Automated Diagnosis of Covid-19 and Pneumonia Using Transfer
Learning and Custom Segmentation on Chest X-Ray Images
ADVAIT K ASOK, LIDIYA LILLY THAMPI
Department of Computer Science & Engineering,
Indian Institute of Information Technology, Kottayam (IIITK),
Valavoor P.O, Pala Kottayam,
INDIA
Abstract: -
The key goals of this study are to discover, demonstrate, and quantify advancements in deep
learning approaches for classifying healthy, pneumonia of the community-acquired and viral types, and
COVID-infected lungs from X-ray images and to learn how the pre-trained models react to the training
with custom segmented images.
The proposed model uses the dataset pre-processed to generate unique
masks and segment the lung region to train a convolutional neural network with a transfer learning model
using VGG16 and VGG19 architecture.
The accuracy and F1 score results for 3-way classification with
custom processing are high for VGG19 with custom segmentation. In contrast, the results for the 4-way
classification were stable with and without custom processing for both VGG16 and VGG19 models.
Key-Words: -
Lung Segmentation, Transfer Learning, Chest X-Ray Image, Pneumonia, COVID-19, Image
Processing.
Received: January 4, 2024. Revised: April 9, 2024. Accepted: May 11, 2024. Published: June 25, 2024.
1 Introduction
Getting a chest x-ray (CXR) has always been the
best way to tell if someone has pneumonia or a
COVID-19 infection. Radiography’s benefits
include its familiarity with patients, low cost
compared to newer modalities like CT, seamless
incorporation into electronic medical records, and
widespread acceptance among medical
professionals. Low sensitivity in detecting a variety
of diseases, such as pulmonary edema and
pneumothorax, are two of its major downsides, [1].
Bacterial pneumonia is caused by streptococcus
pneumoniae, while viral pneumonia is caused by
viruses, characterized by fever, cough, muscle
aches, and sore throat, [2].
CXR image requires a professional and takes
up a lot of time for analysis as the X-rays of
bacterially acquired pneumonia typically show a
white concentrated patch
of porosity, while viral acquired pneumonia
typically shows widespread lung involvement,
[3]. This can lead to misdiagnosis, delay in treatmen
t, and increased cost of treatment [4]. RT-PCR
method is used to diagnose COVID-19 but requires
expensive equipment and at least 24 hours for a
viable result, [5]. Several studies have combined
three-way categorization of X-ray images along
with four-way categorization to distinguish between
COVID-19, viral and bacterial pneumonia, and
healthy lungs [6]. In this paper, CXR pictures are
used for transfer learning using VGG16 and
VGG19, as well as several image processing
techniques including adaptive histogram
equalization, grey scaling, binary threshold, black-
hat morphology, feature extraction, and contouring
to categorize and improve diagnostic accuracy by
forming a custom processed dataset by separating
cases of COVID infected lung, viral and bacterial
pneumonia infected lung, and a healthy lung. The
VGG19 method highly improved the model’s
precision, reliability, and F1 score compared to
previous segmentation and categorization methods.
2 Related Work
Using CXR images to diagnose lung diseases like
COVID-19 and pneumonia, machine-learning
algorithms have recently acquired traction in the
medical world.
In an analysis with 40 analogue chest CXRs from
patients with normal and pneumonia situations, [7],
images were cropped and the lung region was
extracted using methods developed in-house. To
differentiate between normal lung tissue and
diseased haze, Otsu thresholding was used to detect
pneumonia clouds. Healthy lung area as a
percentage of total lung area has been proposed to
be calculated to conclude.
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2024.12.32
Advait K Asok, Lidiya Lilly Thampi
E-ISSN: 2415-1521
328
Volume 12, 2024
By analyzing chest X-rays, suggested a deep-
learning CNN model be used to differentiate
between cases of healthy people, COVID-19, and
viral pneumonia in the lungs, [8]. The proposed
model uses VGG16-based pre-trained ImageNet
weights. The model was improved by adding new,
custom-made layers. VGG16 has an accuracy score
of 92.7%. A system was developed for transfer
learning and tested several deep-learning models
using data from COVID-19, viral pneumonia, and
pictures of normal lung CXR with counts at 423,
1485, and 1579, [9]. To classify the two
methodologies, the networks were trained using
healthy and COVID-induced pneumonia-infected
lung CXR pictures as well as normal, viral, and
COVID-induced pneumonia-infected X-ray images
with and without image augmentation. Both systems
were 99.7 and 97.1 percent accurate in their
classifications, respectively, while their specificity
was 99.55 and 98.1 percent, respectively. However,
it did not meet the criteria for bacterial pneumonia.
In a recent study, multiple Convolutional Neural
Networks were trained to classify x-ray images as
either pneumonia or non-pneumonia by altering the
network’s parameters, hyper-parameters, and
number of convolutional layers, [10]. Six distinct
models are covered in the research. One similarity
between the two models is the presence of two
convolutional layers. There are also four alternative
trained models available, including VGG16,
VGG19, ResNet50, and Inception-v3. The
validation accuracy for the first model is 85.26%,
while the second model is 92.31%. 87.28 percent is
the accuracy of VGG16, 88.46 percent of VGG19,
77.56 percent of ResNet50, and 70.99 percent of
Inception-v3.
To the best of my knowledge, this study is the
first to examine how segmented lung areas from X-
ray images might be used as input for classification
algorithms to perform better
3 Materials and Methods
DL methods require data retrieval, processing,
training, analyzing, optimizing, and classifying.
After the models have been trained on the novel
auto-segmented dataset, we’ll test how well the
models trained with and without custom
segmentation methods can distinguish between
healthy, COVID-19, bacterial, and pneumonia
caused by viruses in CXR pictures. Figure 2
represents the methods we follow in this study.
3.1 Chest X-Ray Image Dataset
Images of the chest X-rays were taken from the
Mendeley data repository. Figure 1 depicts the
current data set of 9208 images which includes
1,281 images of COVID-19 infected, 3270 images
of healthy, 1656 images of viral pneumonia-
infected, and 3001 images of bacterial pneumonia
infected lung CXR, [11]. For the study, we will
consider a total of 5000 images, with 1250 images
chosen at random from each group.
Fig. 1: Images of posterior-anterior chest X-rays
from COVID-19: A Curated Dataset (X-Rays),
[11]
3.2 Custom Lung Segmentation
The original images were of high resolution and
varied sizes. As per the observations from Table 1,
to make the computation and training faster and to
generate a better-segmented image of CXR the
images were resized to 256 x 256 pixels, [12]. To
guarantee the achievement of the dynamic signal
intensity, imaging data must be standardized and
converted to a gray-scale range of [0, 1] or [-1, 1],
[13].
Implementation of the custom segmentation
workflow is explained in Figure 3 (A) vertical line
with a center offset is drawn on the image at the
place of the maximum pixel sum to improve the
lung boundary identification, [7], as shown in
Figure 3 (B). The images with the center line erased
will be subjected to Contrast-Limited Adaptive
Histogram Equalization to boost image quality, as
depicted in Figure 3 (C), [14], [15].
The “Blackhat” morphological operation is
applied to the CLAHE-processed CXR pictures to
boost the image’s dark regions, [16], [17], as shown
in Figure 3 (D). The average pixel intensity value
as the threshold is applied to the image with black-
hat morphology, setting all pixels with intensities
greater than the mean to 255 and all pixels with
values less than the mean to 0, [18]. The result is
shown in Figure 3 (E).
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2024.12.32
Advait K Asok, Lidiya Lilly Thampi
E-ISSN: 2415-1521
329
Volume 12, 2024
Fig. 2: Proposed architecture diagram
Table 1. The image size and custom segmentation
time comparison
The corner mask image is then returned with the
detected corners highlighted in white and the rest of
the image in black, as shown in Figure 3 (F). The
Corner-circle mask, which is a binary mask image
with circular areas around the corners, and the
binary thresholded image will be multiplied element
by element to make the array mask as portrayed in
Figure 3 (G). The median blur is applied as shown
in Figure 3 (H). The list of contours is retrieved,
[19], as depicted in Figure 3 (I). This morphological
dilation operation is applied, [20], as depicted in
Figure 3 (J). Overlaying the mask on the CXR
creates a new image, as shown in Figure 3 (K), by
performing an arithmetic operation that adds two
input images together using their relative weights.
Created a new image in which the areas outside the
mask are blackened out while the areas within the
mask remain unchanged, as shown in Figure 3 (L).
3.3 CXR Classification with Transfer
Learning
In this study, the custom segmented image dataset
will be trained using
transfer learning
methodologies, and there will be two
categorizations for the
chest X-rays: Lung health
Fig. 3: Proposed custom segmentation workflow
is classified as either (i) three classes (healthy
lungs, covid infected lung, and viral pneumonia)
(ii) four classes (healthy lungs, covid infected
lung, viral pneumonia, and bacterial pneumonia)
using the pre-trained VGG16 and VGG19 models
Data
count
Image
shapes
Segmentati
on time (s)
5000
(499, 406),
(645, 715),
(450. 462),
(300, 400),
etc.
933.028
(256, 256)
177.905
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2024.12.32
Advait K Asok, Lidiya Lilly Thampi
E-ISSN: 2415-1521
330
Volume 12, 2024
Additionally, the efficacy of the VGG16 and
VGG19 models trained without a segmented
image dataset is compared.
4 Results and Outcomes
4.1 Experimental Design
The experiment was built using the Google Colab
platform and a Python 3 (GPU) backend for Google
Compute Engine. Multiplying arrays of input data
and output weights by a matrix during CNN training
is a computationally intensive process, [8]. The use
of Kaggle notebooks facilitated rapid training,
computing, and data administration.
4.1.1 Dataset Splitting
As shown in Figure 4, a random reordering is
performed on all the picture filenames in the root
Curated dataset directory with a total of 5000
images with 1250 images in each class and sorted
into three distinct piles based on the specified train
ratio (60%), validate ratio (10%), and test ratio
(30%).
4.1.2 Performance Measurement
The starting settings for the learning rate, epochs,
and number of batches are 0.0001 or 20 and 32 or
128, respectively
,
[21].
Fig. 4: CXR image distribution on test,
train, and validation
The efficacy of the trained models is
evaluated using accuracy and sensitivity. The
confusion matrix is used to compare the
proportions of true positives, true negatives, false
negatives, and false positive detections to assess
the accuracy of a model.
4.2 Evaluation of Three-Class Classification
with Custom Segmented Data
Using the custom segmented image dataset
generated in the earlier stage
of this study, we are
currently performing the initial stage of training a
VGG16 and VGG19 model to distinguish
between COVID-19, viral pneumonia, and a
healthy lung. The results are compared between
training on non-segmented plain X-ray pictures
and training on CXR images that have
been pre-
processed utilizing the unique segmentation method
developed for
this research, [22].
4.2.1 Transfer Learning with VGG16
For 100 epochs, it is observed that the model was
able to accurately categorize 97.33% of the data
points at an early stopping of 30 epochs as
shown
in
Figure 5
(A). With a precision of 0.9734, the
model correctly predicted 97.34% of the cases it
labeled as positive. The model has a recall of
0.9733, therefore it was able to properly identify
97.33% of all true positives. The model successfully
found 98.66% of false negatives and 1.33 percent of
negatives as positives. The model acquired 0.9733
as the F1-score representing a fair harmony between
the precision and the recall.
4.2.2 Transfer Learning with VGG19
According to the classification discussed in
Figure 5
(C)
, the F1 score of 0.9742 was reached
by the model VGG19 in terms of accuracy, recall,
and precision. A score of 0.0138 is received as
the false positive rate and a specificity score is
0.9862. Top scores for these tests suggest that the
VGG 19 model is effective in categorizing
COVID-19, two types of pneumonia, and normal
healthy lung images from a chest X-ray trained on
the custom segmentation methods. It can be
assumed that with more training on larger, clearer
CXR images and a more varied dataset, the
model’s efficiency and universality could be
improved.
4.3 Evaluation of Four-Class Classification
with Custom Segmented Data
Performed the second stage of training a VGG16
and VGG19 model to distinguish between
COVID-19, bacteria-induced pneumonia, viral
pneumonia, and a healthy lung.
4.3.1 Transfer Learning with VGG16
For 100 epochs, VGG16 has a false positive rate
of 0.052 and an F1- score of 0.8441, accurately
predicting 84.46% at an early stopping of
17.5
epochs as seen in
Figure 5
(B). A precision of
0.8435, recall of 0.8446, and specificity of 0.9482
suggest that the model is missing some true
positives. The custom segmented data training
did not result in a high improvement in the
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2024.12.32
Advait K Asok, Lidiya Lilly Thampi
E-ISSN: 2415-1521
331
Volume 12, 2024
performance of VGG-16 classifying the images
into the 4 categories.
4.3.2 Transfer Learning with VGG19
As per the classification for 100 epochs, an
accuracy of 0.844 suggests that almost 84 percent
of the dataset was properly identified by the
VGG19 model in 20 epochs as depicted in
Figure
5 (
D). The model’s precision is 84.16
%, which
means that more than three-quarters of the positive
samples were indeed positive. This statistic helps
gauge how well the model suppresses spurious
positive results. Out of all the true positive
samples in the dataset, the model accurately
identified almost 84 percent of them. Since
94.8
percent of the true negative samples in the training
set were accurately identified as such, the model
has a specificity of 0.948. Approximately 5.2
percent of the true negative samples in the
training set were mistakenly labeled as positive.
The model’s F1-Score, a measure of its overall
performance in terms of accuracy and recall equal
to the harmonic mean of these two metrics, is
0.8428. Based on these measures, it appears that
the VGG19 model did a decent job of
categorizing the samples, but it might use some
more work, especially in the areas of false
positives and false negatives.
4.4 Performance Comparison with Non-
Segmented Images
4.4.1 Transfer Learning with VGG16
For a 3-class classification, it is observed that with
an accuracy of 97.51%, the model was successful
in classifying all the cases. The model’s precision
was 0.9751, which suggests that 94.97% of positive
cases were indeed positive.
With a specificity of
0.9875, the model successfully identified the false
negatives, and 1.2% of negative cases were
incorrectly classified as positives.
The model had an F1-score of 0.9751, which is
the harmonic mean of precision and recall and
represents how well they are balanced.
Collectively, these findings are indicative of the
model’s low false positive rate and high accuracy
and precision in instance classification. According
to the F1-score, the model also struck a reasonable
balance between precision and recall, and the
curves are shown in Figure 6 (A).
Fig. 5: Segmented Dataset Classification
A) VGG-16 3 class B) VGG-16 4 class
C) VGG-19 3 class D) VGG-19 4 class
For the 4-way classification, the model has an
F1-score of 0.8556 and an accuracy of 85.53 %
after an early stop at 20 epochs as shown in
Figure 6
(B). From the metrics depicted in Table 2,
we can infer that a respectable fraction of true
positives is being identified by the model and that it
can accurately identify a sizable fraction of the data.
Table 2. VGG-16 performance metrics
Classifier
VGG-16
Parameters
3 Class
4 Class
non-
segmented
segmen
ted
non-
segmented
segmen
ted
Accuracy
97.51
97.33
85.53
84.46
Precision
97.51
97.34
85.59
84.35
Recall
97.51
97.33
85.53
84.46
Specificity
98.75
98.66
95.17
94.82
False Positive
0.0133
0.025
0.0482
0.0517
F1-Score
0.9751
0.973
0.8556
0.8441
A very small proportion of samples may have
been misclassified by the model as diseased due to
the false positive rate of 0.025 in the 3-way
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2024.12.32
Advait K Asok, Lidiya Lilly Thampi
E-ISSN: 2415-1521
332
Volume 12, 2024
classification than the 4-way classification with
segmented data with the score of 0.0517.
Fig. 6: Non-segmented dataset classification
A) VGG-16 3 class B) VGG-16 4 class
C) VGG-19 3 class D) VGG-19 4 class
The model works efficiently with the 3-class
categorization meanwhile the 4-class categorization
accuracy slightly decreases when the dataset is
processed through custom processing, resulting in
lower categorization precision.
4.4.2 Transfer Learning with VGG19
The model successfully labeled 96.97% of the test
dataset at an early stopping of 35 epochs as depicted
in Figure 5 (C). The F1-score of 0.9698 and the
other metrics shown in Table 3 further indicate that
the model struck an adequate balance between
accuracy and recall. All things considered, VGG19
appears to be a very efficient model for dividing
chest X-rays into three distinct groups when the
custom segmentation process is applied to the
dataset.
For a 4-way classification with an accuracy of
85.67% and F1 score of 0.855, the VGG19 model
proved capable of correctly categorizing the test
data. The model was accurate 85.51% of the time
when predicting a class label (precision= 0.8551).
From the metrics shown in Table 3 and Figure 5
(D), it seems that 4.78 percent of erroneous
negatives were mislabeled as true positives. When
we compare the algorithms for the two types of
classification, we can conclude that 3-class
categorization improves when custom segmentation
is applied to the dataset, whereas 4-class
categorization accuracy slightly decreases when the
dataset is processed through custom processing,
resulting in lower categorization precision.
Table 3. VGG-19 performance metrics
Classifier
VGG-19
Parameters
3 Class
4 Class
Non-seg
Seg.
Non-
seg.
Seg.
Accuracy
96.9
97.2
85.6
84.40
Precision
96.9
97.2
85.5
84.16
Recall
96.9
97.2
85.7
84.40
Specificity
98.5
98.6
95.2
94.80
FPR
0.02
0.01
0.05
0.052
F1-Score
0.96
0.97
0.85
0.842
5 Discussion
High levels of accuracy, precision, recall, and F1-
score were achieved in VGG16 than VGG19 when
used to categorize X-rays into three groups: viral
pneumonia, COVID-19 infected, and healthy lungs.
More specifically, VGG19 outperformed a bit more
than VGG16 in the categorization of X-rays into
three groups job, suggesting it is more suited to it.
However, there is room for improvement in the
custom segmentation approach used in this research.
Particularly bacterial pneumonia and some chest X-
rays with comparatively large clouded lung regions
tend to produce less accurate segmentation. We
need more research to see how other state-of-the-art
models, like EfficientNet, [23], stack up against
VGG19. The impressive architecture and
computational elements, the EfficientNet results in
impressive accuracy in the image categorization
tasks. It would be an interesting experiment to
compare the metrics generated with EfficientNet,
VGG19, and VGG16.
6 Conclusion
The classification with VGG19 resulted in a
higher accuracy score for the 3-way
categorization of the X-ray images. It suggests
that the custom segmentation discussed in this
paper enhances the performance of the model.
The 3-way categorization without the custom
segmentation resulted in the F1 score of 97.51
percent and 0.9751 respectively for VGG16 and
96.97 percent and 0.9698 for VGG19. The results
for the 4-way classification were 85.53 percent and
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2024.12.32
Advait K Asok, Lidiya Lilly Thampi
E-ISSN: 2415-1521
333
Volume 12, 2024
0.8556 for VGG16 and 85.66 percent and 0.8558 for
VGG19. Scores of 97.33 percent, 0.9733 for
VGG16 and 97.24 percent, 0.9724 for VGG19 in 3-
way classification, and 84.46 percent, 0.8441 for
VGG16 and 84.40 percent, 0.8428 for VGG19 in 4-
way classification were observed using custom
segmentation.
Future studies may refine the custom
segmentation approach and apply it to the dataset
without ground truth to better isolate lung areas and
minimize image noise, improving multi-
classification of lung X-rays. We may use Jaccard
similarity coefficients and Dice coefficients to
compare manual and automated segmentation.
The results of the study show that multi-class
categorization of chest X-rays, which can assist in
the early identification and treatment of respiratory
disorders, can be much improved by employing
transfer learning techniques in conjunction with
bespoke lung segmentation.
References:
[1] Hunton, Ryan. “Updated concepts in the
diagnosis and management of community-
acquired pneumonia.” JAAPA: Official
Journal of the American Academy of
Physician Assistants, vol. 32, no. 10, pp. 18-
23, 2019, doi:
10.1097/01.JAA.0000580528.33851.0c.
[2] S. Serte and A. Serener, "Discerning COVID-
19 from mycoplasma and viral pneumonia on
CT images via deep learning," 2020 4th
International Symposium on Multidisciplinary
Studies and Innovative Technologies
(ISMSIT), Istanbul, Turkey, pp. 1-5, 2020,
doi: 10.1109/ISMSIT50672.2020.9254970.
[3] Popovsky, Erica Y., and Todd A. Florin.
“Community-Acquired Pneumonia in
Childhood.” Encyclopedia of Respiratory
Medicine, pp. 119–131, 2022, doi:
10.1016/B978-0-08-102723-3.00013-5.
[4] Abdullahi Umar Ibrahim, Mehmet Ozsoz,
Sertan Serte, Fadi M. Al-turjman, and
Polycarp Shizawaliyi Yakoi. “Pneumonia
Classification Using Deep Learning from
Chest X-ray Images During COVID-
19.” Cognitive computation, pp. 1-13, 4 Jan.
2021, doi: 10.1007/s12559-020-09787-5.
[5] Islam, N., Salameh, J., Leeflang, M.M.,
Hooft, L., McGrath, T.A., van der Pol, C.B.,
Frank, R.A., Kazi, S., Prager, R., Hare, S.S.,
Dennie, C., Spijker, R., Deeks, J.J., Dinnes, J.,
Jenniskens, K., Korevaar, D.A., Cohen, J.F.,
Van den Bruel, A., Takwoingi, Y., van de
Wijgert, J.H., Wang, J., & McInnes, M.D.
“Thoracic imaging tests for the diagnosis of
COVID-19.” The Cochrane database of
systematic reviews, vol. 3, no. 3, p.
CD013639. 16 Mar. 2021, doi:
10.1002/14651858.CD013639.pub4
[6] Tiwari, Alok, Taresh Sarvesh Sharan, Shiru
Sharma, and Neeraj Sharma. "Deep learning-
based automated multiclass classification of
chest x-rays into covid-19, normal, bacterial
pneumonia and viral pneumonia". Cogent
Engineering, vol. 9, no. 1, 2022,
https://doi.org/10.1080/23311916.2022.21055
59.
[7] A. Sharma, D. Raju and S. Ranjan, "Detection
of pneumonia clouds in chest X-ray using
image processing approach," 2017 Nirma
University International Conference on
Engineering (NUiCONE), Ahmedabad, India,
pp. 1-4, 2017, doi:
10.1109/NUICONE.2017.8325607.
[8] Verma, Dhirendra Kumar, Gaurav Saxena,
Amit Paraye, Alpana Rajan, Anil Rawat, and
Rajesh Kumar Verma. “Classifying COVID-
19 and Viral Pneumonia Lung Infections
through Deep Convolutional Neural Network
Model using Chest X-Ray Images.” Journal
of medical physics, vol. 47, no. 1, pp. 57-64,
2022, doi: 10.4103/jmp.jmp_100_21.
[9] Chowdhury, M.E., Rahman, T., Khandakar,
A., Mazhar, R., Kadir, M.A., Mahbub, Z.B.,
Islam, K.R., Khan, M.S., Iqbal, A., Al Emadi,
N. and Reaz, M.B.I. "Can AI Help in
Screening Viral and COVID-19 Pneumonia?,"
in IEEE Access, vol. 8, pp. 132665-132676,
2020, doi: 10.1109/ACCESS.2020.3010287.
[10] Rachna Jain, Preeti Nagrath, Gaurav Kataria,
V. Sirish Kaushik, D. Jude Hemanth,
“Pneumonia detection in chest X-ray images
using convolutional neural networks and
transfer learning". Measurement, Vol. 165, p.
108046, 2020, doi:
10.1016/j.measurement.2020.108046.
[11] Sait, Unais; Lal KV, Gokul; Prakash
Prajapati, Sunny; Bhaumik, Rahul; Kumar,
Tarun; Shivakumar, Sanjana; Bhalla , Kriti,
“Curated Dataset for COVID-19 Posterior-
Anterior Chest Radiography Images (X-
Rays).”, Mendeley Data, V3, 2020, doi:
10.17632/9xkhgts2s6.1.
[12] H. T. Thanh, P. H. Yen and T. B. Ngoc,
"Pneumonia Classification in X-ray Images
Using Artificial Intelligence
Technology," 2020 Applying New Technology
in Green Buildings (ATiGB), Da Nang,
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2024.12.32
Advait K Asok, Lidiya Lilly Thampi
E-ISSN: 2415-1521
334
Volume 12, 2024
Vietnam, pp. 25-30, 2021, doi:
10.1109/ATiGB50996.2021.9423017.
[13] S. Walvekar and S. Shinde, "Efficient Medical
Image Segmentation Of COVID-19 Chest CT
Images Based on Deep Learning
Techniques," 2021 International Conference
on Emerging Smart Computing and
Informatics (ESCI), Pune, India, 2021, pp.
203-206, doi:
10.1109/ESCI50559.2021.9397043.
[14] R. I. Bendjillali, M. Beladgham, K. Merit, A.
Taleb-Ahmed, “Illumination-robust face
recognition based on deep convolutional
neural networks architectures,” Indonesian
Journal of Electrical Engineering and
Computer Science, vol. 18, no. 2, May 2020,
pp. 1015–1027, doi:
10.11591/ijeecs.v18.i2.pp1015-1027.
[15] B. K. Umri, M. Wafa Akhyari and K. Kusrini,
"Detection of Covid-19 in Chest X-ray Image
using CLAHE and Convolutional Neural
Network," 2020 2nd International Conference
on Cybernetics and Intelligent System
(ICORIS), Manado, Indonesia, pp. 1-5, 2020,
PMID: 11701515, doi:
10.1109/ICORIS50180.2020.9320806.
[16] A. Mordvintsev, K. Abid, “Opencv-python
tutorials documentation,” 2017, [Online].
https://opencv24-python-
tutorials.readthedocs.io/_/downloads/en/stable
/pdf/ (Accessed date: April 12, 2022).
[17] Konkol, Marek & Śniatała, Konrad &
Śniatała, Paweł & Wilk, Szymon &
Baczyńska, Beata & Milecki, Piotr,
"Computer Tools to Analyze Lung CT
Changes after Radiotherapy". Applied
Sciences, vol. 11, no. 04, MDPI, p. 1582,
2021, doi: 10.3390/app11041582.
[18] Pham, Dzung L., Chenyang Xu, and Jerry L.
Prince. “Current methods in medical image
segmentation.” Annual Review of Biomedical
Engineering, vol. 2, no. 1, pp. 315-337, 2000,
doi: 10.1146/annurev.bioeng.2.1.315.
[19] M. M. Hasan, M. Md. Jahangir Kabir, M. R.
Haque and M. Ahmed, "A Combined
Approach Using Image Processing and Deep
Learning to Detect Pneumonia from Chest X-
Ray Image," 2019 3rd International
Conference on Electrical, Computer &
Telecommunication Engineering (ICECTE),
Rajshahi, Bangladesh, December 26, 2019,
pp. 89-92, doi:
10.1109/ICECTE48615.2019.9303543.
[20] M. N. Saad, Z. Muda, N. S. Ashaari and H. A.
Hamid, "Image segmentation for lung region
in chest X-ray images using edge detection
and morphology," 2014 IEEE International
Conference on Control System, Computing
and Engineering (ICCSCE 2014), Penang,
Malaysia, 2014, pp. 46-51, doi:
10.1109/ICCSCE.2014.7072687.
[21] Pramanik, R., Sarkar, S. and Sarkar, R., “An
adaptive and altruistic PSO-based deep
feature selection method for Pneumonia
detection from Chest X-rays.” Applied Soft
Computing, vol. 128, 2022, p. 109464, doi:
10.1016/j.asoc.2022.109464
[22] Wang, L., Zhang, Y., Wang, D., Tong, X.,
Liu, T., Zhang, S., Huang, J., Zhang, L.,
Chen, L., Fan, H. and Clarke, M., “Artificial
Intelligence for COVID-19: A Systematic
Review.” Frontiers in medicine, vol. 8, p.
704256. 30 Sep. 2021, PMID: 34660623,
PMCID: PMC8514781, doi:
10.3389/fmed.2021.704256.
[23] Talukder, M.A., Layek, M.A., Kazi, M.,
Uddin, M.A. and Aryal, S., “Empowering
COVID-19 detection: Optimizing
performance through fine-tuned EfficientNet
deep learning architecture.” Computers in
Biology and Medicine, vol. 168, 2024, p.
107789, doi:
10.1016/j.compbiomed.2023.107789.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed to the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2024.12.32
Advait K Asok, Lidiya Lilly Thampi
E-ISSN: 2415-1521
335
Volume 12, 2024