Automated Detection of COVID-19 in Chest Radiographs: Leveraging Machine Learning
Approaches
BatoolRaheela1
RazaGhulam Musa2
KhalidUsman3
KimByung-Seo2
-
( Department of Business Administration, Pusan National University, Korea batoolrahee@pusan.ac.kr)
-
( Department of Software & Communication Engineering, Hongik University, Sejong, Korea
ghulammusaraza96@gmail.com, jsnbs@hongik.ac.kr)
-
( Department of Computer Science, Comsats University Islamabad, Sahiwal, Pakistan
usmanchoudhary0000@gmail.com)
Copyright © The Institute of Electronics and Information Engineers(IEIE)
Keywords
COVID-19 pandemic, Machine learning models, Chest X-ray classification, Automated identification, Medical diagnosis
1. Introduction
The coronavirus that is causing the COVID-19 pandemic has spread quickly around the
world, leading the World Health Organization (WHO) to proclaim a global health emergency
on March 11, 2020 [1]. Declarative action highlighted the startling rate and severity of COVID-19, designating
it as the first coronavirus-caused pandemic. To lessen its effects, governments all
around the world put in place a variety of policies, including social separation,
flight restrictions, border restrictions, and public awareness campaigns. There has
been a notable increase in the number of cases and deaths. Specifically, on January
11, 2020, there were 41 confirmed cases; on February 11, 2020, there were 43,109;
and by May 11, 2020, there were 4.04 million cases. Despite precautions like donning
masks and avoiding social interactions, the number of cases increased threefold during
the next two months.
The World Health Organization (WHO) has recommended some tests to prevent the pandemic.
These tests include the Cobas SARS-CoV-2 test, compatible with the Cobas 6800/8800
systems, and the Genesis Real-Time PCR Coronavirus (COVID-19) test [5]. Although conducting these tests requires a significant investment of both time and
money, CNN (Convolutional News Network) technology has the potential to automate the
detection of infected people. This strategy could ultimately result in financial and
time savings and perhaps protect lives. It might also be used as an extra means of
validation, addressing the inherent shortcomings of the accuracy of the testing that
is done now.
There remains a gap in the practical application of convolutional neural networks
(CNNs) for COVID-19 detection from medical pictures, despite breakthroughs in their
use. Only a small amount of research incorporates automated tools into routine clinical
procedures, emphasizing the necessity for investigations that confirm model effectiveness
across a range of contexts. Although accuracy has been emphasized in recent studies,
little is known about the real-time use and computing efficiency of CNNs in hectic
clinical settings. Future research is required to build frameworks for interoperability
since the smooth integration of these tools into healthcare systems is impeded by
the lack of common protocols. While studies demonstrate great accuracy, to guarantee
inclusivity, it is necessary to evaluate model resilience over a range of groups.
There is a dearth of studies on the ethical and legal implications of using AI tools
in therapeutic settings, which raises issues like patient permission and data protection.
While the literature currently in publication emphasizes AI independence, future studies
ought to focus on cooperative decision-making between AI systems and human specialists.
Since current research only offers snapshots, longitudinal studies that take virus
evolution and imaging variance into consideration are necessary to assess the durability
of the model over time. To ensure the efficacy, dependability, and moral application
of COVID-19 detection techniques in actual healthcare settings, these gaps must be
filled. To this end, the proposed methodology operates within a multiclassification
framework. This novel categorization scheme represents a significant departure from
previous work, thereby enriching the diagnostic landscape. Moreover, the proposed
scheme not only achieves accurate COVID-19 detection but also effectively distinguishes
between various respiratory conditions with commendable performance.
2. Literature Review
Numerous scientists are working hard to examine data from photos of people who have
COVID-19. While some specialists concentrate on CT scans, others suggest a variety
of computer models made to learn and classify X-ray pictures from chest scans. An
advanced computer method was created in a study by the authors [6] to evaluate the presence of COVID-19 in X-ray chest pictures. As experts, three different
computer models (DenseNet201, Resnet50V2, and Inceptionv3) were used; each one learned
from photos on its own and produced predictions on its own. The models then produced
a result by combining their predictions in a novel way, similar to blending colors.
The researchers divided a dataset of 538 X-ray pictures from people with COVID-19
and 468 images from people without the virus into learning and practice sets to evaluate
their methodology.
In a different study [7], scientists suggested using a person's chest X-ray picture to help computers recognize
COVID-19 in that person. A dataset was used that included 194 X-ray scans of verified
COVID-19 patients and 194 images of healthy people. Owing to the scarcity of COVID-19
photos, the authors utilized "transfer learning," which allowed the computer to take
advantage of prior information from a variety of image datasets. These computer structures
were combined with additional advanced methods, including support vector machines
(SVM), k-nearest neighbor, Bayes, random forest, and multilayer perceptron’s (MLP).
Convolutional Neural Networks (CNN) were used in a method that was introduced in [8] to distinguish between five groups (COVID-19, tuberculosis, viral pneumonia, bacterial
pneumonia, and healthy) based on CT and X-ray pictures. The accuracy of this CNN adaptation,
called Mobile Net, was an astounding 98.80%. Similarly, a CNN system showed 100.0%
accuracy in classifying 150 photos into three categories (COVID-19, pneumonia, and
healthy) in the investigation of Bassi and Attux [9].
When authors in [10] used the GLSZM method to find parts, they were able to get a high level of accuracy
of about 98.77% after taking features from images and feeding them into a support
vector machine (SVM). An automated technique for detecting COVID-19 was proposed in
[11]. It involved preparing images, extracting features using the Residual Exemplar Local
Binary Pattern (ResExLBP), and selecting features using the iterative Relief (IRF)
method. A variety of validation techniques were used to assess the accuracy of the
classifiers, including decision trees, linear discriminants, support vector machines,
k closest neighborhoods, and subspace discriminants.
Researchers combined results in a thorough study [12] to improve the accuracy of COVID-19 testing and offer more details on the severity
of the condition. The procedure isn't flawless, as radiologists and specialized experts
typically analyze chest X-ray images, which is a difficult task. Utilizing three alternative
methods to produce the images and evaluating the efficacy of the computer models,
the researchers tested over 79,500 X-ray images, including 8,500 from COVID-19 instances,
and explored various methodologies utilizing a deep neural network. Another potential
work [13] introduces a novel deep learning architecture aimed at analyzing lung conditions,
such as COVID-19 and pneumonia, utilizing chest CT scans and X-ray images (CXR).
Termed as optimized DenseNet201 for lung diseases (LDDNet), this framework incorporates
additional layers, including 2D global average pooling, dense and dropout layers,
and batch normalization, built upon the base DenseNet201 model. LDDNet comprises 1024
dense layers activated by ReLU and 256 dense layers employing the sigmoid activation
method. Model hyper-parameters, such as learning rate, batch size, epochs, and dropout
rate, underwent meticulous tuning. Subsequently, three distinct datasets of lung diseases
were curated from various open-access repositories. Study [14] constructed a deep learning model employing transfer learning, leveraging optimized
DenseNet-169 and DenseNet-201 models for a three-class classification task, with the
Nadam optimizer. Modifications were made to the conventional DenseNet architecture,
and hyperparameters were fine-tuned to enhance the model's efficacy. Evaluation of
the model was conducted on a novel dataset comprising 3312 X-ray images sourced from
publicly available repositories. Performance assessment utilized various metrics including
accuracy, recall, precision, F1-score, and area under the receiver operating characteristics
curve (AUC-ROC).
In essence and contrary to existing work which is mainly focused on binary classification
, the proposed methodology is working on the multiclassification (normal, COVID-19,
viral pneumonia, and lung opacity (non-COVID lung infection)) of chest x-ray images.
The proposed methodology categorized the images into four subcategories, which is
the main novelty of the proposed scheme, besides detecting COVID-19 with good scores.
3. Proposed Methodology
The primary objective of this project is to develop a machine-learning model capable
of accurately classifying chest X-ray pictures into four distinct categories: normal,
COVID-19, viral pneumonia, and lung opacity (non-COVID lung infection). The methodology
section provides a comprehensive outline of the actions taken to achieve the stated
objective. It encompasses various areas, including the collection of data, the techniques
employed for data preparation, the logic underlying the selection of models, and the
metrics employed for evaluation.
Developing a machine learning model to accurately classify chest X-ray pictures into
four groups such as normal, COVID-19, viral pneumonia, and lung opacity (non-COVID
lung infection) is the primary objective of this research. This methodology section
provides a comprehensive summary of the actions taken to accomplish this goal. It
includes the methods for gathering data, preparing the data, selecting the model,
and using metrics for assessment.
3.1 Data Preprocessing
To begin the preparation phase, we standardize the size of all photos in the collection
to 150x150 pixels. This is critical to ensuring that the machine learning model always
receives consistent data. The photographs are also converted to grayscale. To lessen
the computational complexity of the model, color is typically not necessary for the
classification of medical X-ray images. Although picture classification tasks frequently
make use of techniques like rotation, flipping, and zooming to increase the amount
and variety of the training dataset, these were not included in this study. The decision
was made to protect the originality of the X-rays, which is essential for medical
diagnostics. After the images have been resized and converted to grayscale, a label
reflecting their classification is applied. One-hot encoding is then used to convert
these labels into a categorical representation for use in training. Eighty percent
of the dataset is then divided into training and validation sets, while twenty percent
is set aside for testing. This distribution allows the model to be trained on a large
dataset while still leaving room for validation and testing, leading to a more accurate
assessment of the model's effectiveness.
3.2 Model Architecture
The VGG-16 architecture, proposed by the Visual Geometry Group (VGG) at the University
of Oxford, is renowned for its simplicity and effectiveness. It consists of 16 layers,
including 13 convolutional layers and 3 fully connected layers. The convolutional
layers are primarily comprised of 3x3 filters with a stride of 1, resulting in a receptive
field that captures fine spatial details while preserving spatial dimensions. This
configuration allows the network to learn intricate features at multiple scales, leading
to superior performance in image classification tasks.
Proposed VGG-16 is its uniformity in design, where each convolutional block consists
of multiple stacked convolutional layers followed by a max-pooling layer. This uniformity
simplifies the implementation and interpretation of the model while promoting feature
reuse across different layers.
Because of its proven success in these types of picture identification tasks, the
VGG16 architecture was chosen for this investigation. The VGG-16 network has 16 hidden
layers and 138 million parameters, making it a deep convolutional neural network.
While this parameter count contributes to the model's expressive power, it also necessitates
substantial computational resources for training and inference. Techniques such as
transfer learning, where pre-trained VGG-16 weights are fine-tuned on specific datasets,
are often employed to mitigate the computational burden and expedite training convergence.
Intricate picture classification jobs are no match for their power, even though they
are very huge. To preserve the spatial dimensions of feature maps, the architecture
uses 3x3 convolutional layers with a stride of 1, utilizing the same padding. Dimensional
and computational costs are further reduced by the inclusion of 2x2 max-pooling layers
with a stride of 2. An alternative convolutional neural network is built alongside
the standard VGG16 model for evaluation purposes. Our bespoke model incorporates state-of-the-art
features, including separable convolutions and batch normalization to improve computational
efficiency and model generalization.
Overfitting is a prevalent problem in deep learning models, although it can be prevented
by carefully designed dropout layers. The proposed architecture incorporates dropout
layers. Dropout randomly deactivates a fraction of neurons during training, effectively
introducing noise and preventing the network from relying too heavily on specific
features or representations. This regularization technique encourages the model to
learn more robust and generalizable features, leading to improved performance on unseen
data. The Adam optimizer, characterized by its adaptive learning rate capabilities
and momentum-based updates, offers several advantages for training deep neural networks.
By dynamically adjusting the learning rate for each parameter. The multi-label classification
problem is addressed by using the binary cross-entropy loss function. Training covers
10 epochs with a batch size of 32, utilizing the training dataset. To prevent overfitting,
the validation dataset is used to fine-tune the model's parameters. The methodology
is completed with an analysis of the efficiency of the trained model. Accuracy, measured
on the training dataset, is the major indicator of success. Due to the medical nature
of the evaluation, extra measures like precision, recall, and F1-score are considered.
Important for making a medical diagnosis, these metrics shed light on the model's
ability to correctly identify each class.
Fig. 1. Chest X-ray Image.
Fig. 2. X-ray Classification: ML Methodology.
Fig. 3. Model Architecture.
Fig. 4. Training Loss and Accuracy Dataset.
4. Results
The goal of this section is to provide a comprehensive evaluation of the machine-learning
model developed to classify chest X-ray pictures into four groups: normal, COVID-19,
viral pneumonia, and lung opacity. Loss, precision, MAE, and the confusion matrix
are just a few of the measures used to assess the model's efficacy. The effects of
these measurements, variations in learning rates, and graphical representations of
the model's forecasts will all be investigated. There will be an analysis of the metrics
employed during the training and validation processes.
4.1 Epoch-wise Analysis
We trained the model for 10 iterations and collected metrics at the end of each iteration
for both the training and validation sets. The training accuracy in the first time
was about 72.76 percent, whereas the validation accuracy was about 16 percent. The
low validation accuracy in the first epoch indicated that the model did not generalize
well to unknown data. Training and validation accuracy, however, got much better as
more time passed between epochs. The model's training accuracy reached 91.54% at the
10th epoch, while the validation accuracy reached 87.48%.
4.2 Loss and Accuracy
Binary cross-entropy was utilized as the loss function in the first model, whereas
mean squared error was used in the second. The loss values showed a gradual decline
from the first to the last epoch, indicating that the model had successfully learned.
In the first model, the final loss was 0.1184 on the training set and 0.1653 on the
validation set. With the second model, the loss dropped to 0.0354 on the training
set and 0.0457 on the validation set, showing a better fit.
4.3 Mean Absolute Error
The second model also tracked the mean absolute error to measure the typical discrepancy
between forecasted and observed values. By the 10th epoch, the model showed a mean
absolute error of roughly 0.0695 on the training set and 0.0761 on the validation
set, both of which are smaller than the ideal values of 0.08 and 0.09, respectively.
4.4 Test Accuracy and Loss
After being put through its paces on the test dataset, the model performed with an
accuracy of 88.14 percent and a loss of 0.0449. These numbers coincide with the validation
numbers, proving that the model can successfully apply its findings to novel data
sets.
4.5 Confusion Matrix Analysis
The information from the confusion matrix clarified the model's effectiveness. It
found that the model had better-than-average performance in identifying COVID-19 and
normal instances and that it had higher precision and recall overall. The model's
robustness was further confirmed by the fact that the F1-score, which is a balanced
combination of precision and recall, was high across all classes.
4.6 Learning Rate Adjustments
A tailored step decay function was used to dynamically alter the training rate. The
initial learning rate was 0.0001, but it has been adjusted based on past losses. The
convergence was so successful that the final learning rate was close to 0.000106.
The model's remarkable success is attributed in part to the fact that the training
rate may be fine-tuned to allow for subtle weight updates.
4.7 Visual Interpretation of Predictions
To provide a more intuitive understanding of the model's efficacy, visual representations
of its predictions were developed. Color-coded bar graphs and graphics were deployed
to demonstrate the alignment between the model's predictions and the actual labels.
The abundance of blue prediction bars served as proof that most predictions were accurate.
Red bars indicated times when the model made a mistake, providing useful context for
understanding the kinds of faults to which the model is most susceptible.
Fig. 5. Training Loss and mean absolute error on Dataset.
Fig. 6. Confusion Matrix.
Fig. 7. Accuracy score report.
Fig. 8. Model Predictions Visualization.
5. Significance and Implications
The exceptional precision of the model holds substantial significance for the field
of healthcare, specifically in the prompt and precise identification of ailments such
as COVID-19. The tool's strong performance across multiple parameters indicates its
potential relevance as a resource for healthcare practitioners. Nevertheless, it is
imperative to recognize that while the model demonstrates exceptional performance
with this dataset, it is imperative to do extensive testing on a broader spectrum
of X-ray pictures to thoroughly evaluate its effectiveness. The assessment of the
model's performance highlights its notable precision and dependability in categorizing
chest X-ray pictures into four distinct classes. Significantly, the model demonstrated
high performance not only on the training and validation sets but also exhibited robust
generalization abilities on the test set. When adaptive learning rate modifications
and mean absolute error were added as evaluation metrics, they added more dimensions
of analysis, which made the model more resilient overall. The encouraging findings
indicate that the model has the potential to be a valuable instrument in aiding medical
diagnosis, particularly for urgent illnesses such as COVID-19 that require immediate
attention.
6. Conclusion
Lots of countries around the world are seeing a decrease in resources while COVID-19
cases keep going up every year. It's really important to find positive cases, making
sure none are missed. Considering this viewpoint and understanding the pressing nature
of the situation, we propose a technique using deep learning to detect t the COVID-19
virus through chest X-rays. We introduce a machine-learning model designed to classify
chest X-ray images such as normal, COVID-19, viral pneumonia, and lung opacity. This
showcases the effectiveness of the model in aiding medical diagnosis, particularly
in time-sensitive scenarios like COVID-19.
7. Future Direction
In future directions, we have aim to explore integrating multimodal data such as textual
reports alongside chest CT scans and X-ray images to enhance predictive capabilities.
Fine-tuning the model for specific lung pathologies, leveraging transfer learning
with larger datasets, and deploying explainable AI techniques for clinical decision
support are also promising avenues. Additionally, longitudinal studies and prognostic
analysis could extend the model's utility in predicting disease progression and patient
outcomes. These directions aim to advance medical image analysis and improve the diagnosis
and management of lung diseases in clinical practice.
ACKNOWLEDGMENTS
This work was supported by the National Research Foundation (NRF), Korea, under project
BK21 FOUR and in part by a 2024 Hongik University innovation support program fund.
REFERENCES
WHO. (2020). WHO Director-General's opening remarks at the media briefing on COVID-19
- 11 March 2020.
Y.-C. Wu, C.-S. Chen, and Y.-J. Chan, ``The outbreak of covid-19: An overview'' Journal
of the Chinese medical association, vol. 83, no. 3, p. 217, 2020. Article
W. H. O., ``Modes of transmission of virus causing covid-19: implications for ipc
precaution recommendations: scientific brief, 27 march 2020'' World Health Organization,
Tech. Rep., 2020.
W. H. O. coronavirus disease (COVID-19) dashboard. Geneva: World health organization,
2023.
W. H. O. coronavirus disease (COVID-19), 12 October 2020.
Das, A. K., Ghosh, S., Thunder, S., Dutta, R., Agarwal, S., & Chakrabarti, A. (2021).
Automatic COVID-19 detection from X-ray images using ensemble learning with convolutional
neural network. Pattern Analysis and Applications, 24, 1111-1124.
E. F. Ohata et al., "Automatic detection of COVID-19 infection using chest X-ray images
through transfer learning," in IEEE/CAA Journal of Automatica Sinica, vol. 8, no.
1, pp. 239-248, January 2021, etc: 10.1109/JAS.2020.1003393.
Jia, G.; Lam, H.-K.; Xu, Y. Classification of COVID-19 chest X-Ray and CT images using
a type of dynamic CNN modification method. Comput. Biol. Med. 2021, 134, 104425.
Bassi, P.R.; Attux, R. A deep convolutional neural network for COVID-19 detection
using chest X-rays. Res. Biomed. Eng. 2021, 1-10.
Cristianini, Nello, and John Shawe-Taylor. An introduction to support vector machines
and other kernel-based learning methods. Cambridge university press, 2000.
Tuncer, T., Dogan, S. and Ozyurt, F., 2020. An automated Residual Exemplar Local Binary
Pattern and iterative ReliefF based COVID-19 detection method using chest X-ray image.
Chemometrics and Intelligent Laboratory Systems, 203, p. 104054.
J. D. Arias-Londo\~{n}o, J. A. G\'{o}mez-Garc\'{i}a, L. Moro-Vel\'{a}zquez and J.
I. Godino-Llorente, "Artificial Intelligence Applied to Chest X-Ray Images for the
Automatic Detection of COVID-19. A Thoughtful Evaluation Approach," in IEEE Access,
vol. 8, pp. 226811-226827, 2020,
Podder, P.; Das, S.R.; Mondal, M.R.H.; Bharati, S.; Maliha, A.; Hasan, M.J.; Piltan,
F. LDDNet: A Deep Learning Framework for the Diagnosis of Infectious Lung Diseases.
Sensors 2023, 23, 480.
Podder, P.; Alam, F.B.; Mondal, M.R.H.; Hasan, M.J.; Rohan, A.; Bharati, S. Rethinking
Densely Connected Convolutional Networks for Diagnosing Infectious Diseases. Computers
2023, 12, 95.
Raheela Batool obtained a Bachelor of Science in Business Administration from BUITEMS
University, Pakistan, in 2018. She pursued a degree in human resource management during
her undergraduate years. Later, she obtained her Master of Science degree in Business
Administration with a specialization in marketing from the University of Ulsan, South
Korea, in 2021. During her master's program, she conducted research mostly on consumer
behavior and digital marketing. From 2020 to 2021, Raheela held the position of sales
and marketing representative at an automobile trading company located in South Korea.
She is currently undertaking her doctoral studies in management information systems
at Pusan National University in South Korea.
Ghulam Musa Raza received his BS in Computer Sciences from Comsats University Islamabad
in 2019 majoring Intelligent Robotics. He received his MS in Computer Sciences from
SEECS, NUST Islamabad in 2021. His research interest at the time was Natural Language
Processing (Artificial Intelligence). From 2017 to 2019, he worked as a Software Engineer
for Snaky Solutions Pvt Limited. He served as a Machine Learning Research Assistant
in the TUKL lab, NUST Islamabad, at the start of 2021. He was a Lecturer at Alhamd
Islamic University, Islamabad, from 2021 to 2022. His major interests now are Natural
Language Processing, the Internet of Things (IoT), Information-centric Networking,
and Named Data Networking. He is pursuing a PhD in the Department of Communication
and Software Engineering Graduate School, Hongik University, South Korea.
Usman Khalid received his BS in Computer Sciences from Comsats University Islamabad
in 2020 majoring in artificial intelligence. From 2021 to 2022, he worked as a Software
Engineer for Tech Solutions Ltd, where he specialized in developing AI-driven applications.
Currently, he is working as a freelance Software Engineer, focusing on online projects
in the fields of machine learning and data science. His research interest is machine
learning, deep learning, computer vision and natural language processing.
Byung-Seo Kim received his BS in electrical engineering from Inha University, Incheon,
Korea, in 1998 and obtained his MS and PhD in electrical and computer engineering
from the University of Florida in 2001 and 2004, respectively. His PhD was supervised
by Dr. Yuguang Fang. Between 1997 and 1999, he worked for Motorola Korea Ltd., PaJu,
Korea, as a computer integrated manufacturing (CIM) engineer in advanced technology
research and development (ATR&D), and he was the chairman of the Department of Software
and Communications Engineering, Hongik University, South Korea, where he is currently
a professor. He served as the General Chair for the 3rd IWWCN 2017 and was the TPC
member for IEEE VTC 2014-Spring and the EAI FUTURE 2016 and the ICGHIC 2016 and 2019
conferences. He has served as guest editor of special issues for the International
Journal of Distributed Sensor Networks (SAGE), IEEE Access, and Journal of the Institute
of Electrics and Information Engineers. His work has appeared in 167 publications
and he holds 22 patents. He is an IEEE Senior Member and Associative Editor of IEEE
Access. His research interests include the design and development of efficient wireless/wired
networks, including link-adaptable/cross-layer-based protocols, multiprotocol structures,
wireless CCN/NDN, mobile edge computing, physical layer design for broadband PLC,
and resource allocation algorithms for wireless networks.