Image and Vision Computing 142 (2024) 104910

Contents lists available at ScienceDirect

Image and Vision Computing

journal homepage: www.elsevier.com/locate/imavis

A novel approach for breast cancer detection using optimized ensemble

learning framework and XAI
Raafat M. Munshi a, Lucia Cascone b, Nazik Alturki c, Oumaima Saidani c, Amal Alshardan c,
Muhammad Umer d, *
a
Department of Medical Laboratory Technology (MLT), Faculty of Applied Medical Sciences, King Abdulaziz University, Rabigh, Saudi Arabia
b
Department of Computer Science, University of Salerno, Fisciano, Italy
c
Department of Information Systems, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, P.O. Box 84428, Riyadh 11671,
Saudi Arabia
d
Department of Computer Science & Information Technology, The Islamia University of Bahawalpur, Bahawalpur 63100, Pakistan

A R T I C L E I N F O A B S T R A C T

Keywords: Breast cancer (BC) is a common and highly lethal ailment. It stands as the second leading contributor to cancer-
Breast cancer detection related deaths in women worldwide. The timely identification of this condition is of utmost importance in
Image processing mitigating mortality rates. This research paper presents a novel framework for the precise identification of BC,
Healthcare
utilising a combination of image and numerical data features with explainable Artificial Intelligence (XAI). The
Transfer learning
Ensemble learning
utilisation of the U-NET transfer learning model is employed for image-based prediction. Additionally, an
Deep convoluted features ensemble model is constructed by integrating characteristics from a customised convolutional neural network
(CNN) model with an ensemble comprising random forest (RF) and support vector machine (SVM). The exper­
iments aim to evaluate the influence of original features compared to convoluted features. A comparative
analysis is carried out to assess the efficacy of various classifiers in accurately detecting BC, utilising the Wis­
consin dataset. The model under consideration exhibits promising capabilities in enhancing BC diagnosis, with a
remarkable accuracy rate of 99.99%. The present study contributes to the advancement of BC diagnosis by
introducing a novel strategy based on machine learning and discussing the interpretation of the variables using
XAI. The primary objective of this approach is to get a notable level of precision, hence facilitating the early and
reliable identification of BC. Ultimately, the implementation of this approach is expected to enhance patient
outcomes.

1. Introduction critical in reducing the death rate associated with this condition. Ac­
cording to Globocan 2018 data [5], breast cancer is found in one out of
Cancer is a worldwide flagellum, leading to millions of deaths per every four cases of cancer in women, making it the sixth most prevalent
year [1]. As per GLOBOCAN 2020 estimates of cancer incidence and cause of death globally. Breast cancer often develops in the breast tissue,
mortality produced by the International Agency for Research on Cancer especially the inner linings of milk ducts or lobules. Changes or muta­
[2], approximately 2,261,419 new cases of BC were diagnosed world­ tions in Deoxyribonucleic acid (DNA) or Ribonucleic acid (RNA) cause
wide in 2020; there were 24.5% of cancer cases and 15.5% of cancer cancer cell formation. Air pollutants, bacteria, nuclear radiation, fungi,
deaths worldwide among women [3]. Saudi Health Council data in­ physical cell-level damage, viruses, parasites, high temperatures, water
dicates that BC accounted for 15.9% of all reported cancers among Saudi contaminants, electromagnetic radiation, dietary elements, free radi­
citizens and 28.7% among women of all ages. According to the World cals, aging of DNA and RNA, and genetic changes are all factors that can
Health Organization (WHO), breast cancer affects approximately 2.3 lead to breast cancer. Most breast cancers are no special type (NST), also
million women each year, resulting in 685,000 deaths, accounting for called invasive ductal carcinomas (IDCs), followed by invasive lobular
13.6% of all female cancer-related fatalities [4]. Prompt detection is carcinomas (ILCs) [6], inflammatory breast cancer (IBC) [7], and

* Corresponding author.
E-mail addresses: [email protected] (R.M. Munshi), [email protected] (L. Cascone), [email protected] (N. Alturki), [email protected] (O. Saidani),
[email protected] (A. Alshardan), [email protected] (M. Umer).

https://doi.org/10.1016/j.imavis.2024.104910
Received 24 September 2023; Received in revised form 23 November 2023; Accepted 11 January 2024
Available online 20 January 2024
0262-8856/© 2024 Elsevier B.V. All rights reserved.
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

lobular breast cancer (LBC) [8]. The subsequent sections of this study are outlined as follows: Section
Breast cancer, a dangerous disease with high death rates, accounts 2 encompasses recent literature on breast cancer diagnosis and detec­
for 2.5% of all fatalities, impacting one out of every thirty-nine women tion. Section 3 elaborates on the dataset, the proposed methodology, and
[9]. Early identification and treatment of breast cancer are critical since the machine learning classifiers employed. Section 4 comprises results
the disease can spread to other parts of the body if left untreated. Early and a comparative analysis. Discussions are presented in Section 5.
detection and treatment can increase the chance of survival by up to Lastly, Section 6 encapsulates the conclusion and outlines avenues for
80%. This emphasizes the critical need of early identification and future research.
treatment of breast cancer. Various procedures and methodologies, such
as screening examinations, self-assessments, and frequent visits with 2. Related work
healthcare specialists, help to detect breast cancer early [10].
Mammography is a popular and efficient approach for detecting breast Many approaches for BC detection and diagnosis have emerged last
cancer early. Several studies have confirmed its efficacy in detecting years. Recent research has primarily focused on identifying malig­
early-stage breast cancer. Another often-used approach for identifying nancies through CAD and decision support systems. While the majority
breast cancer is a biopsy. Tissue is removed from the afflicted breast of investigations depend on individual models to attain dependable re­
region and examined under a microscope to detect and categorize the sults, a handful explore the utilization of ensemble approaches. This
tumor during a biopsy [11]. Biopsy is an often-utilized approach for section presents an overview of these approaches.
identifying BC. During a biopsy, tissue is removed from the afflicted To achieve precise and accurate breast cancer diagnosis, Yadav and
region of the breast and examined under a microscope in order to di­ Jadhav [17] proposed a machine learning-driven system employing
agnose and categorise the tumour. Although present approaches help to thermal infrared imaging. The researchers evaluated various baseline
reduce breast cancer death rates, there is still room for improvement, models alongside transfer learning models such as VGG16 and Incep­
particularly in terms of more simplified and automated diagnostic pro­ tionV3. Their experiments encompassed both augmented and non-
cesses. Data mining is a technique used to extract important and note­ augmented data scenarios. The findings reveal that the transfer
worthy insights from large data collections. It is recognized as a critical learning model InceptionV3 outperforms other models, achieving an
tool for the early detection of several ailments such as renal disease, accuracy rate of 93.1% without augmentation and 98.5% with
heart disease [12], diabetes [13], and cancer. Using data mining tech­ augmentation. In a separate study [18], researchers leveraged genetic
niques makes it easier to recognize patterns and trends in data, which programming to optimize features for automated breast cancer diag­
aids in the early diagnosis and treatment of certain disorders. Its nosis. They experimented with nine machine learning classifiers. The
importance grows when it comes to recognizing illnesses like cancer, outcomes highlight genetic programming’s effectiveness in amalgam­
when early identification improves survival chances dramatically. ating preprocessing and feature selection, leading to the identification of
Traditional cancer detection consists on three techniques: physical ex­ the best model. The AB classifier attained the highest accuracy score of
amination, pathology testing, and radiological imaging. These old ap­ 98.23%. In another research work, researchers proposed an approach
proaches, however, are time-consuming and prone to false negatives. for BC detection using extreme learning machine based on feature fusion
Machine learning approaches, on the other hand, are gaining popularity with CNN deep features. [19]
due to their superior results. They provide dependability, precision, and The authors devised an automated breast cancer detection system
speed, revolutionizing disease detection beyond traditional methods and employing various models like LR, KNN, SVM, and diverse CNN varia­
giving more consistent findings across a wide range of ailments. tions [20]. During experimentation, the authors scrutinized zones of
Over the past decade, recent advances in machine learning have hostile ductal carcinoma tissue within entire slide images. Results
opened up new areas for the medical fields. Machine learning algorithms indicated that the CNN variant attained the highest accuracy at 87%,
require the capacity to comprehend predictions, which is referred to as surpassing the accuracy of machine learning models by 9%. This sug­
explainable artificial intelligence (XAI). Justification of specific out­ gests that the proposed deep learning system significantly enhances
comes yields valuable, understandable insights, facilitating informed breast cancer detection accuracy. The researchers introduced a voting
decision-making and knowledge generation. Researcher applied XAI in classifier based on ensemble learning for breast cancer detection [21].
detecting breast cancer [14,15]. 1n a recent study authors applied ma­ Their findings demonstrated that the ensemble learning model out­
chine learning with XAI for deting breast cancer in Indonesian patients performed individual machine learning models. A study proposed a
[16]. Due to the aforementioned benefits, this study proposes a machine breast tumor type detection approach and achieved 98.1% accuracy
learning-based approach leveraging XAI for detecting breast cancer with [22]. Another study utilized diverse-density mammograms in breast
the goal of achieving high accuracy. This study makes the following cancer detection, achieving an overall accuracy score of 88.1% [23].
contributions in this regard. Diverse approaches and techniques, including Breast Magnetic
Resonance Imaging (MRI), Mammography, breast ultrasound, biopsy,
• This study presents a complete framework for accurate breast cancer breast self-exam, and clinical breast exam, help in the timely detection
diagnosis that makes use of both image analysis and numerical data of cancer. Breast mammography and ultrasound are typical screening
characteristics. For image-based prognosis, the predictive model methods. [24]. MRI is often used in conjunction with mammography or
employs U-NET transfer learning. as a supplemental method for women with dense breasts or high-risk BC
• An ensemble strategy is proposed in this study to anticipate breast cancer. Ultrasound has played an important role in BC detection for
cancer among patients by merging features from a personalized many years thanks to its portability, availability, and their ability to
convolutional neural network (CNN) model. The proposed ensemble assess a variety of important aspects from multiple planes with high
model combines random forest (RF) and support vector machine resolution. These approaches focused on combining machine learning
(SVM) techniques, with the final prediction determined by a voting methods with various feature extraction techniques to enhance accuracy
process. in breast mass classification. Results showed high accuracy rates, with
• The experiments involve utilizing both the inherent features from the DLA-EABA achieving 97.2% accuracy. Aljuaid et al. proposed a transfer
dataset and those derived from the CNN model. The study aims to learning-based approach in [25] for breast cancer classification. The
assess the influence of the original features in comparison to the authors conducted the experiment in two ways: binary classification and
performance of models utilizing derived features. multi-class classification.
• Moreover, the efficacy of the proposed model is assessed by con­ Mangukiya et al. [26] conducted a study focused on efficient and
trasting its performance with state-of-the-art methodologies con­ accurate early breast cancer diagnosis employing multiple techniques.
cerning accuracy, precision, recall, and the F1 score. The authors employed a range of machine learning algorithms on a

2
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

dataset containing features with diverse units and magnitudes, requiring contains 32 features including ‘Texture SE’, ‘Texture Mean’, ‘Concave
standardization through standard scaling. Their findings highlighted Points Mean’, ‘Concave Points SE’, ‘ID’, ‘Area Worst’, ‘Smoothness
that, with standard scaling, the XGBoost machine learning algorithm Mean’, ‘Symmetry Worst’, ‘Compactness SE’, ‘Radius Mean’, ‘Texture
achieved an accuracy score of 98.24%. Similarly, [27] introduced a deep Worst’, ‘Concave Points Worst’, ‘Perimeter SE’, ‘Fractal Dimension SE’,
ensemble learning model for breast cancer detection using whole slide ‘Area Mean’, ‘Perimeter Worst’, ‘Fractal Dimension Mean’, ‘Compact­
images. They utilized various deep-learning models and proposed an ness Worst’, ‘Compactness Mean’, ‘Radius Worst’, ‘Perimeter Mean’,
ensemble model. The outcomes demonstrated the superiority of the ‘Concavity SE’, ‘Smoothness SE’, ‘Fractal Dimension Worst’, ‘Concavity
ensemble model over other models, achieving an accuracy score of Mean’, ‘Smoothness Worst’, ‘Symmetry Mean’, ‘Symmetry SE’, ‘Area
86.21%. XAI has been used in many cancer detection tasks like ovarian SE’, ‘Radius SE’, ‘Concavity Worst’, ‘Diagnosis’ (target class). Within the
cancer analysis [28], cancer biomarkers evaluation [29], and cervical dataset, two distinct target classes exist: benign and malignant. Sample
cancer cell migration [30]. This research work is a continuation of our distribution indicates a composition of 45% in the malignant class and
previously published paper [31]. Muhammad Umer the co-author of 55% in the benign class. The dataset’s 32 features encompass various
both works helped us in implementing feature based analysis and pre­ classifications like numeric, nominal, and binary, among others.
diction of breast cancer in this research work and further extension of Notably, the target class represents a categorical attribute, while the rest
breast cancer to image based analysis too. of the attributes are numeric.
Numerous investigations have delved into employing machine
learning models to detect breast cancer, seeking to enhance classifica­
tion accuracy and minimize pathological errors in automated diagnosis. 3.2. Machine learning models for breast cancer prediction
Table 1 provides an overview of select literature pertaining to breast
cancer detection through machine learning models. Machine learning classification involves supervised learning,
wherein the system assimilates knowledge from a particular dataset to
3. Materials and methods categorize new observations. The dataset can comprise binary or multi-
class classifications. This section addresses machine learning classifiers
This section covers the dataset utilized for breast cancer detection, employed in breast cancer detection. The implementation of these
outlines the proposed approach, and details the methodology steps. models is executed using the scikit-learn library within the Python
Additionally, it provides a concise overview of the machine learning environment.
classifiers incorporated in this research.
3.2.1. Random Forest
Random Forest (RF) stands as a prevalent ensemble learning tech­
3.1. Dataset for experiments
nique in machine learning, extensively applied for classification and
regression tasks [33,34]. It operates by amalgamating multiple decision
This research applies supervised machine learning models to detect
trees, wherein each tree’s outputs contribute to the ultimate prediction.
breast cancer, aiming to assess their performance. The study proceeds
The core principle involves training numerous decision trees, each on a
through several stages, starting with the acquisition of the dataset [32].
distinct subset of the data, and subsequently amalgamating their pre­
Specifically, the “Breast Cancer Wisconsin Dataset” is sourced from the
dictions to derive the final output.
UCI machine learning repository, accessible to the public. The dataset

3.2.2. Decision Tree

Table 1
Decision Trees (DT) stand among the most prevalent methods for
Summary of the discussed research works.
classification and prediction [35]. Represented as a tree-like structure
Ref. Models Dataset Achieved accuracy akin to a flowchart, a DT showcases logical steps. Within this structure,
[17] Baseline models PROENG dataset 93.1% without internal nodes portray attribute tests, branches signify attribute test
and transfer augmentation; 98.5% outcomes, and leaf nodes indicate class labels. Decision trees offer sig­
learning models with augmentation
nificant advantages in data classification, executing this task swiftly and
(VGG16 and with Inception V3
Inception V3) with minimal computational demands.
[18] k-NN, SVM, GB, Wisconsin Breast Cancer 98.23% with AB
GNB, DT, RF, LR, dataset 3.2.3. K-nearest neighbor
ADA, and LDA The k-NN algorithm, a non-parametric method in machine learning
[20] LR, KNN, SVM, Kaggle 162 H&E 87% CNN model 3;
CNN variants 78.56% SVM
applicable to regression and classification tasks, operates through lazy
[21] RF, KNN, DT, SVM, Breast Cancer Wisconsin 100% with (LR + SGD) learning or instance-based learning [36]. It identifies the k closest
LR, GBM, proposed Dataset training instances to a new data point and determines the majority class
(LR + SGD) among these k neighbors to classify the new data point. Grounded on the
[24] Deep Learning https://wiki. 97.2% using DLA-
similarity concept between input and training data, the algorithm stores
based model (DLA- cancerimagingarchive. EABA
EABA) net/ all available cases and employs a similarity measure, often a distance
[25] ResNet, Inception- BreakHis 99.7% for binary function, to classify new cases. Renowned for its simplicity, the k-NN
V3Net, and classification with algorithm proves straightforward and easily implementable.
ShuffleNet ResNet; 97.81% for
multi-class using
ResNet
3.2.4. Logistic Regression
[26] RF, k-NN, DT, SVM, Wisconsin breast cancer 98.24% using XGboost Logistic Regression (LR) serves as a statistical model employed in
NB, XGBoost, ADA Dataset supervised learning specifically for binary classification problems. Its
[27] CNN, DNN, LSTM, Histopathologic Cancer 86.21% CNN-GRU prevalent use occurs when the outcome variable is binary, such as pre­
GRU, BiLSTM, Detection
dicting the presence or absence of a disease in a patient, or determining
CNN-GRU
[22] DT, SVM, RF, LR, k- The University of 98.1% using logistic if an email is spam or not. LR aims to gauge the probability of a binary
NN, NB and Wisconsin Hospital regression outcome considering specific inputs, utilizing this estimation for pre­
rotation forest dataset dictions. Employing the logistic function, also known as the sigmoid
[23] EfficientNet-B5, Hallym University Sacred 88.1% DenseNet-169 function, LR models the probability of a binary outcome, utilizing the
DenseNet-169 Heart Hospital dataset
resulting output to make predictions [33,37].

3
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

3.2.5. Support Vector Machine medical data processing.

The Support Vector Machine (SVM), a recognized supervised
learning algorithm [38], is employed in machine learning for classifi­ 3.3.1. Multilayer perceptron neural network
cation and regression purposes. SVM’s core principle revolves around For relatively small training sets, along with considerations for ease
identifying the optimal boundary, referred to as a hyperplane, that of implementation, rapidity, and swift outcomes, the Multi-Layer Per­
segregates data points into distinct classes. This boundary is constructed ceptron emerges as the optimal selection [38]. The inherent architecture
to maximize the margin, representing the distance between the bound­ of MLP encompasses three layers: input, output, and hidden layers. The
ary and the nearest data points from each class, often referred to as hidden layer acts as an intermediary linking the input and output layers
support vectors. SVM demonstrates applicability to both linear and non- during neuron processing. The fundamental operation within MLP re­
linear classification tasks. volves around the multiplication of input neurons by weights wij , where
the output yj is the resultant sum. Mathematically, its computation is
3.2.6. Gradient boosting machine expressed as follows:
GBM stands as a machine learning algorithm employed for classifi­ (∑ )
cation and regression tasks, falling under the boosting ensemble learning yj = f wij *Oi ,
family [39]. GBM amalgamates forecasts from multiple weak models,
like decision trees, to generate a robust model. The essence of gradient Within this equation, the weight assignments w correspond to the
boosting involves iteratively training weak models, such as decision gradient descent algorithm, while the hidden layers are symbolized by
trees, and sequentially integrating them into the ensemble. Each new O.
tree is trained to rectify the errors made by the preceding trees, focusing
on the training instances that were previously misclassified. The pre­ 3.4. RNN
dictions from all trees are amalgamated to derive the ultimate
prediction. The Recurrent Neural Network (RNN) is the most popular choice for
investigating sequential neural networks [43]. When processing, the
3.2.7. Extra tree classifier input sequence from one neuron is passed along to the next in a
ETC operates as an ensemble learning technique employing ran­ weighted sequence that mimics the word order in a phrase. The structure
domized trees to produce a conclusive classification outcome by amal­ of RNN sequences is designed to produce the sequence and predict the
gamating independent trees within a decision tree forest [40]. Its next word in the loop.
foundational idea aligns with RF, yet ETC diverges in constructing de­
cision trees within the forest. In ETC, decision-making involves 3.4.1. Convolutional neural network
randomly selecting a subset of the top K features and determining the The CNN is a potent neural network design that is skilled at under­
optimal solution using the Gini index. This methodology guarantees the standing complex correlations between various data properties. In its
creation of uncorrelated trees within ETC. Gini feature importance as­ capacity as a deep learning model, the CNN carefully evaluates incoming
sumes a pivotal role in the process of feature selection. pictures, assigns weight to unique objects and characteristics within
these images, and makes distinctions between them. The CNN archi­
3.2.8. Gaussian Naive Bayes tecture, which consists of input, output, node, and hidden layers, is vital.
GNB is a popular machine learning technique that is used for clas­ In order to provide better results, this study uses a customized CNN
sification tasks based on Bayes’ Theorem. The likelihood of a hypothesis architecture [44]. The eight layers in the suggested architecture—two
(class label) given particular evidence (feature values) equals the like­ dense, two max-pooling, and two convolution layers—play a crucial
lihood of the evidence given the hypothesis multiplied by the prior role. In terms of medical categorization, CNN has exceptional efficacy
probability of the hypothesis, according to this theorem [41]. The word and precision. The CNN model shows its usefulness by using the Sigmoid
‘naive’ in GNB refers to the algorithm’s strong premise, known as class as the error function and being supported by a backpropagation method.
conditional independence, which states that while the class name is CNN is useful for classifying a wide range of illnesses, including heart
known, all characteristics are independent. Although this assumption disease, lung disorders, and brain tumors. At the moment, it is essential
seldom matches real-world conditions, the method frequently gives for handling large-scale medical datasets. There are two pooling possi­
reliable results in actual applications. bilities in CNN’s pooling layer: average pooling works well for extracting
flat features, while maximal pooling works well for sharp feature
3.2.9. Stochastic gradient decent extraction.
SGD serves as an optimization algorithm employed to reduce a
function, especially during model training in machine learning appli­
cations like linear regression, logistic regression, and neural networks 3.5. Long short term memory
[42]. It represents a derivative of the gradient descent (GD) algorithm
and earns its stochastic nomenclature due to utilizing a random data A refined RNN variant called LSTM shows improved performance
subset, termed a mini-batch, for estimating the gradient at every itera­ when managing long-term sequences.CitePascone et al. (2023). With
tion. The primary objective of the SGD algorithm revolves around LSTM, you can overcome the vanishing gradient issue that RNNs have,
adjusting the model parameters in a direction contrary to the gradient of and you can also see better performance and pattern memorization.
the loss function concerning the parameters, employing a constant step With its three key gates—the forget, output, and input gates—LSTM
size designated as the learning rate. represents a noteworthy development. The following equations provide
examples of word sequences.
( )
3.3. Deep learning models for breast cancer prediction it = σ xt U i + ht− 1 W i + bi (1)

Deep learning is an increasingly explored subject of artificial intel­ ot = σ (xt U o + ht− 1 W o + bo ) (2)
ligence that has shown tremendous potential in data modeling. The
( )
incorporation of automated techniques by medical specialists has shown ft = σ xt U f + ht− 1 W f + bf (3)
to be quite beneficial and helpful in illness diagnosis. Deep learning is a
popular approach for dealing with large amounts of data, eliminating where xt is the input sequence, ht− 1 is the preceding hidden state at
the need for manual feature extraction and finding significant use in current step t, it is the input gate, ot is the output gate and ft is the forget

4
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

gate. accompanied by a kernel size of 2 × 2. The ReLU activation function

serves to nullify negative values to zero while preserving all other values
3.5.1. Architecture of convolutional neural network for feature extraction unchanged.
In this research focusing on breast cancer detection, the CNN deep
f (x) = max(0, E)s (7)
learning model is employed as a technique for feature extraction [44].
CNN stands as a widely utilized deep learning framework, primarily The architecture of the used CNN together with the prediction model
applied in classification tasks. Leveraging its feature extraction capa­ are shown in Fig. 1.
bilities, convoluted features are harnessed for breast cancer detection.
The conventional CNN model comprises four layers: the pooling layer, 3.6. Methodology proposed for numerical data features
embedding layer, convolutional layer, and flattening layer. In this breast
cancer detection framework, the initial layer utilized in CNN is the Ensemble models are becoming more popular and have demon­
embedding layer, boasting an embedding size of 20, 000 and an output strated increased effectiveness and accuracy in classification tasks.
dimension of 300. Subsequently, the second layer manifests as the Combining different classifiers can make it possible to outperform in­
convolutional layer, equipped with 5000 filters, a 2 × 2 kernel size, and dividual models. An ensemble learning approach is used in this study to
employing a rectified linear unit (ReLU) as its activation function. improve breast cancer diagnosis. The suggested method uses a voting
Following this, the third layer constitutes the max pooling layer, classifier that uses the soft voting criterion to merge RF and SVM. The
employing a 2 × 2 size for extracting significant feature maps from the class with the greatest number of votes determines the final forecast. The
convolutional layer’s output. Ultimately, a flatten layer is employed to described ensemble model operates as follows, as seen in Algorithm 1:
convert the output into a 1D array conducive for the learning models.
Consider a tuple set (fsi , tci ) originating from the breast cancer ∑ ∑
n n
p = argmax
̂ RFi , SVM i (8)
dataset, where fs denotes the feature set, tc signifies the target class i i
column, and I represents the index of the tuple. To facilitate the trans­
formation of the training set into the necessary input, the embedding
layer is utilized. Algorithm 1. Ensemble Voting Classifier for Breast Cancer Prediction.
EL = embedding layer(Vs, Os, I) (4)

EOs = EL(fs) (5)

Here, the embedding layer is denoted by EL, while the output of the
embedding layer is indicated by EOs . The typical layer receives this
output as its input. Three independent factors control EL: Vs , which is the
vocabulary size; I, which is the input length; and Os, which is the output
dimension. The EL size in this breast cancer detection research is
20, 000, meaning that it can accept inputs in the range of 0 to 20, 000.
The output dimension, Os, is defined as 300, while the input length is set
at 32. All of the incoming data is processed by EL, which also provides
the output that CNN needs for its further processing steps. EOs = (None,
32, 300) is the expression for the output dimension of EL.
1D − Convs = CNN(F, Ks, AF)←EOs (6) The predictive probabilities for individual test samples are generated
∑ ∑
by ni RFi and ni SVMi . These probabilities, as depicted in Fig. 2, un­
The output of the convolutional layer derives from the output of EL.
dergo the soft voting criterion, generating probabilities for each test case
The CNN structure integrates 500 filters denoted as F = 500,
through RF and SVM. The voting process is visualized in Fig. 3.

Fig. 1. Architecture diagram of the CNN with voting classifier (RF + SVM) model.

5
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

Fig. 2. Workflow diagram of the proposed voting classifier (RF + SVM) model.

Fig. 3. Architecture of the proposed voting classifier (RF + SVM) model.

VC(RF + SVM) = argmax(g(x) ) (9) expansive datasets to effectively address new tasks. When applied to
brain tumor classification via image data, two frequently employed
To assess the performance of the proposed VC(RF + SVM) model, it
transfer learning models are U-Net and MobileNet.
undergoes testing on the’Breast Cancer Wisconsin Dataset’ in a two-
stage evaluation. Initially, breast cancer detection is executed using all
3.7.1. U-Net
32 features available in the dataset. In the subsequent experimental
U-Net, brought forth by Olaf Ronnenberg and associates in 2015,
stage, the dataset is preprocessed to employ convolutional features for
stands as a significant model in the realm of image segmentation.
machine learning models. The data is partitioned into two segments,
Originally tailored for biomedical image analysis, U-Net swiftly
with 70% allocated for training purposes and the remaining 30%
garnered acclaim due to its impressive accuracy and efficacy [45]. Its
reserved for testing. This strategy, recognized as the training–testing
notable strength lies in achieving remarkable results, even when con­
split, serves as a common methodology in machine learning, facilitating
fronted with a scarcity of training data, a common challenge in medical
the evaluation of the model’s accuracy on new and unseen data.
image segmentation. The model’s architecture embodies two primary
pathways, resembling an auto-encoder. The initial pathway, termed the
3.7. Transfer learning models contracting or compressive path, functions as the encoder, utilizing a
standard deep CNN network. Conversely, the decoder or expanding path
Transfer learning stands as a widely embraced method in machine (also known as the up-sampling or synthesizing path in some literature)
learning and computer vision, utilizing pre-trained models from integrates deconvolutional and convolutional layers. The contracting

6
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

path reduces the resolution of input images, while the expanding path indicating the minimum and maximum values, respectively.
skillfully restores the original image quality and spatial structure via
TP + TN
optimized techniques, such as concatenated skip connections. Accuracy = (12)
TP + TN + FP + FN
As the network progresses, it assimilates spatial classification infor­
mation, delivering dense predictions at higher resolutions along the TP
ascending path. It also enhances the output image’s resolution, leading Precision = (13)
TP + FP
to final processing through a convolutional layer, yielding a segmented
image mirroring the dimensions of the input image. This process Recall =
TP
(14)
maintains the constant shape of the input image throughout segmenta­ TP + FN
tion, ensuring precise representation. While classification remains
Precision × Recall
essential in medical image analysis, it lacks pixel-level contextual rep­ F1score = 2 × (15)
resentation, providing a single label for the entire image. U-Net, in Precision + Recall
conjunction with subsequent refinement techniques, was explicitly
4. Results
engineered to proficiently handle multi-dimensional tensors, accom­
modating inputs with three or four dimensions. The network produces
A series of experiments were undertaken to identify breast cancer,
an output mirroring the input’s shape. Since its inception, U-Net has laid
employing machine learning models with both the original features and
a robust foundation for extensive exploration in medical image seg­
the convoluted features. The outcomes were analyzed, and further ex­
mentation. Numerous advancements have surfaced, either through ad­
periments were conducted using an ensemble comprising the top four
justments to the original U-Net structure or its integration with
individual machine learning models, applied to both sets of features.
alternative architectures.

3.7.2. Mobilenet 4.1. Experiment set up

MobileNet is a streamlined CNN structure, aims to facilitate efficient
processing on resource-constrained devices, such as mobile phones [46]. Experiments are conducted on a system with an Intel Core i7 pro­
It capitalizes on depthwise-separable convolutions, significantly cessor and a graphics processing unit (GPU) from NVIDIA using the
reducing parameter count and computational complexity while up­ Python 3.8 programming environment. Table 2 contains the specifica­
holding model effectiveness. Due to its efficiency, MobileNet is widely tions of the hardware and software used in the experiment. The pro­
applied in diverse image classification tasks. Leveraging transfer posed model took around 57 min to get trained and making predictions.
learning with a pre-trained MobileNet model enables the extraction of
valuable features from extensive image datasets, thereby enhancing the 4.2. Performance comparison of individual machine learning models on
accuracy of brain tumor classification. The MobileNet architecture is original and convoluted features
meticulously designed to excel in both efficiency and effectiveness,
especially in scenarios with limited computational resources or tasks The present study utilizes a total of nine machine learning models,
requiring minimal features, such as palmprint recognition. Its salient each fine-tuned with optimized hyperparameters to yield improved re­
aspect lies in its depthwise structure, allowing high performance while sults. These parameters are determined through empirical methods to
mitigating computational demands. The consideration of pointwise ensure a high level of accuracy. Notably, RF demonstrates superior
complexity involving a 1 × 1 convolution is pivotal, and ReLU activation performance when employed in conjunction with the original features,
is employed to maintain pointwise connections while generating deep achieving an accuracy score of 91%, while the ETC model attains a score
abstraction layers. of 89%. Among the tested models, the k-NN algorithm exhibits the
Furthermore, a resolution multiplier variable labeled as w is intro­ lowest performance, with an accuracy score of 81%. Refer to Table 3 for
duced, aimed at diminishing the dimensionality of both the input image a detailed presentation of the accuracy scores obtained by all classifiers
and the internal representation within each layer. This variable facili­ when applied using the original features.
tates adjustments in network dimensions while ensuring a consistent Table 4 outlines the classification accuracy of machine learning
value for “w” across all layers, thereby bolstering the overall efficiency models when applied with convoluted features. The results of the ex­
of the model. The feature vector map, marked as Fm in size, corresponds periments indicate that the RF and ETC models outperform other
with a filter size of Fs , while the input and output variables are sym­ models, achieving accuracy scores of 93.75% and 93.74%, respectively.
bolized by p and q respectively. The equation below offers a measure of Similarly, it is evident that both SVM and LR classifiers boast superior
the computational workload for the primary abstraction layers within accuracy compared to the other classifiers.
the architecture, where the variable ce represents the computational
effort. 4.3. Performance of ensemble models using original features

ce = Fs ⋅Fs ⋅w⋅αFm + w⋅ρ⋅αFm ⋅αFm (10) ChatGPT Initially, individual models are employed to analyze both
In the MobileNet framework, the multiplier variable w assumes the original features and convoluted features, with the outcomes pre­
values contextual to the study. When focusing on brain tumor classifi­ sented in Table 3 and Table 4. Among the nine machine learning models
cation experiments, this multiplier typically ranges between 1 and n. considered, it is observed that the RF, ETC, LR, and SVM models
Meanwhile, the resolution multiplier r remains constant at 1. The consistently yield the most favorable outcomes across both feature sets.
equation following calculates the computational workload, represented In this phase of the study, ensembles comprising these machine learning
as cost e .
Table 2
coste = Fs ⋅Fs ⋅w⋅ρ⋅Fm ⋅Fm (11) Experimental setup for the proposed system.
Libraries Scikitlearn, TensorFlow
3.8. Evaluation metrics Language Python 3.8
RAM 8GB
The evaluation of the machine learning models in this study relies on OS 64-bit window 10
metrics like accuracy, precision, recall, and F1 score, computed from the CPU Core i7, 7th Gen with 2.8 GHz processor
GPU Nvidia, 1060, 8GB
values in the confusion matrix. These matrices range between 0 and 1,

7
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

Table 3 Table 6
Accuracy of models with original features. Ensemble model results using convoluted Features set.
Model Accuracy with original features Model Accuracy Precision Recall F1 score

RF 91.78% RF + SVM [31] 99.99% 99.99% 99.99% 99.99%

ETC 89.47% RF + ETC [31] 97.21% 97.65% 98.47% 97.54%
LR 88.59% RF + LR [31] 95.62% 96.81% 97.14% 96.67%
SVM 88.47% ETC + SVM [31] 97.77% 97.45% 97.45% 97.45%
GNB 84.89% ETC + LR [31] 94.39% 95.27% 97.69% 96.44%
KNN 81.77% SVM + LR [31] 96.25% 97.34% 97.74% 97.54%
GBM 85.86%
DT 86.88%
SGD 84.47%
Table 7
Results for k-fold cross-validation of the proposed ensemble model.

Table 4 Fold Number Accuracy Precision Recall F1 score

Classifiers accuracy with convoluted features. Fold-1 99.23% 99.96% 99.94% 99.95%
Model Accuracy with convoluted features Fold-2 99.34% 99.96% 99.95% 99.96%
Fold-3 99.45% 99.97% 99.96% 99.96%
RF 93.75% Fold-4 99.11% 99.94% 100.0% 99.99%
ETC 93.74% Fold-5 99.24% 99.99% 99.98% 99.99%
LR 91.85% Average 99.27% 99.96% 99.96% 99.97%
SVM 92.34%
GNB 89.47%
KNN 86.53% validation results illustrate that the ensemble model proposed in this
GBM 87.84%
study achieves an average accuracy score of 0.996. Additionally, the
DT 90.37%
SGD 90.69% average precision, recall, and F1 scores for the model are 0.998, 0.998,
and 0.997, respectively.

models are evaluated using the original features. The findings from the
5. Discussion
ensemble learning models indicate that the RF + SVM ensemble model
exhibits superior performance in terms of accuracy, achieving a rate of
5.1. Deep learning model performance with original features
95%. This accuracy rate is almost 2% greater than that of the other
ensemble learning models. The subsequent model, SVM + ETC, attains a
The application of deep learning models involves utilizing both
classification accuracy of 92%. The RF + SVM model demonstrates a
original and convoluted features, and the results from these models are
precision of 95%, recall of 98%, and an F1 score of 96% in the context of
outlined in Table 8. Among the three deep learning models, it is evident
breast cancer diagnosis. Table 5 displays the outcomes of the ensemble
that CNN exhibits superior performance on both sets of features. In this
learning models on the original feature set.
phase of the experiments, the proposed model’s validity is assessed by
comparing it to state-of-the-art deep learning models. However, the
4.4. Performance of ensemble model on convoluted features ensemble model outperforms the deep learning models in terms of ac­
curacy. The accuracy of MLP decreases when using CNN features, while
The performance of the ensemble models is also evaluated using the LSTM accuracy improves. This improvement can be attributed to LSTM’s
features extracted by the customized CNN model. The results of these capacity to extract more meaningful features, facilitating sequence
experiments are presented in Table 6. The findings indicate that the RF generation. CNN’s accuracy remains nearly the same because it receives
+ SVM model outperforms alternative models, achieving an accuracy of the same convoluted features and an additional layer for predictions.
99% as well as precision, recall, and F1 scores of 99% each. The ETC +
LR model has demonstrated the least favorable outcomes, achieving an 5.1.1. Results comparison of transfer learning models
accuracy rate of 94%. The utilization of features from the CNN model in This section will undertake an analysis of the outcomes derived from
conjunction with ensemble learning models yields superior outcomes two transfer learning models, namely MobileNet and UNET, in relation
compared to the utilization of the original features alone. to the utilization of image data. Table 9 demonstrates the complete
classification report of the two models and clearly states that the UNET
4.5. Results of K-fold cross-validation model exhibits superior accuracy, precision, recall, and F1 score
compared to the MobileNet model.
K-fold cross-validation is implemented to evaluate the effectiveness
of the proposed model. The objective of cross-validation is to validate 5.1.2. Comparative evaluation of image and feature-based datasets
the results of the presented model and assess its robustness. This tech­ In order to assess the efficacy of the U-Net transfer learning model, a
nique is employed to evaluate the model’s performance across various comparative analysis was undertaken to analyze the performance of
subsets of the data. In this study, a 5-fold cross-validation approach is both the machine learning model and transfer learning models using
utilized, and the corresponding results are detailed in Table 7. The cross- numerical and image data. The findings indicate a significant
enhancement in the efficacy of the transfer learning models. The outputs
Table 5
Ensemble model results using Original Features set. Table 8
Model Accuracy Precision Recall F1 score Accuracy of deep learning models with original and convoluted features.
RF + SVM [31] 95.89% 95.91% 98.54% 96.99% Model Accuracy
RF + ETC [31] 93.34% 93.45% 95.11% 94.37%
Original features Convoluted features
RF + LR [31] 89.55% 90.65% 88.25% 89.17%
ETC + SVM [31] 94.14% 93.78% 95.64% 94.24% MLP 87.69% 84.41%
ETC + LR [31] 90.34% 91.45% 91.67% 91.55% CNN 90.22% 90.70%
SVM + LR [31] 91.73% 92.64% 96.98% 95.74% LSTM 85.95% 88.34%

8
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

Table 9 Table 11
Results comparison of transfer learning models. Performance comparison with state-of-the-art studies.
Model Accuracy Precision Recall F1 score Ref. Technique Accuracy

MobileNet 98.59% 98% 98% 98% [53] K-means clustering 92.01%

U-Net 99.97% 99% 99% 99% [47] PCA features with SVM 96.99%
[49] Quadratic SVM 98.11%
[48] Auto-encoder 98.40%
obtained by the machine learning models in both scenarios are pre­ [54] GF-TSK 94.11%
[50] XgBoost 97.11%
sented in Table 10, providing a comprehensive overview that enables a
[55] Five most significant features with LightGBM 95.03%
full study of their performance. [51] Chi-square features 98.21%
[52] LR with all features 98.10%
Proposed Deep convoluted features with voting classifier (RF + 99.99%
5.2. Performance comparison with existing studies
SVM)

To showcase the effectiveness of the proposed model in comparison

to prior state-of-the-art models, a comprehensive analysis is conducted in supervised learning scenarios with labeled data. Fortunately, the
by contrasting the outcomes with those obtained from existing models. SHAP technique, introduced as a recent breakthrough, provides a
To fulfill the objectives of this study, a meticulous selection process was quantitative remedy to this challenge, allowing us to enhance model
employed to identify the nine most relevant research articles. For interpretability. First proposed by Lee and Lundberg in 2017 and
instance, [47] utilized PCA features with the SVM machine learning expanded upon by Lundberg et al. in 2018 [58,59], this innovation has
model for breast cancer detection, achieving an accuracy score of notably advanced our understanding of the components within the
96.99%. In another study [48], autoencoder was employed, yielding the model.
highest accuracy score of 98.40%. The use of Quadratic SVM by [49] SHAP employs a linear additive feature attribution method rooted in
resulted in an accuracy score of 98.11%. For the same task, [50] applied concepts from cooperative game theory to clarify complex models. This
XgBoost and achieved an accuracy score of 97.11%. Similarly, [51,52] method assigns an importance value to each feature, taking into account
utilized Chi-square features and the LR machine learning model with its impact on the model’s predictions based on its presence or absence
accuracy scores of 98.21% and 98.10%, respectively. The performance during SHAP estimation. Such an explanatory approach simplifies the
comparison between the planned and existing studies is detailed in functioning of intricate models, rendering them more comprehensible.
Table 11. The results underscore the superior performance of the pro­ The principles underlying the linear additive feature attribution tech­
posed model. nique, drawing from cooperative game theory, are expounded upon in
works by Lee and Lundberg (2017) and further developed by Lundberg
5.3. Significance of the proposed model et al. (2020) [58,59].
j
∑
In order to validate the performance of the proposed model, we f (a) = g(a’ ) = ϕ0 + ϕj a’j (16)
tested it on two further independent datasets. The first dataset [56] is j=1

‘Breast Cancer Survival’, which contains 330 patient records with the In this scenario, we denote the original learning model as (a), while
feature Patient_ID, Age, Gender, and expression levels of four proteins the more straightforward explanatory model is referred to as g(a’ ). Here,
(Protein1, Protein2, Protein3, Protein4). The dataset also includes the a’j , where j represents a simplified seismic attribute number, represents
Breast cancer stage of the patient (Tumor_Stage), Histology (type of
these attributes. To calculate SHAP values, labeled as “j,” we consider all
cancer), ER, PR, and HER2 status, Surgery_type, Date of Surgery, Date of
potential input orderings. In the estimation process, a’j indicates the
Last Visit, and Patient Status (Alive/Dead). The second dataset [57]
presence or absence of a specific seismic attribute, using an input vector.
contains 10 Quantitative features to show the presence or absence of
Finally, ϕ0 represents the model prediction when none of the attributes
breast cancer in a patient. The features are Age (years), BMI (kg/m2),
are taken into account during the estimation. The comprehensive
Glucose (mg/dL), Insulin (μU/mL), HOMA, Leptin (ng/mL), Adiponectin
feature importance, calculated using SHAPly, is presented in descending
(μg/mL), Resistin (ng/mL), MCP-1(pg/dL), and Labels (absence or
order in Table 12.
presence). Fig. 4 presents the accuracy comparison of the classifiers used
The SHAP analysis underscores the significance of features in the
in experiments. The proposed model obtained 97.34% accuracy on the
prediction of breast cancer. While SHAP feature importance surpasses
first dataset and 96.67% accuracy on the dataset which greatly shows
traditional methods, relying solely on it provides somewhat constrained
the stability of the proposed model on all kinds of datasets. Important
additional insights. SHAPly XAI provides a more comprehensive and
terms are explained in Table 13.
information-laden portrayal of SHAP values, unveiling not just the
relative importance of features but also their intricate connections with
5.4. Shapley additive explanations the predicted outcome. SHAP explanations furnish valuable insights into
how features contribute to a specific instance. The cumulative sum of
Deciphering the connections between inputs and outputs in learning feature contributions, in conjunction with the bias term, equals the
models is often intricate due to the perceived complexity of these models model’s initial prediction, representing the forecast before applying the
as mysterious, black-box algorithms. This lack of transparency poses a inverse link function.
significant challenge in grasping the significance of features, especially
5.5. Limitations
Table 10
Comparison analysis of image and feature-based dataset techniques. The present study contributes significantly by proposing a machine
Features Based Data Image Data learning-based approach leveraging XAI for breast cancer detection,
aiming for high accuracy. However, certain limitations are evident:
Best ML Performing Models’ Results Best TL Models’ Results

Models Original Features CNN Features Models Accuracy • The study’s effectiveness might be influenced by the size of the
VC (RF + SVM) 95.8% 99.9% U-Net 99.99% dataset used for training and testing the models. A larger and more
VC (ETC + SVM) 94.1% 97.7% MobileNet 98.5%

9
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

Fig. 4. Accuracy comparison of the classifiers.

Table 12 Table 13
Feature importance calculated using SHAPly. Acronym Table.
Weight Feature Abbreviation Description

0.1756 ± 0.0468 Texture SE AB AdaBoost

0.0271 ± 0.0482 Concave Points SE BiLSTM Bidirectional LSTM
0.0464 ± 0.0474 Concave Points Mean CAD computer-aided diagnostics
0.0027 ± 0.0077 Smoothness Mean CNN convolutional neural network
0.0046 ± 0.0141 Area Worst DNA Deoxyribonucleic acid
0.0037 ± 0.0244 Symmetry Worst DCIS Ductal Carcinoma in situ
0.0910 ± 0.0849 Texture Mean DNN deep neural network
0.0048 ± 0.0426 Radius Mean DT Decision Tree
0.0059 ± 0.0074 Compactness SE ETC extra tree classifier
0.0011 ± 0.0241 Texture Worst GNB Gaussian Naive Bayes
0.0011 ± 0.0011 Perimeter SE GBM Gradient boosting machine
0.0011 ± 0.0221 Concave Points Worst GRU gated recurrent unit
− 0.0038 ± 0.0067 Area Mean IDC Invasive ductal carcinoma
− 0.0048 ± 0.0077 Fractal Dimension SE KNN K-nearest neighbor
− 0.0069 ± 0.0245 Perimeter Worst LDA Latent Dirichlet Allocation
− 0.0049 ± 0.0088 Compactness Worst LBC Lobular breast cancer
− 0.0029 ± 0.0054 Fractal Dimension Mean LSTM long short-term memory
− 0.0031 ± 0.0048 Compactness Mean LR Logistic regression
− 0.0068 ± 0.0046 Perimeter Mean MBC Mucinous breast cancer
− 0.0048 ± 0.0345 Radius Worst MRI magnetic resonance imaging
0.0011 ± 0.0261 Concavity SE MTBC Mixed tumor Breast cancer
− 0.0038 ± 0.0077 Fractal Dimension Worst RF Randon Forest
0.0011 ± 0.0011 Smoothness SE RNA Ribonucleic acid
− 0.0038 ± 0.0057 Concavity Mean RNN Recurrent Neural Network
− 0.0039 ± 0.0044 Smoothness Worst SGD stochastic gradient descent
− 0.0039 ± 0.0275 Perimeter Worst SVM Support Vector Machine
− 0.0049 ± 0.0058 Symmetry Mean XAI explainable artificial intelligence
− 0.0048 ± 0.0365 Area SE
− 0.0051 ± 0.0078 Symmetry SE
− 0.0068 ± 0.0056 Concavity Worst Despite these limitations, the study’s contributions in proposing a
− 0.0068 ± 0.0056 Radius SE
comprehensive framework for accurate breast cancer detection using
machine learning techniques, including the ensemble model, and
diverse dataset could offer a more comprehensive understanding of comparing their performance against various benchmarks are valuable
the model’s performance. steps toward advancing diagnostic methodologies.
• The models’ robustness and generalizability could benefit from
external validation on datasets from different sources or de­ 6. Conclusions
mographics, ensuring their applicability beyond the current dataset.
• The computational resources required for implementing some of the This research paper presents a novel framework for the precise
proposed models. identification of BC, utilising a combination of image and numerical
data features with explainable Artificial Intelligence (XAI). The

10
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

utilisation of the U-NET transfer learning model is employed for image- References
based prediction. Additionally, an ensemble model is constructed by
integrating characteristics from a customised convolutional neural [1] WHO. World Health Organization, Cancer: Key Facts. 2022. https://www.who.
int/news-room/fact-sheets/detail/cancer, July 2022. Online; accessed 10 January
network (CNN) model with an ensemble comprising random forest (RF) 2023.
and support vector machine (SVM). The experiments aim to evaluate the [2] Yi-Sheng Sun, Zhao Zhao, Zhang-Nv Yang, Xu Fang, Lu Hang-Jing, Zhi-Yong Zhu,
influence of original features compared to convoluted features. A Wen Shi, Jianmin Jiang, Ping-Ping Yao, Han-Ping Zhu, Risk factors and
preventions of breast cancer, Int. J. Biol. Sci. 13 (11) (2017) 1387.
comparative analysis is carried out to assess the efficacy of various [3] WHO. World Health Organization, Breast Cancer. https://www.who.int/ne
classifiers in accurately detecting BC, utilising the Wisconsin dataset. ws-room/fact-sheets/detail/breastcancer, July 2022. Online; accessed 10 January
The model under consideration exhibits promising capabilities in 2023.
[4] WHO. World Health Organization, Breast Cancer. https://www.who.int/new
enhancing BC diagnosis, with a remarkable accuracy rate of 99.99\%. s-room/fact-sheets/detail/breast-cancer, March 2021. Online; accessed 10 January
The present study contributes to the advancement of BC diagnosis by 2023.
introducing a novel strategy based on machine learning and discussing [5] Freddie Bray, Jacques Ferlay, Isabelle Soerjomataram, Rebecca L. Siegel, Lindsey
A. Torre, Ahmedin Jemal, et al., Globocan estimates of incidence and mortality
the interpretation of the variables using XAI. The primary objective of
worldwide for 36 cancers in 185 countries, CA Cancer J. Clin. 68 (6) (2018)
this approach is to get a notable level of precision, hence facilitating the 394–424.
early and reliable identification of BC. Ultimately, the implementation [6] Amrita Ray Chaudhury, Ranjani Iyer, Kaveri K. Iychettira, A. Sreedevi, Diagnosis of
of this approach is expected to enhance patient outcomes.This study invasive ductal carcinoma using image processing techniques, in: In 2011
International Conference on Image Information Processing, IEEE, 2011, pp. 1–6.
aims to establish a comprehensive framework for effectively identifying [7] Fredika M. Robertson, Melissa Bondy, Wei Yang, Hideko Yamauchi,
BC using explainable Artificial Intelligence (XAI) coupled with image Shannon Wiggins, Samira Kamrudin, Savitri Krishnamurthy, Huong Le-Petross,
and numerical data features, thereby mitigating the associated mortality Luc Bidaut, Audrey N. Player, et al., Inflammatory breast cancer: the disease, the
biology, the treatment, CA Cancer J. Clin. 60 (6) (2010) 351–375.
risk linked to this prevalent reason of death among women. We [8] Serena Masciari, Nina Larsson, Janine Senz, Niki Boyd, Pardeep Kaurah, Michaela
employed U-NET transfer learning model for image-based prediction. In J. Kandel, Lyndsay N. Harris, Hugo C. Pinheiro, Armelle Troussard,
addition, We proposed an ensemble model, leveraging its reported su­ Penelope Miron, et al., Germline e-cadherin mutations in familial lobular breast
cancer, J. Med. Genet. 44 (11) (2007) 726–731.
perior performance in the existing literature. Furthermore, our training [9] American Cancer Society. American Cancer Society, Breast Cancer. https://www.
process employs features derived from the CNN model. The accuracy of cancer.org/content/dam/cancer-org/research/cancer-facts-and-statistics/breast-ca
the suggested model in distinguishing patients from normal ones is an ncer-facts-and-figures/breast-cancer-facts-and-figures-2015-2016.pdf, July 2022.
Online; accessed 10 January 2023.
impressive 0.999. Moreover, the use of CNN-based features in [10] A. Pasumpon, Pandian., Identification and classification of cancer cells using
conjunction with models consistently leads to more optimal outcomes. capsule network with pathological images, J. Artif. Intell. 1 (01) (2019) 37–44.
We demonstrate the efficiency and durability of the provided model by [11] Andrei Chekkoury, Parmeshwar Khurd, Jie Ni, Claus Bahlmann, Ali Kamen,
Amar Patel, Leo Grady, Maneesh Singh, Martin Groher, Nassir Navab, et al.,
doing K-fold cross-validation and comparing its performance to estab­
Automated malignancy detection in breast histopathological images, in: Medical
lished state-of-the-art models. In future studies, we plan to apply our Imaging 2012: Computer-Aided Diagnosis 8315, SPIE, 2012, pp. 332–344.
model to a variety of datasets spanning different domains, including BC [12] Furqan Rustam, Abid Ishaq, Kashif Munir, Mubarak Almutairi, Naila Aslam,
pictures, and analyse the corresponding numerical values produced Imran Ashraf, Incorporating cnn features for optimizing performance of ensemble
classifier for cardiovascular disease prediction, Diagnostics 12 (6) (2022) 1474.
from the microscopic characteristics in these images. [13] Vaibhav Rupapara, Furqan Rustam, Abid Ishaq, Ernesto Lee, Imran Ashraf, Chi-
square and pca based feature selection for diabetes detection with ensemble
CRediT authorship contribution statement classifier, Intell. Automat. Soft Comput. 36 (2) (2023).
[14] Muhammad Idrees, Ayesha Sohail, Explainable machine learning of the breast
cancer staging for designing smart biomarker sensors, Sensors Int. 3 (2022)
Raafat M. Munshi: Data curation, Resources, Writing – original 100202.
draft. Lucia Cascone: Project administration, Writing – review & edit­ [15] M. Rodriguez-Sampaio, Mariano Rincón, Sonia Valladares-Rodríguez, M. Bachiller-
Mayoral, Explainable artificial intelligence to detect breast cancer: A qualitative
ing. Nazik Alturki: Conceptualization, Methodology, Funding, Project case-based visual interpretability approach, in: International Work-Conference on
administration, Writing – review & editing. Oumaima Saidani: the Interplay between Natural and Artificial Computation, Springer, 2022,
Conceptualization, Methodology, Writing – original draft. Amal pp. 557–566.
[16] Fabián Silva-Aravena, Hugo Núñez Delafuente, Jimmy H. Gutiérrez-Bahamondes,
Alshardan: Conceptualization, Methodology, Software, Visualization. Jenny Morales, A hybrid algorithm of ml and xai to prevent breast cancer: a
Muhammad Umer: Methodology, Software, Writing – original draft. strategy to support decision making, Cancers 15 (9) (2023) 2443.
[17] Samir S. Yadav, Shivajirao M. Jadhav, Thermal infrared imaging based breast
cancer diagnosis using machine learning techniques, Multimed. Tools Appl. (2022)
Declaration of competing interest
1–19.
[18] Habib Dhahri, Eslam Al Maghayreh, Awais Mahmood, Wail Elkilani, Mohammed
The authors declare no conflict of interest. Faisal Nagi, et al., Automated breast cancer diagnosis based on machine learning
algorithms, J. Healthcare Eng. 2019 (2019).
[19] Z. Wang, M. Li, H. Wang, H. Jiang, Y. Yao, H. Zhang, J. Xin, Breast cancer detection
Data availability using extreme learning machine based on feature fusion with CNN deep features,
IEEE Access 7 (2019) 105146–105158.
The datasets generated during and/or analyzed during the current [20] Saad Awadh Alanazi, M.M. Kamruzzaman, Md Nazirul Islam Sarker,
Madallah Alruwaili, Yousef Alhwaiti, Nasser Alshammari, Muhammad
study can be downloaded from the following link. https://github. Hameed Siddiqi, Boosting breast cancer detection using convolutional neural
com/MUmerSabir/MDPIDiagnostic network, J. Healthcare Eng. 2021 (2021).
[21] Muhammad Umer, Mahum Naveed, Fadwa Alrowais, Abid Ishaq, Abdullah Al
Hejaili, Shtwai Alsubai, Ala’Abdulmajid Eshmawi, Abdullah Mohamed, and Imran
Acknowledgments Ashraf., Breast cancer detection using convoluted features and ensemble machine
learning algorithm, Cancers 14 (23) (2022) 6015.
This research work was funded by Institutional Fund Projects under [22] Muhammet Fatih Ak, A comparative analysis of breast cancer detection and
diagnosis using data visualization and machine learning applications, Healthcare 8
grant no. (IFPIP: 1328-415-1443) and by Princess Nourah bint Abdul (2) (2020) 111.
rahman University Researchers Supporting Project number [23] Yong Joon Suh, Jaewon Jung, Bum-Joo Cho, Automated breast cancer detection in
(PNURSP2024R333), Princess Nourah bint Abdulrahman University, digital mammograms of various densities via deep learning, J. Personal. Med. 10
(4) (2020) 211.
Riyadh, Saudi Arabia. The authors gratefully acknowledge the technical
[24] Jing Zheng, Denan Lin, Zhongjun Gao, Shuang Wang, Mingjie He, Jipeng Fan, Deep
and financial support provided by the Ministry of Education and King learning assisted efficient adaboost algorithm for breast cancer detection and early
Abdulaziz University, DSR. Jeddah, Saudi Arabia. diagnosis, IEEE Access 8 (2020) 96946–96954.
[25] Hanan Aljuaid, Nazik Alturki, Najah Alsubaie, Lucia Cavallaro, Antonio Liotta,
Computer-aided diagnosis for breast cancer classification using deep neural
networks and transfer learning, Comput. Methods Prog. Biomed. 223 (2022)
106951.

11
R.M. Munshi et al. Image and Vision Computing 142 (2024) 104910

[26] Manav Mangukiya, Anuj Vaghani, Meet Savani, Breast cancer detection with citation analysis using textual features and smote resampling techniques, Pattern
machine learning, Int. J. Res. Appl. Sci. Eng. Technol. 10 (2) (2022) 141–145. Recogn. Lett. 150 (2021) 250–257.
[27] Xiaomei Wang, Ijaz Ahmad, Danish Javeed, Syeda Armana Zaidi, Fahad [43] Lucia Cascone, Saima Sadiq, Saleem Ullah, Seyedali Mirjalili, Hafeez Ur
M. Alotaibi, Mohamed E. Ghoneim, Yousef Ibrahim Daradkeh, Junaid Asghar, Rehman Siddiqui, Muhammad Umer, Predicting household electric power
Elsayed Tag Eldin, Intelligent hybrid deep learning model for breast cancer consumption using multi-step time series with convolutional lstm, Big Data Res. 31
detection, Electronics 11 (17) (2022) 2767. (2023) 100360.
[28] Liang Tang, Jianming Li, Ju Meihua Bao, Yiwei Chen Xiang, Yan Wang, Genetic [44] Ahmad Hameed, Muhammad Umer, Umair Hafeez, Hassan Mustafa,
association between her2 and esr2 polymorphisms and ovarian cancer: a meta- Ahmed Sohaib, Muhammad Abubakar Siddique, Hamza Ahmad Madni, Skin lesion
analysis, Onco Targets Ther (2018) 1055–1066. classification in dermoscopic images using stacked convolutional neural network,
[29] Xiaoliang Xie, Xulin Wang, Yuebin Liang, Jingya Yang, Wu Yan, Li Li, Xin Sun, J. Ambient. Intell. Humaniz. Comput. (2021) 1–15.
Pingping Bing, Binsheng He, Geng Tian, et al., Evaluating cancer-related [45] Yuewu Hou, Zhaoying Liu, Ting Zhang, Yujian Li, C-unet: complement unet for
biomarkers based on pathological images: a systematic review, Front. Oncol. 11 remote sensing road extraction, Sensors 21 (6) (2021) 2153.
(2021) 763527. [46] Parvathaneni Naga Srinivasu, Jalluri Gnana SivaSai, Muhammad Fazal Ijaz, Akash
[30] Sihe Chen, Yongheng Chen, Yu Lanting, Hu. Xiangmei, Overexpression of socs4 Kumar Bhoi, Wonjoon Kim, James Jin Kang, Classification of skin disease using
inhibits proliferation and migration of cervical cancer cells by regulating jak1/ deep learning neural networks with mobilenet v2 and lstm, Sensors 21 (8) (2021)
stat3 signaling pathway, Eur. J. Gynaecol. Oncol. 42 (3) (2021) 554–560. 2852.
[31] H. Karamti, R. Alharthi, M. Umer, H. Shaiba, A. Ishaq, N. Abuzinadah, S. Alsubai, [47] D. Lavanya, Dr K. Usha, Rani., Analysis of feature selection with classification:
I. Ashraf, Breast cancer detection employing stacked ensemble model with breast cancer datasets, Indian J. Comput. Sci. Eng. (IJCSE) 2 (5) (2011) 756–763.
convolutional features, Cancer Biomarkers (2023) 1–16 (Preprint). [48] Sonam Jawahar Singh, Ramanathan Rajaraman, Tanmay Tulsidas Verlekar, Breast
[32] UCI Repository. Uci machine learning repository. URL https://archive.ics.uci.edu/ cancer prediction using auto-encoders, in: International Conference on Data
ml/datasets/Breast+Cancer+Wisconsin+%28Diagnostic%29. Management, Analytics & Innovation, Springer, 2023, pp. 121–132.
[33] Leo Breiman, Bagging predictors, Mach. Learn. 24 (2) (1996) 123–140. [49] O. Ibrahim Obaid, Mazin Abed Mohammed, M.K.A. Ghani, A. Mostafa, Fahad Taha,
[34] Gérard Biau, Erwan Scornet, A random forest guided tour, Test 25 (2) (2016) et al., Evaluating the performance of machine learning techniques in the
197–227. classification of Wisconsin breast cancer, Int. J. Eng. Technol. 7 (4.36) (2018)
[35] Mubariz Manzoor, Muhammad Umer, Saima Sadiq, Abid Ishaq, Saleem Ullah, 160–166.
Hamza Ahmad Madni, Carmen Bisogni, Rfcnn: traffic accident severity prediction [50] Prithwish Ghosh. Breast cancer wisconsin (diagnostic) prediction.
based on decision level fusion of machine and deep learning model, IEEE Access 9 [51] Ravi Kumar Sachdeva, Priyanka Bathla, A machine learning-based framework for
(2021) 128359–128371. diagnosis of breast cancer, International Journal of Software Innovation (IJSI) 10
[36] Afaq Juna, Muhammad Umer, Saima Sadiq, Hanen Karamti, (1) (2022) 1–11.
Ala’Abdulmajid Eshmawi, Abdullah Mohamed, Imran Ashraf, Water quality [52] Muhammet Fatih Ak, A comparative analysis of breast cancer detection and
prediction using knn imputer and multilayer perceptron, Water 14 (17) (2022) diagnosis using data visualization and machine learning applications, Healthcare 8
2592. (2) (2020), https://doi.org/10.3390/healthcare8020111. ISSN 2227-9032. URL,
[37] Elham Besharati, Marjan Naderan, Ehsan Namjoo, Lr-hids: logistic regression host- https://www.mdpi.com/2227-9032/8/2/111.
based intrusion detection system for cloud environments, J. Ambient. Intell. [53] Ashutosh Kumar Dubey, Umesh Gupta, Sonal Jain, Analysis of k-means clustering
Humaniz. Comput. 10 (9) (2019) 3669–3692. approach on the breast cancer Wisconsin dataset, Int. J. Comput. Assist. Radiol.
[38] Samina Sarwat, Naeem Ullah, Saima Sadiq, Robina Saleem, Muhammad Umer, Surg. 11 (11) (2016) 2033–2047.
Ala’Abdulmajid Eshmawi, Abdullah Mohamed, Imran Ashraf, Predicting students’ [54] Allison Murphy, Breast cancer Wisconsin (diagnostic) data analysis using gfs-tsk,
academic performance with conditional generative adversarial network and deep in: North American Fuzzy Information Processing Society Annual Conference,
svm, Sensors 22 (13) (2022) 4834. Springer, 2021, pp. 302–308.
[39] Imran Ashraf, Manideep Narra, Muhammad Umer, Rizwan Majeed, Saima Sadiq, [55] Sami Akbulut, Ipek Balikci Cicek, Cemil Colak, Classification of breast cancer on
Fawad Javaid, Nouman Rasool, A deep learning-based smart framework for cyber- the strength of potential risk factors with boosting models: a public health
physical and satellite system security threats detection, Electronics 11 (4) (2022) informatics application, Med. Bull. Haseki/Haseki Tip Bulteni 60 (3) (2022).
667. [56] Kreesh Rajani. Breast cancer survival dataset. URL https://www.kaggle.com/datas
[40] Muhammad Umer, Saima Sadiq, Michele Nappi, Muhammad Usman Sana, ets/kreeshrajani/breast-cancer-survival-dataset/code.
Imran Ashraf, et al., Etcnn: extra tree and convolutional neural network-based [57] Ankit Khushal Barai. Uci machine learning repository. URL https://www.kaggle.
ensemble model for covid-19 tweets sentiment classification, Pattern Recogn. Lett. com/datasets/ankitbarai507/breast-cancer-dataset/code.
164 (2022) 224–231. [58] Muhammad Aurangzeb Ahmad, Carly Eckert, Ankur Teredesai, Interpretable
[41] Rizwan Majeed, Nurul Azma Abdullah, Muhammad Faheem Mushtaq, machine learning in healthcare, in: Proceedings of the 2018 ACM International
Muhammad Umer, Michele Nappi, Intelligent cyber-security system for iot-aided Conference on Bioinformatics, Computational Biology, and health INFORMATICS,
drones using voting classifier, Electronics 10 (23) (2021) 2926. 2018, pp. 559–560.
[42] Muhammad Umer, Saima Sadiq, Malik Muhammad Saad Missen, Zahid Hameed, [59] Scott M. Lundberg, Su-In Lee, A unified approach to interpreting model
Zahid Aslam, Muhammad Abubakar Siddique, Michele Nappi, Scientific papers predictions, Adv. Neural Inf. Proces. Syst. 30 (2017).

12