographic tumor classification using transfer learning from deep convolutional
neural networks. J. Med. Imaging 2016, 3, 034501. [CrossRef]
15. Guleria, H.V.; Luqmani, A.M.; Kothari, H.D.; Phukan, P.; Patil, S.; Pareek, P.; Kotecha, K.; Abraham, A.; Gabralla, L.A. Enhancing
the Breast Histopathology Image Analysis for Cancer Detection Using Variational Autoencoder. Int. J. Environ. Res. Public Health
2023, 20, 4244. [CrossRef] [PubMed]
16. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. In Proceedings of the
Advances in Neural Information Processing Systems, Lake Tahoe, NV, USA, 3–6 December 2012; pp. 1097–1105.
17. Jiao, Z.; Gao, X.; Wang, Y.; Li, J. A deep feature based framework for breast masses classification. Neurocomputing 2016, 197,
221–231. [CrossRef]
18. Vidhushavarshini, S.; Sathiyabhama, B. Comparison of Classification Techniques on Thyroid Detection Using J48 and Naive
Bayes Classification Techniques. In Proceedings of the International Conference on Intelligent Computing Systems (ICICS 2017)
Organized by Sona College of Technology, Salem, Tamilnadu, India, 15–16 December 2017.
19. Zahoor, S.; Shoaib, U.; Lali, I.U. Breast Cancer Mammograms Classification Using Deep Neural Network and Entropy-Controlled
Whale Optimization Algorithm. Diagnostics 2022, 12, 557. [CrossRef]
20. Sadad, T.; Hussain, A.; Munir, A.; Habib, M.; Ali Khan, S.; Hussain, S.; Yang, S.; Alawairdhi, M. Identification of breast malignancy
by marker-controlled watershed transformation and hybrid feature set for healthcare. Appl. Sci. 2020, 10, 1900. [CrossRef]
21. Badawi, S.M.; Mohamed, A.E.-N.A.; Hefnawy, A.A.; Zidan, H.E.; GadAllah, M.T.; El-Banby, G.M. Automatic semantic segmenta-
tion of breast tumors in ultrasound images based on combining fuzzy logic and deep learning—A feasibility study. PLoS ONE
2021, 16, e0251899. [CrossRef] [PubMed]
22. Zahoor, S.; Lali, I.U.; Khan, M.A.; Javed, K.; Mehmood, W. Breast cancer detection and classification using traditional computer
vision techniques: A comprehensive review. Curr. Med. Imaging 2020, 16, 1187–1200. [CrossRef] [PubMed]
23. Lévy, D.; Jain, A. Breast mass classification from mammograms using deep convolutional neural networks. arXiv 2016,
arXiv:1612.00542.
24. Ting, F.F.; Tan, Y.J.; Sim, K.S. Convolutional neural network improvement for breast cancer classification. Expert Syst. Appl. 2019,
120, 103–115. [CrossRef]
25. Vidhushavarshini, S.; Balasubramaniam, S.; Ravi, V.; Arunachalam, A. A hybrid optimization algorithm-based feature selection
for thyroid disease classifier with rough type-2 fuzzy support vector machine. Expert Syst. 2022, 39, e12811.
26. Huang, S.; Houssami, N.; Brennan, M.; Nickel, B. The impact of mandatory mammographic breast density notification on
supplemental screening practice in the United States: A systematic review. Breast Cancer Res. Treat. 2021, 187, 11–30. [CrossRef]
[PubMed]
Diagnostics 2023, 13, 2746 27 of 28

27. Arevalo, J.; Gonzalez, F.A.; Ramos-Pollan, R.; Oliveira, J.L.; Lopez, M.A.G. Convolutional neural networks for mammography
mass lesion classification. In Proceedings of the 2015 37th Annual International Conference of the IEEE Engineering in Medicine
and Biology Society (EMBC), Milan, Italy, 25–29 August 2015; pp. 797–800.
28. Duraisamy, S.; Emperumal, S. Computer-aided mammogram diagnosis system using deep learning convolutional fully complex-
valued relaxation neural network classifier. IET Comput. Vis. 2017, 11, 656–662. [CrossRef]
29. Chai, J.; Zeng, H.; Li, A.; Ngai, E.W. Deep learning in computer vision: A critical review of emerging techniques and application
scenarios. Mach. Learn. Appl. 2021, 6, 100134. [CrossRef]
30. Borah, N.; Varma, P.S.P.; Datta, A.; Kumar, A.; Baruah, U.; Ghosal, P. Performance Analysis of Breast Cancer Classification from
Mammogram Images Using Vision Transformer. In Proceedings of the 2022 IEEE Calcutta Conference (CALCON), Kolkata, India,
10–11 December 2022; pp. 238–243. [CrossRef]
31. Reenadevi, R.; Sathiyabhama, B.; Vinayakumar, R.; Sankar, S. Hybrid Optimization Algorithm based feature selection for
mammogram images and detecting the breast mass using Multilayer Perceptron classifier. J. Comput. Intell. 2022, 38, 1559–1593.
[CrossRef]
32. Manikandan, P.; Durga, U.; Ponnuraja, C. An integrative machine learning framework for classifying SEER breast cancer. Sci. Rep.
2023, 13, 5362. [CrossRef]
33. Sun, W.; Tseng, T.L.; Zhang, J.; Qian, W. Enhancing Deep Convolutional Neural Network Scheme for Breast Cancer Diagnosis
with Unlabeled Data. Comput. Med. Imaging Graph. 2017, 57, 4–9. [CrossRef]
34. Khan, S.; Khan, M.A.; Alhaisoni, M.; Tariq, U.; Yong, H.-S.; Armghan, A.; Allenzi, F. Human Action Recognition: Paradigm of
Best Deep Learning Features Selection and Serial Based Extended Fusion. Sensors 2021, 21, 7941. [CrossRef]
35. Khan, M.A.; Alhaisoni, M.; Tariq, U.; Hussain, N.; Majid, A.; Damaševičius, R.; Maskeliunas, R. COVID-19 Case Recognition from
Chest CT Images by Deep Learning, Entropy-Controlled Firefly Optimization, and Parallel Feature Fusion. Sensors 2021, 21, 7286.
[CrossRef]
36. Shervan, F.-E.; Alsaffar, M.F. Developing a Tuned Three-Layer Perceptron Fed with Trained Deep Convolutional Neural Networks
for Cervical Cancer Diagnosis. Diagnostics 2023, 13, 686.
37. Tan, Y.N.; Tinh, V.P.; Lam, P.D.; Nam, N.H.; Khoa, T.A. A Transfer Learning Approach to Breast Cancer Classification in a
Federated Learning Framework. IEEE Access 2023, 11, 27462–27476. [CrossRef]
38. Patil, N.S.; Desai, S.D.; Kulkarni, S. Magnification independent fine-tuned transfer learning adaptation for multi-classification of
breast cancer in histopathology images. In Proceedings of the 2022 4th International Conference on Advances in Computing,
Communication Control and Networking (ICAC3N), Greater Noida, India, 16–17 December 2022; pp. 1185–1191. [CrossRef]
39. Sanyal, R.; Kar, D.; Sarkar, R. Carcinoma Type Classification from High-Resolution Breast Microscopy Images Using a Hybrid
Ensemble of Deep Convolutional Features and Gradient Boosting Trees Classifiers. IEEE/ACM Trans. Comput. Biol. Bioinform.
2022, 19, 2124–2136. [CrossRef]
40. Bagchi, A.; Pramanik, P.; Sarkar, R. A Multi-Stage Approach to Breast Cancer Classification Using Histopathology Images.
Diagnostics 2023, 13, 126. [CrossRef] [PubMed]
41. Kumar, D.; Batra, U. Breast Cancer Histopathology Image Classification Using Soft Voting Classifier. In Lecture Notes in Networks
and Systems, Proceedings of the 3rd International Conference on Computing Informatics and Networks, Delhi, India, 29–30 July 2020;
Abraham, A., Castillo, O., Virmani, D., Eds.; Springer: Singapore, 2021; Volume 167, p. 167. [CrossRef]
42. Shahidi, F.; Abas, H.; Ahmad, N.A.; Maarop, N. Breast Cancer Classification Using Deep Learning Approaches and Histopathology
Image: A Comparison Study. IEEE Access 2020, 8, 187531–187552. [CrossRef]
43. Kathale, P.; Thorat, S. Breast Cancer Detection and Classification. In Proceedings of the 2020 International Conference on Emerging
Trends in Information Technology and Engineering (ic-ETITE), Vellore, India, 24–25 February 2020; pp. 1–5. [CrossRef]
44. Farhadi, A.; Chen, D.; McCoy, R.; Scott, C.; Miller, J.A.; Vachon, C.M.; Ngufor, C. Breast Cancer Classification using Deep Transfer
Learning on Structured Healthcare Data. In Proceedings of the 2019 IEEE International Conference on Data Science and Advanced
Analytics (DSAA), Washington, DC, USA, 5–8 October 2019; pp. 277–286. [CrossRef]
45. Rashid, M.; Khan, M.A.; Alhaisoni, M.; Wang, S.H.; Naqvi, S.R.; Rehman, A.; Saba, T.A. Sustainable deep learning framework for
object recognition using multi-layers deep features fusion and selection. Sustainability 2020, 12, 5037. [CrossRef]
46. Sathiyabhama, B.; Gopikrishna, K.; Jayanthi, J.; Sathiya, T.; Ilavarasi, A.K.; Kumar, S.U.; Yuvarajan, V. Automatic Breast Region
Extraction and Pectoral Muscle Removal in Mammogram Using Otsu’s Threshold with Connected Component Labelling. In
Proceedings of the 3rd World Conference on Applied Science, Engineering and Technology, Singapore, 29–30 June 2017.
47. Sathiyabhama, B.; Kumar, S.U.; Jayanthi, J.; Sathiya, T.; Ilavarasi, A.K.; Yuvarajan, V.; Gopikrishna, K. Breast Cancer Histopathol-
ogy Image Analysis using ConvNets and Machine Learning Techniques. In International Conference on Machine Learning, Image
Processing, Network Security and Data Sciences; IETE Springer Series; Springer: Berlin/Heidelberg, Germany, 2019.
48. Hazarika, R.A.; Abraham, A.; Kandar, D.; Maji, A.K. An improved LeNet-deep neural network model for Alzheimer’s disease
classification using brain magnetic resonance images. IEEE Access 2021, 9, 161194–161207. [CrossRef]
49. Sathiyabhama, B.; Kumar, S.U.; Jayanthi, J.; Sathiya, T.; Ilavarasi, A.K.; Yuvarajan, V.; Gopikrishna, K. A Novel Feature Selection
Framework Based on Grey Wolf Optimizer for Mammogram Image Analysis. Neural Comput. Appl. 2021, 33, 14583–14602.
[CrossRef]
Diagnostics 2023, 13, 2746 28 of 28

50. Available online: https://www.kaggle.com/datasets/aryashah2k/breast-ultrasound-images-dataset (accessed on 13 March 2023).

51. Nahid, A.A.; Mehrabi, M.A.; Kong, Y. Histopathological Breast Cancer Image Classification by Deep Neural Network Techniques
Guided by Local Clustering. Biomed Res. Int. 2018, 2018, 2362108. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.