Similarity Report ID: oid:16158:62619656

PAPER NAME

malaria detection using deep-learning sh

akib.pdf

WORD COUNT CHARACTER COUNT

4191 Words 24610 Characters

PAGE COUNT FILE SIZE

11 Pages 621.5KB

SUBMISSION DATE REPORT DATE

Jul 10, 2024 11:55 AM GMT+5:30 Jul 10, 2024 11:56 AM GMT+5:30

17% Overall Similarity

The combined total of all matches, including overlapping sources, for each database.
17% Internet database 1% Publications database

Excluded from Similarity Report

Crossref database Crossref Posted Content database
Submitted Works database Bibliographic material
Quoted material Cited material
Small Matches (Less then 10 words)

Summary
 1. Abstract
Malaria remains a significant health challenge globally, especially in regions like Africa and Asia where
resources are limited. The high incidence rate, coupled with the scarcity of advanced diagnostic tools and
skilled healthcare professionals, calls for innovative solutions. This research introduces machine learning
techniques as a supplementary diagnostic tool to identify malaria from blood smear images. We explored
six distinct features extracted using advanced neural network architectures: VGG16, VGG19, ResNet50,
ResNet101, DenseNet121, and DenseNet201. These features were then employed to train four different
1
classifiers: Decision Tree, Support Vector Machine, Naïve Bayes, and K-Nearest Neighbour. A
1
comprehensive evaluation of their performance was conducted, focusing on metrics such as precision,
recall, F1-score, overall accuracy, and processing time. The findings indicate that the automated system is
capable of detecting malaria parasites in blood specimens with an impressive accuracy rate exceeding 94%,
showcasing a reduction in complexity compared to previous methods documented in scientific literature.

2. Introduction The predominant method for malaria

diagnosis is light microscopy of blood films.
This technique involves placing a drop of the
Malaria, transmitted by the bite of female
patient’s blood on a microscope slide, which
Anopheles mosquitoes, is a significant global
is then stained to highlight the presence of
health concern. The disease is caused by
malaria parasites for visual inspection.
1
various mosquito-borne parasites, including
Plasmodium Ovale, Plasmodium malariae,
In the realm of malaria diagnosis, both thick
Plasmodium vivax, and Plasmodium
and thin blood smears are utilized. Thick
falciparum. According to the 2020 world
smears enhance the detection efficiency of
malaria report, these parasites contributed to
parasites, while thin smears offer the
an estimated 229 million malaria cases and
advantage of easier species identification and
approximately 409,000 deaths worldwide.
parasite stage recognition. Light microscopy,
The most affected demographic is children
the traditional diagnostic method, is cost-
under five, who represent 67% of malaria-
effective and accessible due to its minimal
related deaths globally. The African
equipment needs, relying instead on the
continent bears the brunt of the disease, with
94% of cases and half of the global fatalities expertise of the examiner. However, this
3 method’s drawbacks are significant,
occurring in just six countries: Nigeria
including the extensive training required for
(23%), the Democratic Republic of Congo
personnel to accurately interpret slides,
(11%), Tanzania (5%), Mozambique (4%),
which is both costly and labor-intensive.
Niger (4%), and Burkina Faso (4%). Malaria
also significantly impacts regions in South-
East Asia and the Eastern Mediterranean. Addressing these challenges, our study
introduces an innovative automated approach
for identifying malaria parasites in blood
Despite substantial investments exceeding
smear images. We have developed a set of
USD 3 billion in 2019 for malaria control and
automated procedures leveraging transfer
elimination efforts—with malaria-endemic
learning techniques to pinpoint infections in
countries contributing USD 900 million—the
blood samples. Additionally, we employed a
disease continues to pose a severe threat to 1
public health. suite of deep-learning models for feature
extraction and evaluated four machine-
learning classifiers in detail. Our study also
delves into the computational complexity
involved in each method.
5
The structure of the paper is as follows:
Section 2 presents a review of related
literature; Section 3 details the methodology
employed; Section 4 discusses the
Figure 1 malaria death rates
experimental results; and Section 5 concludes
the findings. Untreated malaria can lead to severe
complications, such as respiratory distress,
The Covid-19 pandemic has significantly pneumonia, anemia, and renal failure.
disrupted malaria control efforts across
Pregnant women are particularly vulnerable,
various regions. The diversion of already
with the infection potentially resulting in
limited financial resources to manage the
pandemic has resulted in underfunded stillbirths, neonatal deaths, and miscarriages.
healthcare services. Additionally, the The brain and lungs are among the organs
pandemic has intensified the demand for most severely affected by malaria.
healthcare workers, exacerbating the existing Consequently, malaria is considered a
shortage. The limited availability of personal medical emergency that requires immediate
protective equipment (PPE) and testing detection. Delays in diagnosis can be fatal,
facilities further increases the risk of making the early identification of malaria in
exposure and infection among healthcare patients imperative.
providers. Isolation, quarantine, and in some
cases, the death of healthcare workers further
aggravate the shortage.
In this study, we harness the power of
machine learning and deep learning to detect
Moreover, Covid-19 containment strategies,
such as lockdowns, have implications beyond malaria in its early stages. By utilizing these
the immediate health response. They disrupt advanced technologies, we aim to improve
global and local supply chains, affecting the the speed and accuracy of malaria diagnosis,
production, procurement, shipping, and potentially saving lives by enabling timely
distribution of commodities, which in turn treatment interventions.
impacts malaria control. The restrictions on
movement and the fear of contracting Covid-
19 may deter patients from seeking timely 3. Literature Review
medical attention for malaria, leading to
delayed treatment and increased risk of Deep learning, a sophisticated extension of
severe illness and death. Recent studies neural networks, has revolutionized
suggest that malaria mortality rates could numerous fields with its ability to
double if the focus remains solely on Covid- autonomously learn intricate data features.
19, highlighting the need for a balanced Unlike human cognition, deep-learning
approach to health crises. models necessitate extensive, high-quality,
annotated datasets to effectively predict
future events. This requirement has
historically hindered the medical sector’s
early adoption of deep learning due to
challenges in acquiring such datasets and two convolutional layers comprised
concerns over privacy. However, the concept 128 and 256 filters, respectively. The
of transfer learning has enabled the 1
CNN’s architecture concluded with
application of deep learning in medicine. three fully connected layers, each
Transfer learning utilizes pre-trained with 256 neurons. The model was
models—deep-learning systems already trained on 90% of the images and
versed in solving analogous problems—to tested on the remaining 10%,
adapt to new, yet similar, tasks. This achieving an accuracy of1 97.37%.
approach is one of three primary methods for 2. Transfer Learning: The second
deploying deep learning, alongside building approach utilized a pre-trained CNN
models from the ground up and refining model, specifically AlexNet, for
existing models. feature extraction, coupled with an
1
SVM (Support Vector Machine) for
Pre-trained models like VGG16, VGG19, classification. This method resulted in
ResNet50, ResNet101, ResNet152, an accuracy of 91.99%.
DenseNet121, DenseNet201, AlexNet, and
Xception have been instrumental beyond the Both approaches demonstrate the potential of
medical domain, aiding
2
in tasks ranging from deep learning in medical diagnostics,
facial recognition to skin burn classification offering efficient and accurate alternatives to
and cancer detection, as well as identifying traditional methods.
financial fraud. Recently, this methodology
has been applied to distinguish blood-smear In a recent study by Huq, A., et al., a dataset
images containing the Plasmodium parasite, of 27,558 images was utilized, comprising an
demonstrating its versatility and potential in equal number of normal and malaria-infected
healthcare diagnostics. images sourced from the National Institutes
of Health. Despite the original variability in
The researchers explored deep-learning image sizes, all were standardized to 224 ×
methodologies to enhance the diagnosis of 224 pixels to align with the VGG16 3pre-
malaria. The primary goal was to trained model’s input requirements. The
differentiate between malaria-positive and dataset was divided into three parts: 70% for
negative blood smear samples. They training, 30% for validation, and 10%
employed 27,578 RGB images, balanced extracted from the training set for testing.
equally between infected and uninfected The training set was then reduced to 60% of
samples, resized to 44×44 pixels. Two the original data.
distinct strategies were implemented:
The research included two distinct
1. Training from Scratch: A 16-layer experiments:
Convolutional Neural Network
(CNN) was designed, featuring six 1. Standard Training: Utilized
convolutional layers and three fully unaltered original images for training,
connected layers. The initial two with adversarial images introduced
convolutional layers contained 32 during the testing phase. The model
filters each, followed by a max- demonstrated a 95.86% accuracy
1
pooling layer. The subsequent two with standard images but was
layers had 64 filters each, followed significantly less effective against
by an average pooling layer. The final
 adversarial images, with an accuracy In a subsequent study, Rajaraman et al.
1
of only 29.40%. explored ensemble1
learning, combining
2. Adversarial Training: Integrated features from three pre-trained models—
adversarial images into the training VGG19, SqueezeNet, and InceptionResNet-
process, which enabled the algorithm V2—with a custom-trained CNN model for
to better recognize and classify these final classification. This approach aimed to
images during testing, achieving an distinguish between malaria-infected
1
and
accuracy of 93.30%. This method healthy blood images. While the ensemble
also maintained a high accuracy of method achieved an accuracy of 99.11%
95.80% with standard images. with a sensitivity of 98.94%, the VGG19
model independently reached a higher
These findings underscore the importance of accuracy of 99.32% and a sensitivity of
including adversarial examples in the training 99.31%, surpassing the performance of both
phase to build more robust deep-learning the ensemble and other pre-trained models.
models capable of withstanding such
perturbations. These studies underscore the effectiveness of
both individual and combined deep learning
The research by Rajaraman et al. provided models in accurately diagnosing malaria
significant insights into the application of from blood-smear images, showcasing the
deep learning for malaria diagnosis. They potential of AI in medical diagnostics.
analyzed 27,580 blood-smear images,
equally divided between infected and 4. Motivation
uninfected samples. The study employed two
main approaches: Our research encompasses a comprehensive
2
examination of feature representation and
Deep Learning from Scratch: A custom extraction, the practicality of various deep
Convolutional Neural Network 2
(CNN) learning approaches, and the classification of
model was developed with three convolution malaria through image analysis. We aim to
layers, each followed by a max-pooling layer, benchmark datasets and evaluate the
two fully connected (FC) layers, and ReLU applicability of these methods in real-world
activation.
1
The model’s input dimension was scenarios, focusing on activity recognition
set to 100 × 100 × 3, and all convolution and anomaly detection outcomes, as well
layers utilized a 3 x 3 filter size. This CNN establishing assessment criteria for image-
achieved a classification accuracy of 92.7%. based studies.
Transfer Learning Approach: The study The motivation behind our study is multi-
leveraged pre-trained models known for their faceted. Firstly, we emphasize the
strong feature extraction capabilities, distinction between traditional methods
including AlexNet, VGG16, ResNet50, that rely on manual feature extraction and
Xception,
1
and DenseNet121. Specific layers modern deep learning techniques,
like fc6, block5_conv2, res5c_branch2c, showcasing the advancements in the
block14_sepconv1, and conv_16_x2 were latter for both activity recognition and
used for their discriminative features. Among malaria diagnosis in images. Secondly,
these, ResNet50 stood out with the highest we address challenges encountered
recognition accuracy of 95.91% and a during image processing. Furthermore,
sensitivity of 94.74%. we conducted a comparative analysis of
 different machine learning and deep predictions, ML models are increasingly
learning methodologies to determine the valued for their interpretative power,
most effective approach for our offering insights into the underlying
objectives. relationships within datasets. This aspect
of ML is known as interpretability.
5. Key concept overview
ML empowers computers to
2 autonomously perform tasks by learning
Indeed, Machine Learning (ML), Deep
Learning (DL), and Convolutional Neural from data, rather than following explicit
Networks (CNNs) are terms that have instructions. It’s a data analysis method
become prevalent in scientific discourse. that involves building and training
6
Machine learning is a subset of artificial models, enabling computers to learn
intelligence (AI) focused on developing through experience and predict outcomes.
algorithms that enable computers to learn In the context of malaria diagnosis, ML
from and make predictions or decisions based analyzes input images, extracting salient
on data. It encompasses a variety of tasks, features to identify the presence of
including: malaria. These features are pivotal in
determining whether the image
 Regression: Predicting continuous represents a normal or infected case,
values based on input data. showcasing the practical applications of
 Classification: Categorizing data ML in healthcare.
into predefined classes.
 Clustering: Grouping similar data Reinforcement learning has indeed seen
points together without predefined
classes.
7
Deep learning, a subset of machine learning,
involves neural networks with many layers 2
(hence “deep”) that can learn complex Figure 2 types of machine learning

patterns in large amounts of data. CNNs are a

type of deep learning model particularly remarkable advancements in the new
effective for tasks involving images, such as millennium, achieving performance levels in
image recognition and classification, due to various domains that are comparable to
11
their ability to automatically and adaptively human intelligence. While it’s a driving force
learn spatial hierarchies of features from in social robotics, its application in the
input images. These technologies are at the medical field is less prevalent. Medical
forefront of AI research and application, decision-making often involves a sequential
driving innovation across various scientific process, requiring the interpretation of test
and industrial fields. results and a thorough diagnostic session to
understand a patient’s illness. However,
many machine learning models in healthcare
6. Machine Learning tend to overlook the dynamic nature of
disease progression, relying instead on static
Machine learning (ML) has indeed data. Consequently, physicians’
garnered significant attention for its understanding of diseases’ conditional
predictive prowess across a myriad of
complex scenarios. Beyond mere
progression is largely based on personal the images were standardized to a resolution
experience. of 224×224×3

In the realm of machine learning,

unsupervised and supervised learning are
two fundamental approaches. Unsupervised
learning algorithms infer patterns from
datasets without labeled responses, meaning
they operate without knowing the expected
outcome. In contrast, supervised learning
aims to identify the relationship between
input features (independent variables) and the
outcome (dependent variable), using labeled
data to train the model. Each type has its own
set of applications and is chosen based on the
nature of the problem and the available data.
Figure 4 sample blood smear image

, which is compatible with 13the input

requirements of several pre-trained
Convolutional Neural Network (CNN)
models, namely VGG16, VGG19, ResNet50,
and ResNet101. The same normalization was
applied for the DenseNet121 and
DenseNet201 models. These models were
initially trained on the ImageNet database,
which is renowned for its extensive
collection of images spanning 1000 different
Figure 3 popular machine kearning techniques object classes, providing a robust foundation
for the recognition and classification tasks in
the study.
1
7.2 feature extraction
7. Materials and Method
17
In the study, the researchers adapted pre-
7.1 Dataset trained Convolutional Neural Network
(CNN) models for the task of malaria
In the referenced study, a collection of blood- detection from blood-smear images. They
smear images from Chittagong Medical modified each CNN model by removing the
Hospital in Bangladesh was utilized. This top-most layers, which are typically
dataset was composed of an equal number of responsible for classification. The remaining
images depicting infected and uninfected lower layers of the CNNs, known for their
erythrocytes, categorized into two distinct feature extraction capabilities, were
classes. For the purpose of feature extraction, preserved. These layers are adept at
identifying patterns and characteristics
within the images that are indicative of the  NB: This algorithm classifies entities
presence of malaria parasites. based on distinct features, utilizing
Bayes’ theorem, expressed as
Subsequently, independent classification 1
algorithms were applied in place of the P(A|B) = \frac{P(B|A)P(A)}{P(B)}
original top layers. This approach allows the
models to leverage the powerful feature , to calculate the probability of ( A )
extraction of the pre-trained networks while given ( B ).
tailoring the final classification to the specific
task of identifying malaria in blood smears.  KNN: As a versatile algorithm, KNN
By doing so, the models could be fine-tuned is used for both classification and
to more accurately recognize and classify the regression problems, grouping data
medical images in the study’s dataset. This points based on similarity in features.
method is a common practice in deep
learning, enabling the application of These algorithms were chosen for their
sophisticated, pre-trained models to new performance and simplicity, aiming to
domains with relatively minimal additional provide a comprehensive comparison across
training. different deep-learning features extracted
from the images. Each algorithm brings a
7.3 Classification different approach to the classification task,

In the study, a selection of classification

algorithms was made to evaluate blood-
smear images for malaria
1
detection. The
chosen algorithms are Decision Tree (DT),
Support Vector Machines (SVM), Naïve
Bayes (NB), and K-Nearest Neighbour
(KNN), each offering unique strengths: Figure 5 training process

 DT: Known for its interpretability, allowing for a diverse evaluation of the
DT classifies by recursively models’ capabilities.
partitioning data into smaller subsets
based on tests at each node. It’s
advantageous for its quick learning
and handling of nonlinear 8. Result and discussion
relationships.
 SVM: A robust algorithm for binary The results in Table 1, derived from the
classification, SVM segregates inputs VGG16 features, provide a comprehensive
by finding an optimal separating evaluation of each classification algorithm’s
hyperplane, maximizing the margin performance in identifying malaria from
between classes. It’s memory- blood-smear images. The table likely
efficient and provides a clear includes the following metrics for each
distinction between classes. algorithm:
  Accuracy: The overall correctness of selection of the most suitable algorithm for
the model in classifying images as practical application in malaria diagnosis.
either containing the malaria parasite Precision and recall are particularly
18
critical in
or not. 9
medical diagnostics, where both false
 Precision: The proportion of true positives and false negatives carry significant
positive predictions in relation to all consequences. The F1-score is a crucial
positive predictions made by the measure as it balances the trade-off between
model. It is calculated using the precision and recall, ensuring that the model
formula does not favor one at the expense of the other.
1
The training time also provides insight into
\text{Precision} = \frac{TP}{TP + the computational efficiency of each
FP} algorithm.

, where ( TP ) is the number of true Table 1 Classionfication result VGG16

positives and ( FP ) is the number of
false positives.

 Recall (Sensitivity): The ability of

the model to identify all relevant
instances correctly. It is calculated as
12
\text{Recall} = \frac{TP}{TP + FN}

, where ( FN ) is the number of false

negatives.
4 The performance metrics from Table 1, based
 F1-Score: The harmonic mean of on VGG16 features, indicate how each
precision and recall, providing a classification algorithm fares in identifying
balance between the two. It is malaria from blood-smear images:
especially useful when the class
distribution is uneven. The F1-score  Decision Tree (DT):
is calculated as o Precision: 89.10%
10 o Recall: 89.11%
F1\text{-Score} = 2 \times o F1-Score: 89.22%
\frac{\text{Precision} \times o Accuracy: 89.24%
\text{Recall}}{\text{Precision} +  Support Vector Machine (SVM):
\text{Recall}} o Precision: 94.26%
o Recall: 95.57%
. o F1-Score: 94.91%
o Accuracy: 94.88%
 Training Time: The duration taken  Naïve Bayes (NB):
to train the model on the dataset. o Precision: 62.83%
o Recall: 96.47%
These metrics collectively offer a nuanced o F1-Score: 76.10%
view of the models’ strengths and o Accuracy: 69.70%
weaknesses, allowing for an informed  K-Nearest Neighbour (KNN):
 o Precision: 88.66% The performance metrics from Table 3, based
o Recall: 96.49% on ResNet50 features, are as follows for each
o F1-Score: 92.41% classification algorithm:
o Accuracy: 92.07%
 Decision Tree (DT):
These results highlight the strengths and o Precision: 86.89%
weaknesses of each algorithm. For instance, o Recall: 87.17%
SVM shows a high balance between o F1-Score: 87.03%
precision and recall, indicating a strong o Accuracy: 87.05%
performance in both correctly identifying  Support Vector Machine (SVM):
positive cases and minimizing false positives. o Precision: 94.22%
On the other hand, NB, while having a high o Recall: 95.53%
recall, shows lower precision, suggesting it o F1-Score: 94.87%
may incorrectly classify more non-parasitic o Accuracy: 94.84%
images as parasitic. The F1-Score and  Naïve Bayes (NB):
accuracy further provide insight into the o Precision: 80.82%
overall effectiveness of the algorithms, with o Recall: 87.67%
SVM and KNN outperforming DT and NB o F1-Score: 84.11%
on these fronts. These insights can guide the o Accuracy: 83.43%
selection of an appropriate algorithm for  K-Nearest Neighbour (KNN):
practical applications in malaria detection. o Precision: 84.69%
o Recall: 94.76%
These results indicate that SVM performed the o F1-Score: 89.44%
best in terms of precision, recall, and overall o Accuracy: 88.81%
accuracy among the four algorithms when using
VGG19 features. Naïve Bayes, while having the These results suggest that SVM is the most
highest recall, shows lower precision and accurate and precise algorithm for classifying
accuracy, which could lead to a higher rate of malaria from blood-smear images when
16
false positives. The F1-Score, which balances using ResNet50 features. It shows a high
precision and recall, is also highest for SVM, level of performance across all metrics,
14
suggesting it is the most reliable model for this indicating its effectiveness in correctly
particular task according to the VGG19 feature identifying positive cases and minimizing
set. KNN also shows strong performance, false positives. DT also shows good
particularly in recall and F1-Score, making it a performance, but SVM outperforms it. NB,
good candidate for scenarios where the cost of while having a decent recall, has lower
false negatives is high. precision, which could lead to more false
Table 2 Classification of VGG19 positives. KNN has a high recall, making it a
strong candidate for scenarios where missing
a positive case would be particularly
detrimental. The F1-Score for each algorithm
provides a balanced view of their precision
and recall, which is crucial for evaluating
their overall performance.
 Table 3 classification result of Resnet50 precision and accuracy, which may lead to a
higher rate of false positives. KNN has a
good balance between precision and recall,
making it a viable option, especially in
scenarios where high recall is critical. The
F1-Score for each algorithm provides a
balanced view of their precision and recall,
which is essential for evaluating their overall
performance.

Table 4 Classification result of DenseNet121

The performance metrics from Table 4, based
on DenseNet121 features, show the
following results for each classification
algorithm:

 Decision Tree (DT):

o Precision: 87.45%
o Recall: 86.82%
o F1-Score: 87.13%
Figure 3 illustrates a comparative analysis of
o Accuracy: 87.08% the accuracy of different classification
 Support Vector Machine (SVM): algorithms. The Support Vector Machine
o Precision: 92.55% (SVM) classifier stands out as the most
o Recall: 96.15% effective, particularly when utilizing VGG16
o F1-Score: 94.32% features, which are noted for their strong
o Accuracy: 94.21% ability to distinguish between classes. This
 Naïve Bayes (NB): suggests that SVM, coupled with VGG16
o Precision: 71.67% features, is highly reliable for classifying
o Recall: 73.79% blood-smear images in the context of malaria
o F1-Score: 72.71%
detection.
o Accuracy: 72.31%
 K-Nearest Neighbour (KNN):
o Precision: 79.77%
o Recall: 89.27%
o F1-Score: 84.26%
o Accuracy: 83.00%
15
These results indicate that SVM outperforms
the other algorithms in terms of precision,
recall, and accuracy when using
DenseNet121 features. It suggests that SVM Figure 6 Comparision between Each classifier and feature
is particularly effective at correctly 2
classifying blood-smear images as either 9. What would it take for
containing malaria parasites or not. DT also radiologists to accept deep
shows strong performance, but SVM’s higher
recall indicates it is better at identifying all learning tools in their everyday
relevant instances. NB has the lowest practice?
 2
Incorporating new data into a training system provided some details about the different
is a common practice in machine learning to architectures of deep learning especially
improve model performance and adapt to convolutional neural network. Also, we presented
new information. When radiologists upload challenges that militate against the performance
of Deep learning models.
new data, there are several approaches to
integrate this data into the existing model:
Training from Scratch: This involves
retraining the model on the entire dataset,
including the new data. While this can be
time-consuming, it ensures that the model
learns from the complete dataset.
Using Pre-trained Networks: A more
efficient approach is to use transfer learning,
where a pre-trained network is fine-tuned
with the new data. This method significantly
reduces training time and computational
resources.
8
To address overfitting, which occurs when a
model learns the training data too well and
19
performs poorly on unseen data, techniques
such as cross-validation, regularization, and
dropout can be employed. Regular updates to
the dataset and retraining the model can help
maintain its accuracy and generalizability.

2
Conclusion and Outlook
Machine learning and deep learning are
significant technologies that have changed the
medical history. These paradigms have
significantly changed how people store, process,
and analyse data to make better decision. These
have also led to research in developing
technology that changes drastically with time.
Machine learning techniques can be used to solve
complicated real-world problems. The relevance
of machine learning especially Deep learning in
resolving the difficult issues of malaria detection
and classification is highlighted in this article.
This paper has presented a detailed review of all
these new paradigms by particularly illustrating
their contributions to the healthcare system. We
 Similarity Report ID: oid:16158:62619656

17% Overall Similarity

Top sources found in the following databases:
17% Internet database 1% Publications database

TOP SOURCES
The sources with the highest number of matches within the submission. Overlapping sources will not be
displayed.

mdpi.com
1 6%
Internet

deepai.org
2 5%
Internet

ncbi.nlm.nih.gov
3 <1%
Internet

nature.com
4 <1%
Internet

mafiadoc.com
5 <1%
Internet

alls-academy.com
6 <1%
Internet

productschool.com
7 <1%
Internet

lup.lub.lu.se
8 <1%
Internet

coursehero.com
9 <1%
Internet

Sources overview
 Similarity Report ID: oid:16158:62619656

link.springer.com
10 <1%
Internet

deepgram.com
11 <1%
Internet

systematicreviewsjournal.biomedcentral.com
12 <1%
Internet

discovery.researcher.life
13 <1%
Internet

ijsrcseit.com
14 <1%
Internet

1login.easychair.org
15 <1%
Internet

arxiv.org
16 <1%
Internet

dro.dur.ac.uk
17 <1%
Internet

fastercapital.com
18 <1%
Internet

medrxiv.org
19 <1%
Internet

Sources overview