SARS-CoV-2 or
COVID-19 is the worldwide epidemic caused by Coronavirus illness in 2019 caused
severe acute respiratory syndrome (SARS) and was found for the first time in
Wuhan, which is located in China. In March 2020, the
World Health Organization (WHO) gave the disease the name COVID-19 and
predicted that it would spread over the world as an epidemic (Cantarero et al. 2021;De Anda-Suarez et al. 2022). COVID-19 negatively
affects the health of people, especially the elderly, people with existing
medical issues, and people whose immune systems are already impaired. By the
middle of July 2020, more than 13 million individuals had contracted COVID-19,
and more than 570,000 people had already died from the epidemic (Sakib et al. 2020). COVID-19's most pervasive clinical characteristics are
a dry cough, fever, and tiredness
(Yan et al. 2021). patients often have the most severe symptoms of
pneumonia after they have been infected with the coronavirus (Rahman, Khandakar, and Member 2021). Despite this, it is possible that some
COVID-19 cases may exhibit no symptoms at all. In severe circumstances,
intensive care unit treatment or oxygen support therapy may be required (Munusamy et al. 2021).
"Reverse
Transcriptase Polymerase Chain Reaction (RT-PCR)" real-time testing was
utilized as a diagnostic method to halt the COVID-19 illness from spreading
after the genome of SARS-CoV-2 was sequenced (Mondal et al. 2022).
The present
pandemic state does not support the low-sensitivity RT-PCR test because RT-PCR
takes considerable time and has a significant number of false-negative results (Ibrahim and Mahmood 2023). A test result that is falsely negative can
result in an illness spreading extensively (Tabik et al. 2020). Sometimes, infected individuals may not be identified
right away and may not receive the proper care (Shah et al. 2021). Consequently, diagnostic imaging, especially chest
X-rays and chest computerized tomographic (CT) (Dai et al. 2020), is frequently used in combination with other tests to
aid in earlier COVID-19 screening and treatment (Tai et al. 2021).
In light of
this, lesions with ground-glass opacities (GGO) and pulmonary fibrosis are
COVID-19 lung imaging features (Dong et al. 2021). CT scans are extremely sensitive, so they help assess
the severity of the illness; however, their main limitation is the time
required for diagnosis; even radiology specialists require approximately 21.5
minutes to parse each case's test results (Wu et al. 2021). For these reasons, automated
diagnosis technologies, such as models of deep
learning (DL), have been created to screen for COVID-19 in CT images recently
(He et al. 2020). According to several recent studies, CNNs are a
particular kind of DL technique (Hussain et al. 2021), having a high capability for COVID-19 identification (Shi et al. 2021). With an acceptable performance prediction using
COVID-19-infected patients' chest CT scan images (Mohammed et al. 2020). These techniques can be applied by fine-tuning the
weights or knowledge gained from the pre-trained CNN model on the large dataset
and adapting it to perform well on specific tasks with limited data and achieve
better results compared to training from scratch (Shukla et al. 2020). As in the dataset of COVID-19 for the diagnosis of
COVID-19 from CT-scan chest images (Sarker et al. 2020). A densely connected
convolutional network (DenseNet), a specialized CNN technique, has the
advantages of mitigating the problem of vanishing gradients, improving feature
reuse, and decreasing parameter usage. Despite these advantages, it has some
limitations, such as computational complexity and the substantial amount of GPU
memory; furthermore, DenseNet-121 is prone to overfitting (Albelwi 2022).
Therefore,
the main focus of this study is on using pre-trained DenseNet-121 models of DL
to automatically identify and diagnose COVID-19 using chest CT-scan data.
The following
is a list of the important contributions that this article has made:
1.
The
DenseNet121-based Deep Transfer Learning (DTL) model was modified by the
exclusion of the last layer (i.e., the top) and the addition of three dense
layers, including Dense128, Dense64, and Dense2, to the model to enhance the
model and achieve high accuracy.
2.
The model
is used as an approach for automatically detecting and diagnosing COVID-19
patients on lung CT-scan images.
3.
To
ensure a precise and effective evaluation of the model, the image datasets go
through a pre-processing stage.
4.
Employing
a modified model on two CT-scan image datasets offers a faster alternative to
the widely used RT-PCR testing kit.
The remaining
parts of this article have been organized into the following sections and
order: The relevant works are listed in section 2. The suggested DTL model and
data description are described in Section 3. In section 4 the findings and
their analysis are discussed. Finally, section 5 presents the points of
conclusion.
Various study
projects have been carried out extensively and fast in order to build AI
strategies for combating the COVID-19 worldwide pandemic. Some relevant
research on using a chest CT scan to diagnose and classify COVID-19 is
reviewed.
Wang, Liu, and Dou (2020) suggested a new joint learning framework to
use different datasets to detect COVID-19 disease. The network has been split
into two parts. The first part has a lightweight structure and four separate
layers of convolution. The second part has more densely packed learning blocks
that connect. The COVID-CT dataset and SARS-CoV-2 were used in this work. The
work results show that the SARS-CoV-2 dataset outperformed the COVID-CT dataset
with an accuracy of 90%, 95% precision, 85% recall, and a 90% F1-score.
Panwar et al. (2020) proposed the VGG model as a type
of DTL algorithm to classify COVID-19 positive or negative cases, that utilized
three distinct datasets: SARS-COV-2 CT-scan, COVID-chest X-ray, and Chest X-ray
radiography. The grad-CAM-based colour representation technique was used by the
researchers to make the suggested model more comprehensible and interpretable.
Achieved an accuracy of 95%, a sensitivity of 94.4%, and a specificity of
95.84.
Silva et al. (2020) presented a model called Efficient CovidNet applied to
lung CT-scan radiography to identify COVID-19 patients, which uses voting
techniques and cross-dataset analytics. Experimental results obtained from
SARS-COV-2 CT-Scan datasets show a sensitivity of 98.80%, a precision of 99.2%,
and an accuracy of 98.99%.
Seum et al. (2020) compared 12 pre-trained CNNs models in classifying
COVID-19 patients by utilizing an available dataset (the SARS-COV-2 CT-Scan). This
work was carried out in two phases: one without segmentation and another with
segmentation. The study results showed in the first phase that the performance
of the DenseNet169 model is higher than other models in terms of F1-scores, and
accuracy with 88.60% and 89.31%, respectively. In the second phase, DenseNet201
and ResNet18 both obtained an accuracy of 89.92%, implying that using
classification after segmentation improves the efficiency of a classification
model.
Ramzy, Sherin, and Karma (2021) suggested (COV-CAF), which is a
computer-aided framework to identify COVID-19 cases through 3D CT-scan pictures.
The researchers combined both DL and traditional techniques in COV-CAF, which
has two stages: one for preparatory work and the second one for extracting and
classifying. The first stage is focused on 3D CT volumes and proposes an
effective cut selection method for selecting useful CT slices, and the second
stage uses a recently developed CNN called Norm-VGG16 to extract features
automatically. They employed two datasets, the SARS-CoV-2 CT-scan and
MosMedData. The findings obtained on the SARS-COV-2 CT scan were 97.59%
accuracy and 98.41% sensitivity; whereas, it gained an overall accuracy of
97.76% and an average sensitivity of 96.73% on the MosMedData dataset.
Biswas et al. (2021) conducted a study to use effective transfer learning
(TL) algorithms to construct an accurate COVID-19 prediction model on
SARS-CoV-2 CT datasets. Initially, they utilized three kinds of (DL) models
separately: ResNet50, Xception, and VGG-16, and then combined these three
models to increase the model's prediction accuracy. The suggested model has a
high F1-score of 99% and an accuracy of 98.79%.
Jaiswal et al. (2021) developed a DTL model to detect COVID-19 infection. It
is a DenseNet201 – based model, which was trained before with a massive dataset
known as the ImageNet dataset. The provided method produced results with an
accuracy of 0.9625, 96.2% recall, 96.2% precision, and a specificity of 96.21%.
Hasan et al. (2021), using CNN, suggested a method to find COVID-19 persons
via images from lung scanning. They use the newest updated CNN model (DenseNet-121).
The findings showed a satisfactory level of performance for the COVID-19
prediction, with an accuracy of 92% and a recall of 95%.
Horry et al. (2020) built a mechanism for COVID-19 detection based on the DL
by utilizing the VGG19 algorithm on various kinds of medical photographs (CT,
x-ray, ultrasound), achieving a precision of 84% for CT scans, 86% for x-rays,
and 100% for ultrasounds, and this system was able to identify COVID-19 in 100%
of the ultrasound images.
Islam and Matin (2020) employed LeNet-5 CNN networks after a simple CNN model
to predict COVID-19 from thoracic images. For training purposes, the obtained
results were an accuracy of 86.06%, an F1-score of 87%, 89% recall, 85%
precision, and a 0.86 ROC.
Pham (2020) showed results from 16 pre-trained CNN models of
COVID-19 classification. According to this study, TL led to higher
classification rates than data augmentation. ResNet-18 had the best sensitivity
(98%), MobileNet-v2 earned the maximum accuracy (95%), and F1-score (96%), and
DenseNet-201 had the highest degree of specificity (96%). This research
utilized various networks, including ShuffleNet, ResNet-18, DenseNet-201, and
MobileNet-v2.
Garain et al. (2021), for COVID-19 screening, created a three-layer DCSNN
using CT images. The proposed method obtained a 99% F1 score for the
potential-based model. For chest CT images, the suggested SNN-based model outperforms
CNN models,
but it takes longer to train. Furthermore, in comparison to older DL models,
the current method is more efficient.
Chen et al. (2021) conducted a study on automatically identifying positive
COVID-19 patients using chest CT scan radiography, a prototype network that has
been pre-trained by a momentum-contrasting learning strategy was applied. They
obtained values of 87% of accuracy, 93.2% of AUC, 88.5% of precision, and
recall 87.4%.
Huang and Liao (2022) proposed LightEfficientNetV2
model and fine-tunes seven CNN techniques involving EfficientNet-B0,
InceptionV3, Xception, MobileNetV2, DenseNet121, EfficientNetV2, and ResNet50V2
to diagnose COVID-19 cases on three different datasets (NIH Chest X-rays, SARS-CoV-2,
and COVID-CT). The proposed LightEfficientNetV2 model obtained an accuracy of
99.47% on the SARS-CoV-2 dataset and 88.67% on COVID-CT.
Kathamuthu et al. (2023) used
several deep transfer learning-based CNN approaches including InceptionV3,
Resnet50, Xception, VGG16, VGG19, and Densenet121 to screen COVID-19 in chest
CT images and achieved an accuracy of 98% with VGG16 model on SARS-CoV-2
CT-scan dataset.
Methodology
In light of
the limited number of COVID-19 radiography pictures in our dataset, it is
difficult to develop a DL model in this situation because DL approaches need a
huge dataset to start from scratch. Transfer learning (TL) could be utilized as
an effective solution to this issue because TL techniques use the knowledge
gained in one field to carry out tasks in other fields. For instance, when
initializing DL models with weights obtained from the ImageNet (Fei-Fei, Deng, and Li 2010), as the ImageNet is a massive dataset
comprised of 14 million images from various resources, by that time it has
learned several features for diverse images, so these newly acquired features
can then be applied to another dataset. For this reason, the weights gained
from pre-trained DenseNet-121 on the ImageNet dataset were employed for the
SARS-COV2 and the COVID19-CT datasets in this study.
3.1 Study Design
The study system's processes start with two distinct
datasets that go through data pre-processing techniques before being split into
training, validation, and test sets. Hence, the data are ready to fit into a
modified model. Figure 1 shows the general study diagram.
3.2 Datasets Description
The two datasets that were utilized in this paper are demonstrated in this
section. It provides details about the source, size, nature, characteristics,
and other information of the datasets.
3.2.1
SARS-CoV-2
CT-scan dataset: This dataset
consists of a total of 2481 CT scans from 120 patients, including 1252 CT scan
images of 60 infected persons (COVID-19), 28 females with 32 males, and 1229 CT
scan images of 60 uninfected persons (non-COVID), females (30) and males (30) (Soares and Angelov 2020). Figure 2 shows samples from each class
(non-COVID and COVID-19). These data were obtained from patients in Sao Paulo
hospitals, Brazil. There are no standard dimensions for CT-scan images; in
general, their sizes are between 104 × 153 and 484 × 416. Figure 3 shows the
different sizes between them. The protocol offered for evaluating models is
split randomly into training, validation, and testing at 80, 10, and 10
percent, respectively, by the train_test_split () method. The distribution of
the dataset is shown in Table 1. The dataset can be found at: www.kaggle.com/plameneduardo/sarscov2-ctscan-dataset.
Figure 2: Non-COVID (left) and COVID-19 (right).
Figure 3: CT-scan
images of different sizes.
3.2.2
COVID19-CT
dataset: The second dataset comprises 216 instances of COVID-19 and 171
cases of non-COVID, with a collection of 746 CT scans. Among the CT scans,
there are 349 images of instances of COVID-19 with 397 samples of non-COVID-19,
including several other disorders of the lungs (He et al. 2020). Positive CT scans had been sourced from preprints on
COVID-19 available on bioRxiv and medRxiv, displaying a range of symptoms. As
the CT scans were obtained from various places, their dimensions varied from
124 × 153 to 1485 × 1853, causing inconsistencies in image contrast, as shown
in Figure 4. The pattern of the spreading of the
dataset is displayed in Table 1. The data are available at : https://www.kaggle.com/datasets/luisblanche/covidct?resource=download&select=CT_NonCOVID. 
Figure 4:
The differences in contrast CT scan images.
Table 1: The
pattern of distribution of the SARS-COV-2 CT-scan and COVID19-CT datasets.
|
Dataset
|
SARS-COV-2
Ct-Scan
|
COVID19-CT
|
|
#
Patients
|
120
|
387
|
|
#
CT-scan images
|
2481
|
746
|
|
Train
sets images
|
1983
|
596
|
|
Valid
sets images
|
249
|
75
|
|
Test
sets images
|
249
|
75
|
3.3 Data Pre-processing
The term used
to describe all the changes made to initial data before feeding it into machine
learning (ML) or DL algorithms named pre-processing. All images are resized
from different resolutions to 224 x 224 resolution using the OpenCV library
interpolation function INTER_AREA. Additionally, the pixels' values might range
from 0 to 255. If a DL model is applied to the image as-is, computing high
numeric values may become more challenging. We can minimize this by adjusting
the values to be
between 0 and 1,
also known as normalization, which is a significant stage to guarantee that
each input parameter (in this case, a pixel) has a uniform data distribution.
This makes convergence faster while training the network.
In data
analysis, data augmentation is utilized to raise the number of training data
sets either by producing drastically altered copies of already existing data or
by developing synthetic data from earlier existing data. This is accomplished
by one of these two methods. This technique enhances the generalization
performance and classification accuracy of the model while also serving as a regularise
to reduce overfitting (Tang, Yuan, and Zhu 2020). Table 2 lists the parameters used for image augmentation in the
current study.
3.5 The Modified DenseNet121
In this study, a
DenseNet121-based Deep Transfer Learning model-COVID-19 is modified and called
(DDTL-COV) to classify COVID-19-infected patients from chest CT-scan images.
The DDTL-COV was trained using TL, with the pre-trained.
DenseNet121 weights on
the ImageNet dataset used as the initial weights. Following this, the model was
fine-tuned on the COVID-19 dataset. Figure 5 shows the architecture of the
DDTL-COV model, which employs DenseNet121 for
accurately classifying COVID-19 patients.
The main parts of the
DDTL-COV model are the feature extraction and classification parts.
In the
feature extraction part, the input layer of the CNN is defined by the shape of
the input image (224,224,3) followed by a convolutional layer of 3 kernels with
size (3 x 3) with padding 'same'. The DDTL-COV model was trained using TL with
the pre-trained DenseNet121 weights on the ImageNet dataset.
Table 2: Parameters for image augmentation
|
Parameters
|
values
|
Functions
|
|
Rotation-range
|
360
|
The degree range of rotations randomly
|
|
Width-shift-range
|
0.2
|
The range of shifting horizontally is
randomly selected.
|
|
Height-shift-range
|
0.2
|
The range of shifting vertically is
randomly selected.
|
|
Horizontal-flip
|
True
|
The input randomly flips in the horizontal
direction.
|
|
Vertical-flip
|
True
|
The input randomly flips in the
vertical direction.
|
|
Zoom range
|
0.2
|
The randomized zooming range.
|
DenseNet
is a logical extension of ResNet and boosts performance by merging the feature
maps at each layer with those of the layer that came before it inside a dense
block. This permits the layers of the network that come after them to directly
use the features of the layers that came before them, which promotes the reuse
of features across the network as shown in Figure 6, which facilitates solving
the vanishing gradient issue and also minimizes the parameters' number.
Depending on
the network version, the DenseNet network is structured into four blocks of
layers, each of which has a particular number of layers. For instance, the
DenseNe-169 has 4 blocks with 6, 12, 32, and 32 layers, respectively, and
DenseNe-121 was used in this study; it is shown in Error! Reference source not found. 7 as having four blocks with a total of 6, 12, 24, and
16 layers, respectively.
Figure 5: A dense block with five layers that shows
the direct connections between layers.
Figure 6: DenseNet121's layered architectural
design.
Layers for Batch
Normalization, Relu functions, consonant layers, and Pooling layers are put
between these blocks.
The DenseNet121
architecture implements direct interconnections between all antecedent and
subsequent layers in furtherance of connectivity.
The feature concatenation's
mathematical formulation is as follows:
(1)
Here, 
is a
transition that is not linear. It consists of BN and ReLU (Yu et al. 2020). And
mathematically can be represented as in Eq (2)
(2)
And followed by a convolution layer of (3 x 3) windows,
is the
feature map string for layers 0 to l-1 that has been concatenated into one
tensor for simplicity of use, for down sampling. In the model architecture,
transition layers made up of a BN layer are used to build and separate dense
blocks, a 2×2 average pooling layer at the end, and a (1×1) convolution layer
between them.
The growth rate,
represented by the hyperparameter k in DenseNet121, indicates how the dense
design achieves superior outcomes. The architecture of DenseNet121 evolves by
observing feature maps as the network's global state. Therefore, despite having a reduced growth rate, it
performs efficiently. Every succeeding layer
of the network is able to access all of the feature maps that were generated by
the layer that came before it. Additionally,
each layer adds k feature maps to the overall state, resulting in a total input
feature map count across the lth layer known as FM:
(3)
Here, k0 is
the symbol for the channels on the input layer. To improve the efficiency
of computation, an (1x1) convolution layer is added before every (3x3)
convolution, decreasing the number of inputs for the feature map, which is typically more than the number of
output feature maps k. The bottleneck layer,
which is a (1x1) convolution layer, is inserted and creates 4k feature maps.
In the classification part, the fully
connected layer at the end pre-trained DenseNet121model (i.e., the top) is
excluded (include_top = False), an additional three dense layers including
(Dense128_Relu_BN_Dropout 0.3), (Dense64_Relu_BN_Dropout 0.25), and last dense
(2) is added to the model after GobalAveragePooling2D layer and BN, hence,
Dense () indicates the layer's neurons, BN represents the Batch Normalization
layer, Relu is the activation function, Dropout () is used to regularize CNNs
in addition, it overcomes the over-fitting issue. For binary classification, the
last Dense layer along the sigmoid is utilized as the output layer since the
sigmoid is an activation function that turns non-normalized outputs into binary
0s and 1s, where (0 = COVID-19, 1= Non_COVID).
The sigmoid function
can be expressed using the formula below:
The history of the model is
presented in Table 3.
Table 3: Model
history.
|
Layer(type)
|
Output shape
|
Parameter
#
|
|
Input layer
|
(None,224,224,3)
|
0
|
|
Conv2D
|
(None,224,224,3)
|
84
|
|
Densenet121 (Functional)
|
(None, None, None, 1024)
|
7037504
|
|
GlobalAveragePooling2d
|
(None, 1024)
|
0
|
|
Batch normalization
|
(None, 1024)
|
4096
|
|
Dense
|
(None, 128)
|
131200
|
|
Batch normalization_1
|
(None, 128)
|
512
|
|
Dropout
|
(None, 128)
|
0
|
|
Dense 1
|
(None, 64)
|
8256
|
|
Batch normalization_2
|
(None, 64)
|
256
|
|
Dropout 1
|
(None, 64)
|
0
|
|
Dense 2
|
(None, 2)
|
130
|
|
Total parameters: 7,182,038
Trainable parameters: 7,095,958
Non-trainable parameters: 86,080
|
3.6 Hyperparameters
The model
completed training around 20 minutes with 75 epochs and a batch size of 32. The
binary cross-entropy is utilized as a loss function that is capable of being
lowered using the backpropagation approach while the model is being trained.
The following equation demonstrates the loss function:
Here, the number of training samples is
represented by n, the number of classes by k, the model parameter is θ,
the actual level of the ith training sample is 
, and the output for the ith training
sample is 
at
the kth node.
The model is optimized
by using the Adam optimizer, which is a subset of the stochastic gradient
descent approach that is based on adaptive estimation of the first moment
estimate 
,
second-moment estimate 
, with
α = 0.003 as the initial learning rate. updating the model weight θ
with Adam's optimizer is illustrated in the below equation:
(6)
Where
(7)
(8)
Here, 
represents
the weights of the gradient concerning the loss function, the weight at time
t−1 is indicated by θt−1, the weight
at time t is represented by θt, β1, and β2
are hyperparameters, the default value of β1 = 0.9 and β2
= 0.999 were used, With the ReduceLROnPlateau callback, the learning rate is
decreased by a factor of 0.70 until it reaches a lower limit (min lr = 0.0001)
when the validation loss remains constant for about 5 epochs (patience = 5).
3
This section demonstrates
the efficacy of the suggested model in identifying if a person has COVID-19
illness or not by using the datasets SARS-CoV-2 CT-scan and COVID19-CT. Initially, we analyzed SARS-CoV-2 CT results then
COVID19-CT and compared them. Additionally, each dataset's results were then
compared with those from earlier studies and research using the same datasets.
The pretraining model that is used in this study is DenseNet-121.
Figure
8: The loss of the DDTL-COV model
throughout training and validation across the course of 75 epochs
.
4.2 Experimental setup
In this study, a Windows
10 Pro 64-bit operating system has been used with the Anaconda compiler
supported by Python 3.7 and the TensorFlow library and Keras library, which
both support GPU hardware acceleration on some COVID-19 CT scans for DL and CNN
purposes, probably for achieving perfect infection detection. And, as for
hardware, the following computer and key specifications were used: I5-10400F
(10th Generation Intel Core i5 Desktop Processor with 6 cores and 12 threads),
GeForce RTX 2070 with 8GB of vRAM, and 16 GB of DDR4 at 2666 MHz (RAM/Memory).
4.3 The results of the
DDTL-COV model from the SARS-CoV-2 CT dataset
Three sets
are created from the dataset for the performance analysis. 80% of the data used
to train the model is in the training
set. 10% of the data
makes up the validation set, which is used to fine-tune the model's
hyperparameters and guard against overfitting. The remaining 10% of the data
makes up the testing set, which is used to assess the model's ultimate performance.
The loss of training and validation decreases gradually through each epoch,
indicating that the model is learning and generalizing well. After epoch 55,
the validation loss stops decreasing, which means that the model has reached
its optimal performance. It is crucial to remember that the ideal number of
epochs may change based on the model's complexity and the quantity of the
dataset. The training and validation accuracy of the DDTL-COV model across
various epochs are arranged in Figure 8, and the loss analyses of the model
over the same number of epochs is illustrated in Figure 9.
Figure 10 shows a confusion matrix that is used to
summarize the results of samples of each class that were successfully and
mistakenly categorized. Based on the confusion matrix, the performance of the
model was evaluated using four metrics: precision, recall, F1 score, and
accuracy. The DDTL-COV model only produced one false positive case and no false
negative cases, indicating a high score of accuracy of 99.6% and an F1-score of
100%. So, it will be of significance in the medical research field, as shown in
Table 4.
Table 4: DDTL-COV
model results from the SARS-CoV-2 CT dataset.
|
Class\metrics
|
precision
|
recall
|
F1-score
|
|
Class 0\COVID-19
|
100%
|
99%
|
100%
|
|
Class 1\non-COVID
|
99%
|
100%
|
100%
|
|
Accuracy overall 100%
|
|
Macro avg
|
100%
|
100%
|
100%
|
|
Weighted avg
|
100%
|
100%
|
100%
|
4.3.1
Comparison
of Results with previous studies
In this
research, several innovative approaches developed for COVID-19 identification
using the SARS-CoV-2 dataset are compared, as shown in Table 5. Some studies
yielded fewer results, such as Panwar et al. (2020), which used the VGG model to obtain an accuracy of 95%
and a recall of 94.4%. Wang, Liu, and Dou (2020) used the CNN model and achieved 95% accuracy
and a 94.4% of F1-scor. As for Ramzy, Sherin, and Karma (2021), with VGG16, accuracy is 97.59%, recall is
98.41%, and so on. On the other hand, some other studies introduced various DL
models and got superior results, as Silva et al. (2020) were capable to obtain an accuracy of 98.99%, a recall
of 98.80%, and a precision of 99.20% through the utilization of Efficient
CovidNet, and Biswas et al. (2021), by merging three previously trained CNNs of models
(Xception, ResNet50, and VGG-16), were able to get an F1-score of 99% and an
accuracy of 98.79%. The overall performance of the
DDTL-COV model is superior to all other studies.
Table 5:
A comparison of the DDTL-COV model with the modern techniques using the
SARS-CoV-2 CT dataset (Acc: accuracy, Pre: precision, Rec: recall, F1-S:
F1-score).
|
Ref.
|
Method
|
Acc.
%
|
Pre.
%
|
Rec.
%
|
F1-S
%
|
|
(Wang, Liu, and Dou 2020)
|
CNN
|
90
|
95
|
85
|
90
|
|
(Panwar et al. 2020)
|
VGG
|
95
|
/
|
94.4
|
/
|
|
(Silva et al. 2020)
|
Efficient
CovidNet
|
98.99
|
99.20
|
98.80
|
/
|
|
(Seum et al. 2020)
|
12
pre-training CNNs
|
89.92
|
99.50
|
86.80
|
89.67
|
|
(Ramzy, Sherin, and Karma
2021)
|
VGG16
|
97.59
|
/
|
98.41
|
/
|
|
(Biswas et al. 2021)
|
Combine
3 pre-trained CNNs models (VGG-16, ResNet50, Xception)
|
98.79
|
/
|
/
|
99
|
|
(Jaiswal et al. 2021)
|
DenseNet201
|
96.21
|
96.25
|
96.2
|
96.2
|
|
(Hasan et al. 2021)
|
DenseNet-121
|
92
|
/
|
95
|
/
|
|
(Huang and Liao 2022)
|
LightEfficientNetV2
|
99.47
|
99.19
|
99.25
|
99.04
|
|
(Kathamuthu et al. 2023)
|
VGG16
|
98
|
97.99
|
97.99
|
97.99
|
|
This
work
|
DDTL_COV
(DenseNet-121)
|
99.6
|
100
|
100
|
100
|
4.4
DDTL-COV model results from the COVID19 CT dataset
As in a previous
experiment, the COVID19 CT dataset is partitioned randomly into three distinct
sets of training, validation, and testing at 80, 10, and 10 percent, respectively.
The model is trained over 75 epochs at a speed of 5 seconds per epoch, and its
convergence is at a noticeably good speed. Figures 11 and 12 illustrate the
accuracy of training and validation, as well as the training loss, and the
validation loss of the DDTL-COV model on the COVID19-CT dataset, respectively.
Figure 10: The
DDTL-COV model accuracy across 75 epochs of training and validation from the COVID19-CT
dataset.
Figure 11: The
loss of the DDTL-COV model throughout training and validation across the course
of 75 epochs using the COVID19 CT dataset.
Common assessment criteria such
as F1-score, Recall (sensitivity), accuracy, and specificity were used to
assess the efficacy of the classification method employed in this research. The
results of the model are shown in Table 6, which finds that it has a recall of
91% and an accuracy of 89%. Furthermore, these findings show that this model
only produced one false negative case and seven false positive cases, as shown
in Figure 13.
Table 6: The
DDTL-COV model's classification from the COVID19-CT dataset.
|
Class \ Metrics
|
Precision
|
Recall
|
F1-score
|
|
Class 0\COVID-19
|
81%
|
97%
|
88%
|
|
Class
1\non-COVID
|
97%
|
84%
|
90%
|
|
Accuracy
overall 89%
|
|
Macro avg
|
89%
|
91%
|
89%
|
|
Weighted avg
|
91%
|
89%
|
89%
|
Figure 12: The
DDTL-COV model confusion matrix utilizing COVID19-CT dataset.
4.4.1
Result
in comparison with previous studies: Compared to the related works using the
dataset COVID19 CT, the provided approach obtains average accuracy and F1- scores
of 89% and 88%, respectively. The results of
some research are fewer (Islam and Matin 2020)
which used the basic CNN model
plus the LeNet-5 CNN model to achieve an accuracy of 86.06%, a precision of
85%, a recall of 89%, and an F1-score of 87%. Also, its performance is superior to Horry et al. (2020) but less than that to Garain et al. (2021), Chen et al. (2021), and Pham (2020), as shown in Table 7.
4.5 A
Comparison of results between the COVID-19-CT dataset and the SARS-CoV-2
dataset
The overall
performance of the COVID19-CT dataset in the manner of accuracy in comparison
to that of the SARS-CoV-2
dataset is less as shown in Table 8. This is explained by the cross-source
variability of the dataset's CT scans. The non-COVID-19 CT images were
obtained from various sources and display a variety of results, making it
challenging to differentiate between COVID-19 and other findings associated
with lung disorders due to the possibility of visual manifestations
overlapping. Another explanation is that the COVID19-CT dataset's CT images
exhibit a significant contrast variation, varying spatial resolution, and other
visual traits that can impair the model's capacity to extract more
discriminative and generalizable features.
Table 7: A
comparison of the DDTL-COV model with the modern techniques using the COVID19-CT dataset (Acc: accuracy, Pre:
precision, Rec: recall, F1-S: F1-score).).
|
Ref.
|
Method
|
Acc.
%
|
Pre.
%
|
Rec.
%
|
F1-S
%
|
|
(Horry et al. 2020)
|
VGG19
|
/
|
84
|
/
|
78
|
|
(Islam and Matin 2020)
|
CNN
+ LeNet-5
|
86.06
|
85
|
89
|
87
|
|
(Pham 2020)
|
Sixteen
pre-trained CNNs
|
95
%
(MobileNet-v2)
|
/
|
98
(ResNet-18)
|
96
(MobileNet-v2)
|
|
(Garain et al. 2021)
|
Three-layer
DCSNN
|
/
|
/
|
/
|
99
|
|
(Chen et al. 2021)
|
prototypical
network
|
93.2
|
88.5
|
87.4
|
/
|
|
(Huang and Liao 2022)
|
LightEfficientNetV2
|
88.67
|
87.28
|
87.43
|
87.55
|
|
This
work
|
DDTL_COV
(DenseNet-121)
|
89
|
89
|
91
|
89
|
Table 8: Accuracy of the DDT-COV model for two datasets
on test data.
|
Datasets
|
Accuracy
|
|
SARS-CoV-2 CT
|
99.6%
|
|
COVID19-
CT
|
89%
|
This study suggests the DDTL-COV
model, a deep transfer learning model based on DenseNet121, using weights from
the ImageNet dataset to categorize patients on CT scans as either COVID or
non-COVID. The SARS-CoV-2 CT scan and COVID19-CT datasets were utilized to
assess the model, with image enhancement in pre-processing using the OpenCV
library interpolation function INTER_AREA and data augmentation techniques. The
suggested model achieves outstanding performance in the SARS-CoV-2 CT-scan
dataset, with an accuracy of 99.6%. However, on the COVID19-CT dataset, it
performed with an accuracy rate of 89%. The limitation of this model, its
performance is not good on the COVID19-CT dataset.
The future study will design a
structure combining DL models that have already been pre-trained. Additionally,
the technique of k-fold cross-validation will be used as a partitioning
strategy for the dataset
References
Albelwi, Saleh A. 2022. “Deep Architecture Based on DenseNet-121
Model for Weather Image Recognition.” International Journal of Advanced
Computer Science and Applications 13(10): 559–65.
De Anda-Suarez, Juan et al. 2022. “A Novel Metaheuristic
Framework Based on the Generalized Boltzmann Distribution for COVID-19 Spread
Characterization.” IEEE Access 10: 7326–40.
Biswas, Shreya et al. 2021. “Prediction of Covid-19 from
Chest Ct Images Using an Ensemble of Deep Learning Models.” Applied Sciences
(Switzerland) 11(15).
Cantarero, Ruben et al. 2021. “COVID19-Routes: A Safe
Pedestrian Navigation Service.” IEEE Access 9: 93433–49.
Chen, Xiaocong et al. 2021. “Momentum Contrastive
Learning for Few-Shot COVID-19 Diagnosis from Chest CT Images.” Pattern
Recognition 113: 107826. https://doi.org/10.1016/j.patcog.2021.107826.
Dai, Hui et al. 2020. “High-Resolution Chest CT Features
and Clinical Characteristics of Patients Infected with COVID-19 in Jiangsu,
China.” International Journal of Infectious Diseases 95: 106–12.
https://doi.org/10.1016/j.ijid.2020.04.003.
Dong, Di et al. 2021. “The Role of Imaging in the
Detection and Management of COVID-19: A Review.” IEEE Reviews in Biomedical
Engineering 14: 16–29.
Fei-Fei, L., J. Deng, and K. Li. 2010. “ImageNet:
Constructing a Large-Scale Image Database.” Journal of Vision 9(8):
1037–1037.
Garain, Avishek et al. 2021. “Detection of COVID-19 from
CT Scan Images: A Spiking Neural Network-Based Approach.” Neural Computing
and Applications 1. https://doi.org/10.1007/s00521-021-05910-1.
Hasan, Najmul, Yukun Bao, Ashadullah Shawon, and Yanmei
Huang. 2021. “DenseNet Convolutional Neural Networks Application for Predicting
COVID-19 Using CT Image.” SN Computer Science 2(5): 1–11.
https://doi.org/10.1007/s42979-021-00782-7.
He, Xuehai et al. 2020. “Sample-Efficient Deep Learning
for COVID-19 Diagnosis Based on CT Scans.” IEEE Transactions on Medical
Imaging XX(Xx): 10.
Horry, Michael J. et al. 2020. “COVID-19 Detection
through Transfer Learning Using Multimodal Imaging Data.” IEEE Access 8:
149808–24.
Huang, Mei Ling, and Yu Chieh Liao. 2022. “A Lightweight
CNN-Based Network on COVID-19 Detection Using X-Ray and CT Images.” Computers
in Biology and Medicine 146(March): 105604.
https://doi.org/10.1016/j.compbiomed.2022.105604.
Hussain, Emtiaz et al. 2021. “CoroDet: A Deep Learning
Based Classification for COVID-19 Detection Using Chest X-Ray Images.” Chaos,
Solitons and Fractals 142: 110495.
https://doi.org/10.1016/j.chaos.2020.110495.
Ibrahim, Walat Ramadhan, and Mayyadah Ramiz Mahmood.
2023. “COVID-19 Detection Based on Convolution Neural Networks from CT-Scan
Images : A Review.” 29(3): 1668–77.
Islam, Md Rakibul, and Abdul Matin. 2020. “Detection of
COVID 19 from CT Image by the Novel LeNet-5 CNN Architecture.” ICCIT 2020 -
23rd International Conference on Computer and Information Technology,
Proceedings: 19–21.
Jaiswal, Aayush et al. 2021. “Classification of the
COVID-19 Infected Patients Using DenseNet201 Based Deep Transfer Learning.” Journal
of Biomolecular Structure and Dynamics 39(15): 5682–89.
https://doi.org/10.1080/07391102.2020.1788642.
Kathamuthu, Nirmala Devi et al. 2023. “A Deep Transfer
Learning-Based Convolution Neural Network Model for COVID-19 Detection Using
Computed Tomography Scan Images for Medical Applications.” Advances in
Engineering Software 175(October 2022): 103317.
https://doi.org/10.1016/j.advengsoft.2022.103317.
Mohammed, Mazin Abed et al. 2020. “Benchmarking
Methodology for Selection of Optimal COVID-19 Diagnostic Model Based on Entropy
and TOPSIS Methods.” IEEE Access 8: 99115–31.
Mondal, Arnab Kumar, Arnab Bhattacharjee, Parag Singla,
and A. P. Prathosh. 2022. “XViTCOS: Explainable Vision Transformer Based
COVID-19 Screening Using Radiography.” IEEE Journal of Translational
Engineering in Health and Medicine 10(November 2021): 1–10.
Munusamy, Hemalatha et al. 2021. “FractalCovNet
Architecture for COVID-19 Chest X-Ray Image Classification and CT-Scan Image
Segmentation.” Biocybernetics and Biomedical Engineering 41(3): 1025–38.
https://doi.org/10.1016/j.bbe.2021.06.011.
Panwar, Harsh et al. 2020. “A Deep Learning and Grad-CAM
Based Color Visualization Approach for Fast Detection of COVID-19 Cases Using
Chest X-Ray and CT-Scan Images.” Chaos, Solitons and Fractals 140:
110190. https://doi.org/10.1016/j.chaos.2020.110190.
Pham, Tuan D. 2020. “A Comprehensive Study on
Classification of COVID-19 on Computed Tomography with Pretrained Convolutional
Neural Networks.” Scientific Reports 10(1): 1–8.
https://doi.org/10.1038/s41598-020-74164-z.
Rahman, Tawsifur, Amith Khandakar, and Senior Member.
2021. “Development and Validation of an Early Scoring System for Prediction of
Disease Severity in COVID-19 Using Complete Blood Count Parameters.” IEEE
Access 9: 120422–41.
Ramzy, Mohamed, Ibrahim Sherin, and M Youssef Karma.
2021. “Abnormality Detection and Intelligent Severity Assessment of Human Chest
Computed Tomography Scans Using Deep Learning : A Case Study on SARS ‑ COV ‑ 2 Assessment.” Journal of Ambient Intelligence and Humanized
Computing (0123456789). https://doi.org/10.1007/s12652-021-03282-x.
Sakib, Sadman et al. 2020. “DL-CRC: Deep Learning-Based
Chest Radiograph Classification for Covid-19 Detection: A Novel Approach.” IEEE
Access 8(July): 171575–89.
Sarker, Laboni et al. 2020. “COVID-DenseNet: A Deep
Learning Architecture to Detect COVID-19 from Chest Radiology Images.” Preprints
(May).
https://github.com/mmiemon/COVID-DenseNet%0Ahttp://proxy.library.stonybrook.edu/login?url=https://search.proquest.com/docview/2414040424?accountid=14172%0Ahttps://www.preprints.org/manuscript/202005.0151/v1/download%0Ahttp://linksource.ebsco.com/linking.a.
Seum, Ashek, Amir Hossain Raj, Shadman Sakib, and Tonmoy
Hossain. 2020. “A Comparative Study of CNN Transfer Learning Classification
Algorithms with Segmentation for COVID-19 Detection from CT Scan Images.” Proceedings
of 2020 11th International Conference on Electrical and Computer Engineering,
ICECE 2020: 234–37.
Shah, Vruddhi et al. 2021. “Diagnosis of COVID-19 Using
CT Scan Images and Deep Learning Techniques.” Emergency Radiology 28(3):
497–505.
Shi, Feng et al. 2021. “Review of Artificial
Intelligence Techniques in Imaging Data Acquisition, Segmentation, and
Diagnosis for COVID-19.” IEEE Reviews in Biomedical Engineering 14:
4–15.
Shukla, Prashant Kumar et al. 2020. “Efficient
Prediction of Drug–Drug Interaction Using Deep Learning Models.” IET Systems
Biology 14(4): 211–16.
Silva, Pedro et al. 2020. “COVID-19 Detection in CT
Images with Deep Learning: A Voting-Based Scheme and Cross-Datasets Analysis.” Informatics
in Medicine Unlocked 20: 100427. https://doi.org/10.1016/j.imu.2020.100427.
Soares, Eduardo, and Plamen Angelov. 2020. “A Large
Dataset of Real Patients CT Scans for COVID-19 Identification.” Harv.
Dataverse 1: 1–8.
Tabik, S. et al. 2020. “COVIDGR Dataset and COVID-SDNet
Methodology for Predicting COVID-19 Based on Chest X-Ray Images.” IEEE
Journal of Biomedical and Health Informatics 24(12): 3595–3605.
Tai, Yonghang et al. 2021. “Intelligent Intraoperative
Haptic-AR Navigation for COVID-19 Lung Biopsy Using Deep Hybrid Model.” IEEE
Transactions on Industrial Informatics 17(9): 6519–27.
Tang, Shengnan, Shouqi Yuan, and Yong Zhu. 2020. “Data
Preprocessing Techniques in Convolutional Neural Network Based on Fault
Diagnosis towards Rotating Machinery.” IEEE Access 8: 149487–96.
Wang, Zhao, Quande Liu, and Qi Dou. 2020. “Contrastive
Cross-Site Learning with Redesigned Net for COVID-19 CT Classification.” IEEE
Journal of Biomedical and Health Informatics 24(10): 2806–13.
Wu, Yu Huan et al. 2021. “JCS: An Explainable COVID-19
Diagnosis System by Joint Classification and Segmentation.” IEEE
Transactions on Image Processing 30: 3113–26.
Yan, Qingsen et al. 2021. “COVID-19 Chest CT Image
Segmentation Network by Multi-Scale Fusion and Enhancement Operations.” IEEE
Transactions on Big Data 7(1): 13–24.
Yu, Yongbin et al. 2020. “RMAF: Relu-Memristor-Like
Activation Function for Deep Learning.” IEEE Access 8: 72727–41.
