Combined Unet and CNN image classification model for COVID disease detection using CXR/CT imaging

Abstract

Accurate SARS-CoV-2 screening is made possible by automated Computer-Aided Diagnosis (CAD) which reduces the stress on healthcare systems. Since Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) is highly contagious, the transition chain can be broken through an early diagnosis by clinical knowledge and Artificial Intelligence (AI). Manual findings are time and labor-intensive. Even if Reverse Transcription-Polymerase Chain Reaction (RT-PCR) delivers quick findings, Chest X-ray (CXR) imaging is still a more trustworthy tool for disease classification and assessment. Several studies have been conducted using Deep Learning (DL) algorithms for COVID-19 detection. One of the biggest challenges in modernizing healthcare is extracting useful data from high-dimensional, heterogeneous, and complex biological data. Intending to introduce an automated COVID-19 diagnosis model, this paper develops a proficient optimization model that enhances the classification performance with better accuracy. The input images are initially pre-processed with an image filtering approach for noise removal and data augmentation to extend the dataset. Secondly, the images are segmented via U-Net and are given to classification using the Fused U-Net Convolutional Neural Network (FUCNN) model. Here, the performance of U-Net is enhanced through the modified Moth Flame Optimization (MFO) algorithm named Chaotic System-based MFO (CSMFO) by optimizing the weights of U-Net. The significance of the implemented model is confirmed over a comparative evaluation with the state-of-the-art models. Specifically, the proposed CSMFO-FUCNN attained 98.45% of accuracy, 98.63% of sensitivity, 98.98% of specificity, and 98.98% of precision.

Keywords

COVID-19 classification deep Learning U-Net Convolutional Neural Network (CNN)Moth Flame Optimization (MFO)

1 Introduction

The zoonotic diseases induced by the transmission of pathogens from animals to people are infectious, and the coronavirus is generally a zoonotic virus as well as a Ribonucleic Acid (RNA) virus of the family Coronaviridae [1]. It is a group of viruses that are contagious through respiration. Because of the crown-shaped spikes on its outer layer, the virus is known as a coronavirus [2]. A type of illness called COVID-19 causes respiratory problems in people. In Wuhan, China, in December 2019, an unidentified species transmitted the novel coronavirus (COVID-19), which had its origins in bats. [3]. This virus may affect people who have high temperatures, cough, sore throats, fatigue, muscle pains, throat pain, breathing difficulties, etc. On January 30, 2020, the World Health Organization (WHO) designated COVID-19 as a Public Health Emergency of International Concern, and on March 11, 2020, it was identified as a global pandemic [4, 5]. The four genera of SARS-CoV-2 strains are alpha, beta, gamma, and delta. While gamma and delta were descended from birds and peccary gene provision, alpha, and beta were descended from bats. Civet cats and camels serve as carriers for SARS as well as the Middle East Respiratory Syndrome (MERS) [6]. The most common symptom associated with the COVID-19 illness is pneumonia which appears to be the most frequent sign of contamination along with other major indications like fever, cold, body ache, coughing, sore throat, and epistaxis. RT-PCR and gene sequencing of respiratory or blood samples are the primary screening techniques for COVID-19. The use of RT-PCR on throat swab samples, which has a reported positive rate of 30 to 60 percent, may infect a significant number of healthy people even if the infected individuals themselves have not been diagnosed [7, 8]. Chest CT is one method that may be used to diagnose Covid-19. Visual indices that are associated with COVID-19 can be seen on X-ray images. According to the investigations, CXR revealed multilobar involvement as well as opacities in the peripheral airways. Radiography and radiology photographs may sometimes be used to diagnose sickness more rapidly [9].

Since COVID-19 is a pandemic and highly contagious, early diagnosis is necessary. For this reason, the efficacy of AI is employed in disease detection. Later, the researchers focused on CAD, which became the major research area in the medical field as well as clinical prognostics [10]. CNN has recently achieved improvements in object recognition, semantic segmentation, as well as picture classification in image processing. Furthermore, CNN in recent studies provides an optimum solution for COVID-19 classification [11]. Effective object recognition from medical images could save clinicians a lot of time, give accurate quantitative evaluations, and hasten the diagnosis process [12]. Detection of sparsely distributed things from large-scale data, however, remains a challenging task. From a DL standpoint, grouping applications by input data type will make it easier for readers to comprehend typical research frameworks [13, 14]. Although these DL algorithms appeared to be accurate, fresh data from several hospitals revealed that they were ineffective. Because of how these DL systems gather data, according to previous research, it’s nearly impossible to learn false shortcuts [15, 16]. Due to the coronavirus’s explosive global expansion, researchers have faced several difficulties in creating and putting into use CAD methods to detect and forecast sickness. The availability of datasets, changes to the architectural design of DL networks, and hardware improvements are the biggest barriers to coronavirus diagnosis. A large portion of the dataset is used for binary class or multiclass [17]. It is accepted that the categorization for COVID-19 diagnosis is incredibly unbalanced. The performance of a trained model may be biased as a result of the skewness of the dataset. Due to the challenge of obtaining sufficient high-quality photos, especially in the early phases of the pandemic [18]. To solve the issues, the competence of DL-assisted CAD algorithms attracted the focus of researchers. Table 1 summarizes the recent research surveys in COVID-19 detection and classification using AI algorithms. However, the studies exposed the challenges, future, and opportunities of AI algorithms support in COVID-19 classification. Though AI supports COVID-19 diagnosis and yields more precise findings than RT-PCR, CAD-assisted AI in COVID-19 handling is still in its infancy and needs to be improved in terms of data secrecy, accuracy, computational complexity, and so on.

Table 1
Recent Research and Advancements in COVID-19 Detection and Classification using Various AI Algorithms

Authors Published Year/Type Methods Image Modality Findings

Manoj Kumar M et al. [2] 2022/Survey ML/DL CXR/CT AI can provide better results than RT-PCR

Abhishek Agnihotri, and Narendra Kohli [9] 2023/Survey DL CXR/CT Exposed the challenges in implementing a real-time COVID classification system

Bhattacharya S et al. [13] 2021/Survey DL CXR/CT/MRI Still need improvements concerning data privacy, accuracy, and transparency

Francesco Piccialli et al. [15] 2021/Survey AI(ML/DL) CXR/CT/ Lung Ultrasound (LUS) Still need improvements in overall AI support to handle the COVID-19 pandemic

Nayal, K et al. [23] 2022/Survey AI – Explores the role of AI adoption under the disruption of the COVID pandemic in several fields

O. Tutsoy et al. [24] 2022/Research AI Real Data based on Big Data Established a set of policies to minimize COVID casualties and morbidity

Authors	Published Year/Type	Methods	Image Modality	Findings
Manoj Kumar M et al. [2]	2022/Survey	ML/DL	CXR/CT	AI can provide better results than RT-PCR
Abhishek Agnihotri, and Narendra Kohli [9]	2023/Survey	DL	CXR/CT	Exposed the challenges in implementing a real-time COVID classification system
Bhattacharya S et al. [13]	2021/Survey	DL	CXR/CT/MRI	Still need improvements concerning data privacy, accuracy, and transparency
Francesco Piccialli et al. [15]	2021/Survey	AI(ML/DL)	CXR/CT/ Lung Ultrasound (LUS)	Still need improvements in overall AI support to handle the COVID-19 pandemic
Nayal, K et al. [23]	2022/Survey	AI	–	Explores the role of AI adoption under the disruption of the COVID pandemic in several fields
O. Tutsoy et al. [24]	2022/Research	AI	Real Data based on Big Data	Established a set of policies to minimize COVID casualties and morbidity

However, DL algorithms exposed impressive methods for extracting high-dimensional features from medical images [19]. First of all, this prevents the need to continuously describe how text or images are fed into a CNN. Second, programs that use the same kind of input data share a lot of characteristics [20]. For image synthesis tasks, CNN’s U-Net encoder and decoder are made to perform better. The pooling layers are replaced by the stride and fractional strides of convolution in the decoder, which enables the model to obtain its up-sampling and down-sampling during training [21]. Additionally, batch normalization is applied to the encoder and decoder during the training phase in order to address issues with poor initialization. As a result, the aforementioned enhancements to U-Net reduce the current instability of training a standard CNN [22]. With these efficiencies in Unet and CNN, this paper aims to introduce a novel U-Net and CNN fusion algorithm to enhance disease classification accuracy. The main contribution of this context is given as follows.

In order to develop an automated and precise COVID-19 diagnosis, a novel fusion algorithm is introduced using CNN and Unet.

The input images from multiple datasets are first gathered and subjected to an image filtering approach namely the Weiner filter to remove any noises, and then data augmentation is used to solve the over-fitting and imbalanced data.

Now, image segmentation is applied to the augmented images using U-Net. The U-Net technique is utilized for extracting and segmenting the important features.

In addition, the U-Net architecture is fused with CNN to classify the images, and the weights of U-Net and CNN have been tuned to improve accuracy performance via the CSMFO algorithm.

To confirm the competence of the suggested method, a comparative evaluation is conducted using the classed results by means of accuracy, sensitivity, specificity, and precision.

The rest of this article is ordered as follows. Section II represents a brief description of the recent COVID-19 disease classification algorithms and their features. The proposed architecture of the novel COVID-19 disease classification is explained in Section III. Section IV deals with the DL and optimization concepts. Further, Section V gives the obtained results and their experimentation details. Finally, Section VI concludes the paper.

2 Literature review

2.1 Related works

In 2021, Sharmila, and Florinabel [25] developed a COVID-19 classification model using Deep Convolutional Generative Adversarial Network (GAN)-CNN (DCGAN-CNN). Here, the CXR images were classified through 8 convolutional layers, 4 max-pooling layers, as well as 2 Fully Connected (FC) layers. In this research, synthetic images were produced to the extent the data used and extracted the deep features. From the investigation results, it revealed competitive performance over AlexNet and GoogLeNet). In 2021, Dan Nguyen et al. [26] trained 9 CNN-based models for COVID-19 disease classification via 3-D Computed Tomography (CT) images. The data for classification was gathered from various countries containing COVID-19+ and negative people. Herein, 72%, 8%, and 20% of data have been used for training, validating, as well as testing the data. The efficacy was recorded via accuracy and Area Under Curve (AUC). For accuracy, 97% to 98% performance was attained. Similarly, for AUC, 82% to 87% performance was achieved. In 2022, A. Chaddad et al. [27] established a CNN with Gaussian Mixture Model (GMM) named (GMM-CNN) strategy to encode the features of COVID-19 disease classification. CT and CXR images were employed for experimentation. Besides, the Random Forest (RF) model was employed for classification. The performance was accessed via accuracy and Receiver Operator Characteristics (ROC). In comparison with traditional CNN, GMM-CNN attained improved performance. In 2020, S. Sakib et al. [28] established a COVID-19 disease diagnosis model through DL-based Chest Radiograph Classification (DL-CRC). The dataset contained CXR of normal, pneumonia, and COVID-19+ patients’ data. Besides, Data Augmentation of Radiograph Images (DARI) and GAN strategies were employed to create synthetic data for training as well as testing the proposed DL-CRC technique. The empirical study exposed the efficiency of the DL-CRC model by outrunning the conventional CNN-based models. In 2021, A. Castiglione et al. [29] introduced a COVID-19 classification scheme through an optimized CNN named ADECO-CNN model. However, CT images were employed for classification and compared with various CNN-based models to validate the performance of ADECO-CNN. From the experimentation, the implemented ADECO-CNN model accomplished 99% of accuracy. In 2022, Pir Masoom Shah et al. [30] developed a COVID-19 disease detection model via CNN with Gated Recurrent Unit (GRU) methods. Here, CXR images were used with normal, pneumonia, and COVID-19+ people’s data. CNN was employed for feature extraction and GRU was utilized to classify the data. The experimental study exposed the performance by 96% of precision and 95% of F1-score. In 2021, A. Abbas et al. [31] developed a Self-Supervised Super Sample Decomposition for Transfer Learning (4S-DT) framework for COVID-19 disease detection via CXR images. Through the use of a downstream learning method and a Class-Decomposition (CD) layer for streamlining the data’s local structure, 4S-DT aided in enhancing the resilience of knowledge transformation. Through experimentation, the 4S-DT model proved its efficiency by obtaining 99% of accuracy. In 2021, El-Sayed M. El-Kenawy et al. [32] developed a COVID-19 disease classification framework using an Advanced Squirrel Search Optimization Algorithm (ASSOA) with a Multi-layer Perceptron (MLP) named ASSOA+MLP model. Initially, the data augmentation and feature learning processes were carried out by CNN. ASSOA was employed for selecting the significant features and optimizing the weights of MLP. MLP was used to classify the CXR images. The investigation validated the ASSOA+MLP model through a comparative study and achieved 99% accuracy.

2.2 Review

From the literature, it is clear that plenty of research has been done on COVID-19 disease classification using various DL algorithms.

In contrast to CT and MRI, the most popular radiological diagnostic imaging is CXR since it is less expensive, takes less time to process, and exposes patients to less radiation. It is critical to rapidly isolate suspected individuals during pandemics like the recent one so that they can receive the appropriate care. Controlling epidemics also requires quick diagnosis of these patients. Additionally, confirmatory laboratory tests may be used in conjunction with AI-based illness classification. Similarly, implementing AI-based quick and time-saving solutions to help with prognosis and evaluation of recovering/follow-up patients may be a great idea. Researchers have recently undertaken attempts to create COVID-19 classification or identification systems based on CXR/CT images, each with a particular set of capabilities. However, in order to create more dependable and accurate classification models, the studies have certain fundamental flaws that must be addressed. To name a few shortcomings, only a small number of studies have used CXR/CT pictures (in various classes) from vastly varied lung diseases. The process facilitates categorization (for DL algorithms), but at the same time, it is biased since normal vs. infected (with high severity) groups’ lungs differ in pixels. On the other hand, only a few research have used augmentation approaches with a small number of augmentation types to raise the sizes of the datasets and to generalize the models. The AI-based models trained on fewer pictures run the risk of overfitting in the classification models since there are only a limited number of CXR/CT images available for COVID-19. Lack of these sorts of selection criteria may result in models being trained incorrectly and performing poorly in practical applications. Based on the prior literature, there were variations in overall performance, and the classification tasks for COVID-19 differed in terms of factors linked to the accuracy of findings. Previous research has only addressed ways to speed up classification processes, increase accuracy, or even just improve performance. Additionally, there are changes between earlier literature’s categorization approaches, stages, and processes. The implemented DL algorithms exposed competitive performance over the standard DL concepts. However, computational complexities, overfitting, high computational time, imbalanced data, too many processing steps, noisy data, limited data in datasets, difficulties in acquiring proper datasets, and poor understanding of the DL algorithms restrict the COVID-19 disease classification framework to be implemented in a real-world scenario. Intending to overwhelm the previous complexities, the researchers develop novel DL concepts yet, complexities in computation exist. Even though DL algorithms aid disease detection and classification, coping with real-time scenarios is difficult. However, the performance of the proposed CSMFO-FUCNN model on different datasets disproves any claim that they were overfitting. Some of the features and challenges of the existing methods are briefly described in

3 A Novel COVID-19 Disease Classification Framework

3.1 Proposed Architecture

Figure 1 depicts the block diagram of the implemented design for the COVID-19 disease classification model. Initially, the input images from the dataset have been collected and have undergone an image filtering technique called the Weiner filter to remove any noises. Generally, the Wiener filter can be described as an MSE-optimal stationary linear filter for images with additional noise and blur. It is necessary to assume that the signal and noise processes are 2nd-order stationery in order to calculate the Wiener filter (in the random process sense). Then, data augmentation is employed like random scaling, random resize, and random rotation to enlarge the dataset so as to cope with the imbalanced data. Now, the augmented images are subjected to segmentation using standard U-Net.

Fig. 1

Block Diagram of Implemented Design for COVID-19 Disease Classification Model.

Usually, U-Net is a semantic segmentation method that was first presented for segmenting medical images. This DL segmentation model dates back to the beginning and efficiently works with CNN, key features are extracted and segmented using the U-Net approach. The segmented photos are given to FUCNN for categorization. Additionally, the CSMFO method has been used to fine-tune the weights of the FUCNN to increase accuracy performance. Using the findings from the classification, a comparison study is done to assess the efficacy of the implemented algorithm through measures like accuracy, sensitivity, specificity, and precision.

3.2 Weiner filter

Generally, an ideal low pass and nonlinear filter is the Wiener filter [33]. Now, the input image I is given to the Weiner filter. The Weiner filter is crucial for a number of applications, including image processing, medical imaging, echo cancellation, signal restoration, and linear prediction to reduce the average squared distance among the desired pixel and the filter output. It is frequently used because of its efficiency and simplicity. Also, image restoration as well as lowering the Mean Square Error (MSE) number are the primary goals of the Weiner filter. This follows the classical method of least squares in order to decrease the MSE. It is regarded as simple since it computes a set of ideal filter weights that lower the incoming signal’s noise level using a system of linear equations. To determine these weights and give a precise estimate of the undistorted deterministic signal under Gaussian noise, it calculates cross-correlation and covariance matrices of noisy signals. A set of ideal filter weights is chosen by first estimating the noise statistics. A signal deterministic component is then calculated by processing a fresh input signal with the best filter weights and noise characteristics. In situations when the noise distribution is Gaussian, this approach performs well. Additionally, its execution only needs a small number of computing steps, which are processed relatively quickly. Specifically, Wiener filters possess the ability to concurrently decrease blur and eliminate noise in an image. The initial phase of the projected approach involves running a noisy, blurry image through a Wiener filter and constructing a filter. The Wiener filter W in the Fourier domain is implemented as shown in Equation (1). $I_{W} = \frac{D^{*} P_{i}}{{| H |}^{2} P_{i} + G}$ (1)

Dividing Equation (1) by P_i alters its behavior as shown in Equation (2). $I_{W} = \frac{D^{*}}{{| H |}^{2} + \frac{G}{P_{i}}}$ (2)

Herein, H indicates the degradation function, D^* specifies the degradation function’s complex conjugate, and G signifies the power spectral density of noise, where $\frac{G}{P_{i}} = 1 / SNR$ (reciprocal of SNR). SNR stands for Signal to Noise Ratio. Also, P_i means for power spectral density of un-degraded images. Based on statistics inferred from a pixel’s immediate neighborhood, the Wiener filter is used to eliminate noise from a distorted discovered picture. This filter is reliant on noise power, or the variation in noise in a distorted picture. When the variance is high, the filter does not do much smoothing; when the variance is low, the filter does more smoothing. Here, Gaussian noise is added with a mean of 0 and standard variance of 0.024 which makes the filter do more smoothing. At this point, the output I_W is given to data augmentation. The sample input images are displayed in Fig. 2. Figure 2(a), demonstrates the original chest X-ray images of normal people, pneumonia, and COVID-19+ correspondingly. Additionally, Fig. 2(b) shows examples of the filtered images from the aforementioned input images using the Weiner filter technique. The input images have now been improved and the extra noise has been removed.

Fig. 2

Sample Images displaying (a) Original Chest X-ray Images of Normal People, Pneumonia, and COVID-19+, and (b) Improved Images of Chest X-ray Images of Normal People, Pneumonia, and COVID-19+ using Weiner Filter.

3.3 Data augmentation

Here, the filtered image I_W is given as input. The normal images are augmented to improve training performance by techniques like cropping, rotation, and scaling. Specifically, when the dataset is insufficient and to enlarge the size of the training and testing data. Here, one full-size image is flipped vertically, rotated left and right, and cropped using the image-editing techniques, and increasing the training dataset’s size is not what was originally intended. Figure 3 shows sample images that have been augmented to enlarge the dataset. Now, the training and testing data size has been enlarged so as to attain improved classification accuracy. The output images such as the original image I₀, vertically flipper image I_v, left rotated image I_l, right-rotated image I_r, and cropped image I_c are subjected to U-net segmentation.

Fig. 3

Sample Images displaying (a) Original Chest X-ray Image, Pneumonia, and COVID-19+, (b) Vertically Flipped Images, (c) Rotate –90⁰, (d) Rotate +90⁰, and (e) Cropped Image.

4 Lung Disease Classification by Proposed DL Model

4.1 U-Net Image Segmentation

The implemented architecture of COVID-19 disease classification is discussed. The input image (I_o (n) , I_v (n) , I_l (n) , I_r (n) , I_c (n)) is given as input for U-Net segmentation. Here, the images are initially normalized. Generally, U-Net [34, 35] is a modified variant of fully CNN that can function effectively even with small datasets. The U-Net is built with an encoder-decoder architecture, where the encoder is in charge of input features and only takes into account the most important ones, while the decoder is in charge of learning the features from the encoder and producing better results for the actual input or desired result. Two important U-Net parts, the contracting and expansive paths, are situated on the left and right sides and compute down- and up-sampling, respectively. Plotting initial weights from a Gaussian distribution with a standard deviation of $\sqrt{2 / n}$ , in which n represents the neuron’s input node count. Consider a 3 × 3 convolution with 64 feature channels in U-net architecture (cross-convolution with ReLU layers), n is equivalent to 9 × 64 = 576. Moreover, the loss function of the cross-entropy as well as the softmax prediction class value of all pixels on the last feature map are combined to create the energy function E_x (v), which is defined as given in Equation (3). $E_{x} (v) = \frac{exp (u_{x} (v))}{\sum_{x^{'} = 1}^{x} exp (u_{x} (v))}$ (3)

Here, v signifies the categories count, u_x (v) represents the feature channel v’s activation value, in which the pixel is in y, (y ∈ ϑ). The notion ϑ refers to the true label set of the pixel. The energy function is expressed in Equation (4), with the goal of cross-entropy C_E being to penalize the nonconformity of ρ_q(y) (y) and q at every point. $C_{E} = \sum_{y \in ϑ} ω (y) \log (ρ_{q (y)} (y))$ (4)

The notation, q : ϑ→ { 1, …, v } states every pixel’s true label, ω : ϑ → Z indicates the pixel training’s weight map, and ρ_q(y) (y) points out the probability of the softmax function. Table 3 shows the parametric description of the FUCNN employed for segmentation and classification. The weight ω of the standard U-Net design and weight W of CNN is now optimized with CSMFO. The graphical depiction of FUCNN optimization utilizing the CSMFO Model is shown in Fig. 4. The fundamental design of U-Net for 32 × 32 pixel resolution is shown in Fig. 5.

Fig. 4

Graphical Depiction of U-Net Optimization utilizing CSMFO Model.

Fig. 5

Fundamental Design of U-Net for 32 × 32 Pixel.

4.2 Conventional MFO

Generally, the MFO algorithm was inspired by how moths navigate [36]. The moth population is expressed as provided in Equation (5), where m denotes the number of moths and r denotes dimensions. According to Eq, the fitness values for moths are kept in an array (6). $𝕄 = [\begin{matrix} \begin{matrix} α_{11} α_{12} \begin{matrix} \dots α_{1 r} \end{matrix} \end{matrix} \\ \begin{matrix} α_{21} α_{22} \dots \\ \dots \dots \dots \\ α_{m 1} α_{m 2} \dots \end{matrix} \begin{matrix} α_{2 r} \\ \dots \\ α_{mr} \end{matrix} \end{matrix}]$ (5) $𝔽 𝕄 = [\begin{matrix} \begin{matrix} 𝔽 𝕄_{1} \\ 𝔽 𝕄_{2} \\ \dots \end{matrix} \\ \dots \\ F 𝕄_{m} \end{matrix}]$ (6)

Here, Equation (7) defines the flames of a moth, and Equation (8) shows the fitness of flames. $F = [\begin{matrix} \begin{matrix} β_{11} β_{12} \begin{matrix} \dots β_{1 r} \end{matrix} \end{matrix} \\ \begin{matrix} β_{21} β_{22} \dots \\ \dots \dots \dots \\ β_{m 1} β_{m 2} \dots \end{matrix} \begin{matrix} β_{2 r} \\ \dots \\ β_{mr} \end{matrix} \end{matrix}]$ (7) $F F = [\begin{matrix} \begin{matrix} F F_{1} \\ F F_{2} \\ \dots \end{matrix} \\ \dots \\ F F F_{m} \end{matrix}]$ (8)

The three-tuple representation of the global optimal in MFO is shown in Equation (9), where A and B are the parameters used to form the moth population and their corresponding fitness values, T is the termination procedure, and it returns either true or false. $mfo = ABT$ (9) $𝕄_{i} = S (𝕄_{i}, F_{j})$ (10)

The transverse orientation of moths is a key inspiration for MFO, and this model is realized by updating the moths with their corresponding flames as shown in Equation (10), where $S$ denotes spiral operation. The logarithmic spiral of the moths is explained by Equation (11), where $h$ depicts the i^thmoth and j^th flame distance, $c$ is a constant, and $g$ is any integer between $[w, 1]$ , where $w$ decreases linearly from -1to - 2. $S (𝕄_{i}, F_{j}) = h_{i} \cdot e^{c g} \cdot cos (2 π g) + F_{j}$ (11)

The distance $h$ is estimated as given in Equation (12). $h_{i} = | ℱ_{j} - M_{i} |$ (12) $𝔽 ℂ = round (n - y * \frac{n - 1}{t})$ (13)

According to Equation (13), as the iteration count rises, the moths with better flames are taken into account for further processing and the rest are ignored. Here, $n$ denotes the maximum flames count, $y$ denotes the present iteration count, and t denotes the maximum iteration. Simply explained, A develops the initial population and calculates the fitness values. In addition, B runs repeatedly until T returns true.

4.3 Proposed CSMFO Algorithm

The conventional MFO is effective at tackling discrete optimization issues as well as complex issues. However, the performance is constrained by difficulties in achieving the best convergence and problems with local optima. By substituting a chaotic system-based update function for the logarithmic spiral of moths provided in (11), the classic MFO is modified with the intention of overwhelming the problems and attempting to resolve the imprecise COVID-19 disease classification systems. The enhanced moths’ update is defined by Equation (14), where $z$ is initially set to zero and increases by one after each iteration. In addition, $K$ is a random number between [0, 1] . $S (𝕄_{i}, F_{j}) = h_{i} \cdot e^{c z} \cdot cos (2 π z) + F_{j}$ (14) $z = K z (1 - tanh (z))$ (15)

This improvement made the traditional MFO perform well and helps to achieve better accuracy performance and is thus named CSMFO. Algorithm 1 summarizes the proposed algorithm of the CSMFO model.

Algorithm 1: Pseudocode of Proposed CSMFO Model
Begin
Initialize the population and flames
Estimate the fitness values
Apply three-tuple fashion
$𝕄 = A ()$
while $T (𝕄) = false$
for i = 1 : m
for j = 1 : r
$𝕄 (i, j) = (ul (i) - ll (i) * rand () + ll (i))$
End
End
Update the position of moths and flames as per
Equation (14)
$𝔽 𝕄 = objectivefunction (𝕄)$
if iter = =1
$F = sort (M)$
$F F = sort (F M)$
else
$F = sort (M_{z - 1}, M_{z})$
$F F = sort (M_{z - 1}, M_{z})$
end
for i = 1 : m
for j = 1 : r
Update $w$ and $z$
Calculate the distance as shown in Equation (12)
Update the moths with respect to Equation
(10) and (11)
End
end
$𝕄 = B (𝕄)$
end
end

4.4 Objective function

The major aim of this paper is to improve the accuracy of COVID-19 disease classification. For this reason, the excellence of MFO is utilized and tuned via chaotic system concepts. Algorithm 2 defines the fitness estimation of moth population $𝕄$ . Here, the initial $𝕄$ comprised of mmoths, training data Train_dataand validation data Valid_data, and an empty vector f to collect each fitness of all m moths. Fitness evaluation is carried out by assessing all mmoths as per algorithm 2 and converts each m individual of $𝕄$ into new solution. Now, training and validating the network is performed using Train_data and Valid_data. Herein, the performance of every iteration is calculated in terms of validation accuracy Accuracy_valid and stored in f. Eventually, fit_funreturns the moth population $𝕄$ with mmoths and their corresponding fitness. Equation (16) portrays the fitness function fit_fun of CSMFO and the objective of this research, in which i indicates iteration count and j is the input image given for classification. ${fit}_{f} un = max {({Accuracy}_{valid})}_{i}^{j}$ (16)

Algorithm 2: Pseudocode of Fitness Estimation of Proposed CSMFO Model
Begin
Initialize the population $𝕄$ with m moths for fitness estimation, Train_data and Valid_data
f =∅ ; // empty vector for storing fitness
for each m in $𝕄$ do
$𝕄 = A ()$
Sol = Convert the mmoths to its corresponding solution
Train and validate the network using Train_dataand Valid_data
f = Accuracy_valid of all solutions
end
Return $𝕄$ with all mmoths and their corresponding fitness f

4.5 Solution Encoding

U-Net exposes Fully Convolutional Network (FCN) computational properties because it has a distinct Neural Network (NN) architecture. Even while U-Net offers noteworthy performance, there is always room for enhancement in terms of the accuracy of the results in the field of image processing. With the goal of improving accuracy for COVID-19 disease classification, the final convolution layer of U-Net’s weight parameter ωis improved using the CSMFO model. The contraction process specifically carries out down-sampling, which encodes the input data and omits irrelevant features. It thus narrows the feature space and significantly trains the network. Here, the down-conv4 layer’s weight parameter is tuned with a constant ρ to achieve noteworthy performance, where ρ = 0.5. The solution encoding of CSMFO is shown in Fig. 6, where ω is the weight parameter of U-Net, and W denotes CNN’s weight.

Fig. 6

Solution Encoding of Proposed COVID-19 Disease Classification using CSMFO.

Figure 7 represents the segmented CXR images using the proposed CSMFO-FUCNN model. Figure 8 signifies the flowchart of the proposed CSMFO.

Fig. 7

Sample Images displaying (a) CXR Images of Normal People, Pneumonia, and COVID-19+, and (b) Segmented Images of CXR Images of Normal People, Pneumonia, and COVID-19+ using using Proposed CSMFO-FUCNN.

Fig. 8

Flow chart of Proposed CSMFO for COVID-19 Disease Classification.

5 Simulation results

5.1 Simulation Setup

The suggested COVID-19 disease classification system via the modified U-net and CNN approach was implemented in MATLAB on Intel® core i7 processor 1065G7@3.9 GHz, 32 GB RAM, 64-bit operating system. Here, CXR/CT Dataset was employed for classification which is available in the link COVID-19 Radiography Database | Kaggle and ieee8023/covid-chestxray-dataset: We are building an open database of COVID-19 cases with chest X-ray or CT images. (github.com) [Access date: 27-10-2022]. The efficiency and novelty of the implemented model were recorded via the simulation results. Here, the evaluation was implemented through various performance parameters such as accuracy, sensitivity, specificity, and precision. In addition to this, a comparative study is carried out to assess the performance of the suggested model over several traditional methods such as K-Nearest Neighbor (KNN) [37], Deep belief Network (DBN) with hosted Cuckoo optimization Algorithm (COA) [38], CNN with Bayesian optimization (BO) [39], and CNN-Multi-Objective Emperor Penguin Optimizer (CNN-MOEPO) [40].

5.2 Algorithmic analysis

The implemented COVID-19 disease classification scheme via the modified U-Net model is described in this section. The important features were retrieved using the U-Net methodology and classified using the suggested method. Figure 9 displays the performance of the proposed model concerning the convergence graph of the implemented CSMFO and other conventional models such as the Whale Optimization Algorithm [41], Grey Wolf Optimization [42], Duffing Equation-based Tuna Swarm Optimized Resnet 101 (DETS-Optimized Resnet 101) [43], Euler-Lagrangian Equation-Assisted WOA-DCNN (EEA-WOA-DCNN) [44], and conventional MFO in Fig. 9(a). Here, the proposed CSMFO attained 7.89%, 6.14%, 4.56%, 1.01%, and 3.11% better than WOA-FUCNN, GWO-FUCNN, MFO-FUCNN, DETS-Resnet 101, and EEA-WOA-DCNN respectively for accuracy performance. similarly, Fig. 9(b) shows the Receiver Operating Characteristic (ROC) curve in which the Area Under Curve (AUC) is 0.931. Figure 9(c) represents the accuracy and loss curves of the proposed model. In the initial epochs, the proposed model shows minimum accuracy and at the end of the computation, the accuracy is considerably increased and attained 98.45%. correspondingly, the loss is high in the beginning and at the end, the proposed model becomes stable with minimized error. Moreover, Fig. 10 displays the performance evaluation of the suggested CSMFO-FUCNN for COVID-19 disease classification over other conventional models by varying the training percentage. The accuracy of proposed CSMFO-FUCNN attained improved performance over other models which is 8.52% better than KNN, 4.67% better than COA-DBN, 5.33% improved than BO-CNN, 3.38% improved than CNN-MOEPO, 1.94% better than DETS-Resnet 101, and 2.11% better than EEA-WOA-DCNN for 70% of training data in Fig. 10(a). Figure 10(b) depicts the sensitivity performance of the proposed CSMFO-FUCNN which is 6.88%, 5.66%, 5.68%, 4.03%, 2.02%, and 2.88% better than KNN, COA-DBN, BO-CNN, CNN-MOEPO, DETS-Resnet 101, and EEA-WOA-DCNN, respectively for 30% of training data. For 50% of training data, the proposed CSMFO-FUCNN accomplished 8.12% better than KNN, 4.67% better than COA-DBN, 5.33% improved than BO-CNN, 3.02% improved than CNN-MOEPO, 1.45% better than DETS-Resnet 101, and 2.01% better than EEA-WOA-DCNN for specificity in Fig. 10(c). Finally, Fig. 10(d) demonstrates the precision results of the proposed CSMFO-FUCNN for 80% of training data which reached 7.98%, 4.14%, 5.32%, 2.98%, 0.98%, and 1.56% better than KNN, COA-DBN, BO-CNN, CNN-MOEPO, DETS-Resnet 101, and EEA-WOA-DCNN, respectively. From the experimentation, it is evident that the proposed CSMFO- FUCNN attained promising results and outperformed the existing models.

Fig. 9

Performance indicating (a) Convergence Graph of Proposed CSMFO for COVID-19 Disease Classification over other Traditional Models concerning Accuracy, (b) ROC Plot, and (c) Accuracy and Loss of proposed CSMFO-FUCNN Model.

Fig. 10

Performance Evaluation of Proposed CSMFO-FUCNN for COVID-19 Disease Classification over other Conventional Models by means of (a) Accuracy, (b) Sensitivity, (c) Specificity, and (d) Precision.

Table 2 summarizes the performance evaluation of the proposed CSMFO-FUCNN concerning training percentages like 30%, 50%, 70%, and 80% respectively, and varying epochs of DL algorithms such as 30, 50, 60, and 70 respectively. For 30 epochs of accuracy, the proposed CSMFO-FUCNN accomplished 65.47%, 76.89, 81.62, and 88.85% for 30%, 50%, 70%, and 80% of training data respectively. For 70% training data, the suggested method’s accuracy for epoch = 50, 60, and 70, are 88.27, 95.74, and 97.78 respectively, while its accuracy for Epoch = 70 is better at 97.78 %. For 80% of training data, the sensitivity of the proposed technique at epoch 80 is 9.76%, and 5.58% better than at epochs 30, and 50 respectively. Similar to this, when epoch 70, the proposed approach has a high specificity of 98.98% which is 1023%, and 5.81% improved than epochs 30, and 50 respectively. For 80% of training data and epoch 70, the proposed CSMFO-FUCNN reached its highest precision performance of 98.98%. This investigation helps to identify the efficacy of the proposed DL algorithm CSMFO-FUCNN and proves its significance in COVID-19 disease classification.

Table 2

Features And Challenges Of COVID-19 Disease Classification Strategy Using Various DL Methods

Authors	Methods	Datasets	Features	Challenges
Sharmila V J, and Jemi Florinabel D [25]	The DCGAN-CNN Model	CXR	•Exposed competitive performance •The computationally effective architecture was developed	•Raw data were used for experimentation •Limited data was used
Dan Nguyen et al. [26]	9 CNN-based Models	3D CT	•Achieved better accuracy and AUC •Various CNN-based models were tested	•Too limited testing data •Computationally complex
A. Chaddad et al. [27]	The GMM-CNN Model	CT and CXR	•Attained improved accuracy and ROC •Significant features were used for experimentation	Exposed high computation time due to many DL algorithms
S. Sakib et al. [28]	The DL-CRC Model	CXR	•Provided competitive results •No pre-processing was performed	•Noisy data limits the performance •Limited testing data
A. Castiglione et al. [29]	The ADECO-CNN Model	CT	•Reached 99% of accuracy •No pre-processing was performed	•Limited data was used •High time-consuming experimentation
Pir Masoom Shah et al. [30]	The CNN-GRU Model	CXR	•No pre-processing was performed •Accomplished competitive accuracy	•Complex expressions were used •Limited data were employed for the investigation
A. Abbas et al. [31]	The 4S-DT Model	CXR	•Attained 99% of accuracy •No pre-processing was performed	•Complex expressions were used •Exposed computationally complex architecture
El-Sayed M. El-Kenawy et al. [32]	The ASSOA+MLP Model	CXR	•Obtained improved accuracy •Outperformed the conventional models	•Small testing dataset •Revealed time-consuming execution

Table 3

Parametric Description of Developed FUCNN Model

Parameters	Values
Input layer	1 × 1 ×1 conv
Activation	relu
Conv	3 × 3 ×3
ReLU	3 × 3 ×3 conv
Dropout	0.25, and 0.5
Sigmoid	3 × 3 ×3 conv
Max pooling	2 × 2 ×2
Batch Normalization	512
Deconv	3 × 3 ×3
ReLU	3 × 3 ×3 conv
Dropout	0.25, and 0.5
Sigmoid	3 × 3 ×3 conv
Max pooling	2 × 2 ×2
Strides	2 × 2 ×2
Output	1 × 1 ×1 conv
Activation	Sigmoid
Batch Normalization	512

5.3 Discussion

Table 4
Performance Evaluation of Proposed CSMFO- FUCNN concerning Training Percentage

Accuracy (%)

Training percentage CSMFO-FUCNN Epoch = 30 CSMFO-FUCNN Epoch = 50 CSMFO-FUCNN Epoch = 60 CSMFO-FUCNN Epoch = 70

30 65.47 76.24 88.58 92.14

50 76.89 81.74 92.34 95.65

70 81.62 88.27 95.74 97.78

80 88.85 92.44 98.45 98.45

Sensitivity (%)

30 63.41 75.89 88.11 92.97

50 75.68 80.74 92.95 96

70 81.62 89.27 95.02 98

80 89 93.12 98.63 98.63

Specificity (%)

30 66.14 77.35 89.65 93.57

50 77.28 82.68 93.68 96.75

70 82.95 89.91 98 98

80 88.85 93.22 98.98 98.98

Precision (%)

30 65.47 76.13 88.58 93.45

50 76.89 81.74 92.34 95.65

70 83.82 89.27 98 98

80 89.85 93.44 98.98 98.98

Accuracy (%)
Training percentage	CSMFO-FUCNN Epoch = 30	CSMFO-FUCNN Epoch = 50	CSMFO-FUCNN Epoch = 60	CSMFO-FUCNN Epoch = 70
30	65.47	76.24	88.58	92.14
50	76.89	81.74	92.34	95.65
70	81.62	88.27	95.74	97.78
80	88.85	92.44	98.45	98.45
Sensitivity (%)
30	63.41	75.89	88.11	92.97
50	75.68	80.74	92.95	96
70	81.62	89.27	95.02	98
80	89	93.12	98.63	98.63
Specificity (%)
30	66.14	77.35	89.65	93.57
50	77.28	82.68	93.68	96.75
70	82.95	89.91	98	98
80	88.85	93.22	98.98	98.98
Precision (%)
30	65.47	76.13	88.58	93.45
50	76.89	81.74	92.34	95.65
70	83.82	89.27	98	98
80	89.85	93.44	98.98	98.98

The implemented COVID-19 disease classification scheme using CXR/CT images with CSMFO-FUCNN attained enhanced accuracy performance. Prompt identification of COVID-19 patients can assist in maximizing the use of the resources that are already available, including skilled human resources, for all the supporting measures needed for confirmed patients. The proposed model has shown that automated DL-based intelligent CXR categorization offers untapped promise for this unmet requirement. Similarly, implementing DL-based quick and time-saving solutions to help with prognosis and evaluation of recovering/follow-up patients may be a great idea. Nevertheless, in order to create a more dependable and accurate classification model, this study has certain fundamental flaws that must be addressed. The process facilitates categorization for DL algorithms, but it also introduces bias because of the disparities in lung capacity between children and adults. Instead of using unhealthy vs. normal or any other criterion, it is likely that fresh, unknown photographs belonging to a certain age group will be automatically allocated to a class of images that have images comparable to that age group. The AI-based models trained on fewer pictures run the risk of overfitting in the classification models since there are only a limited amount of CXR images available for COVID-19. However, the performance of our model on different datasets disproves any claim that they were overfitting. In the future, the model training will be improved even further by including a huge number of photos to create classification models that are more reliable and scalable. The amount of CXR data for COVID-19 is growing quickly, thus it is possible to update the classification models and improve their accuracy and usefulness in real-time settings.

6 Conclusion

This paper introduced a novel Dl algorithm for the COVID-19 disease classification model named CSMFO-FUCNN. Initially, the input images I from the dataset were collected and underwent an image filtering technique called the Weiner filter to remove any noises and produce filtered images I_W. Besides, data augmentation was utilized to enlarge the dataset size as (I_o (n) , I_v (n) , I_l (n) , I_r (n) , I_c (n)) so as to enhance the classification performance. Now, the augmented photos were subjected to segmentation by U-Net. The key features were extracted and segmented using the U-Net approach. The segmented images were given to optimized FUCNN for categorization. The CSMFO method has also been used to adjust the U-Net weights ω and CNN weight W to increase accuracy performance. Using the findings from the classification, a comparison analysis was done to measure the efficacy of the developed scheme concerning accuracy, sensitivity, specificity, and precision. The convergence of proposed CSMFO attained 7.89%, 6.14%, 4.56%, 1.01%, and 3.11% better than WOA-FUCNN, GWO-FUCNN, MFO-FUCNN, DETS-Resnet 101, and EEA-WOA-DCNN respectively for accuracy performance. The accuracy of the proposed CSMFO-FUCNN attained improved performance over other models which was 8.52% better than KNN, 4.67% better than COA-DBN, 5.33% improved than BO-CNN, 3.38% improved than CNN-MOEPO, 1.94% better than DETS-Resnet 101, and 2.11% better than EEA-WOA-DCNN for 70% of training data. For 70% of training data, the suggested method’s accuracy for epoch = 50, and 70 are 88.27, and 95.74, respectively, while its accuracy reached the highest performance at epoch = 80 which was better at 97.78 %. From the accomplished results, the proposed CSMFO-FUCNN model proved its efficacy and significance in COVID-19 disease classification. In the future, DL algorithms will be used for data augmentation to avoid specific dataset enlargement procedures. Also, the model training will be improved even further by including a large number of images in order to create classification models that are more consistent and scalable.

References

Z.W.

, Yuan

, Yuen

K.S.

, Fung

S.Y.

, Chan

C.P.

, Jin

D.Y.

, Zoonotic origins of human coronaviruses, Int J Biol Sci 16(10) (2020), pp. 1686–1697.

Manoj Kumar

M.V.

, Atalla

, Almuraqab

, Moonesar

I.A.

, Detection of COVID-19 Using Deep Learning Techniques and Cost Effectiveness Evaluation: A Survey, Front. Artif. Intell. 27 May 2022.

Alanagreh

, Alzoughool

, Atoum

, The Human Coronavirus Disease COVID-19: Its Origin, Characteristics, and Insights into Potential Drugs and Its Mechanisms, Pathogens 9(5), 2020.

Shereen

M.A.

, Khan

, Kazmi

, Bashir

, Siddique

, COVID-19 infection: Origin, transmission, and characteristics of human coronaviruses, J Adv Res 24 (2020),pp. 91–98.

Kassahun

, Berhanu

, Damtew

Damtew, Reviewon Coronavirus, a Middle East Respiratory Syndrome (MERS-CoV), Online Journal of Animal and Feed Research, 2020.

Joseph Larsen

, Margaret Martin

, John Martin

, Kuhn

, Hicks

J.B.

Modeling the Onset of Symptoms of COVID-19, Front. Public Health, 13 August 2020.

Cucinotta

, Vanelli

, WHO Declares COVID-19 a Pandemic, Acta Biomed 91(1) (2020),pp. 157–160.

Singhal

, Review of Coronavirus Disease-(COVID-19), Indian J Pediatr 87(4) (2020),pp. 281–286.

Agnihotri

, Kohli

, Challenges, opportunities, and advances related to COVID-19 classification based on deep learning, Data Science and Management 6(2) (2023) 98–109 .

10.

Chan

H.P.

, Hadjiiski

L.M.

, Samala

R.K.

, Computer-aided diagnosis in the era of deep learning, Med Phys 47(5) (2020),pp. e218–e227 .

11.

Sultana

, Sufian

, Dutta

, Advancements in Image Classification using Convolutional Neural Network, Conference: 2018 Fourth International Conference on Research in Computational Intelligence and Communication Networks (ICRCICN), 2018.

12.

Yamashita

, Nishio

, Gian

R.K.

, Do and K. Togashi, Convolutional neural networks: An overview and application in radiology, Insights into Imaging 9 (2018), 611–629.

13.

Bhattacharya

, Reddy Maddikunta

P.K.

, Pham

Q.V.

, Gadekallu

T.R.

, Krishnan

S.S.R.

, Chowdharya

C.L.

, Alazab

, Jalil Piran

, Deep learning and medical image processing for coronavirus (COVID-19) pandemic: A survey, Sustain Cities Soc 65, 2021.

14.

Sarker

I.H.

, Deep Learning: A Comprehensive Overview on Techniques, Taxonomy, Applications and Research Directions, SN Comput Sci 2(6), 2021.

15.

Piccialli

, di Cola

V.S.

, Giampaolo

, Cuomo

, The Role of Artificial Intelligence in Fighting the COVID-19 Pandemic, Information Systems Frontiers 23 (2021),pp. 1467–1497.

16.

Iqbal Sarker

, Machine Learning: Algorithms, Real-World Applications and Research Directions, SN Computer Science 2(160), 2021.

17.

Sharma

, Jain

, Mishra

, An Analysis Of Convolutional Neural Networks For Image Classification, Procedia Computer Science 132 (2018),pp. 377–384.

18.

Krawczyk

, Learning from imbalanced data: Open challenges and future directions, Progress in Artificial Intelligence 5 (2016), 221–232.

19.

Nirthika

, Manivannan

, Ramanan

, Wang

, Pooling in convolutional neural networks for medical image analysis: A survey and an empirical study, Neural Computing and Applications 34 (2022),pp. 5321–5347.

20.

Lou

, Guan

, Loew

, DC-UNet: Rethinking the U-Net Architecture with Dual Channel Efficient CNN for Medical Images Segmentation, Available Online 2020.

21.

Siddique

, Sidike

, Elkin

, Devabhaktuni

, U-Net and its variants for medical image segmentation: Theory and applications, IEEE Access (2021), pp. 1–1.

22.

Soomro

T.A.

, Afifi

A.J.

, Gao

, Hellwich

, Paul

, Zheng

, Strided U-Net Model: Retinal Vessels Segmentation using Dice Loss, 2018 Digital Image Computing: Techniques and Applications (DICTA) (2018), pp. 1–8.

23.

Nayal

, Raut

, Priyadarshinee

, Narkhede

B.E.

, Kazancoglu

, Narwane

, Exploring the role of artificial intelligence in managing agricultural supply chain risk to counter the impacts of the COVID-19 pandemic, The International Journal of Logistics Management 33(3) (2022),pp. 744–772.

24.

Tutsoy

, Pharmacological, Non-Pharmacological Policies and Mutation: An Artificial Intelligence Based Multi-Dimensional Policy Making Algorithm for Controlling the Casualties of the Pandemic Diseases, IEEE Transactions on Pattern Analysis and Machine Intelligence 44(12) (2022),pp. 9477–9488.

25.

Sharmila

V.J.

, Jemi

, Deep Learning Algorithm for COVID-19 Classification Using Chest XRay Images, Computational and Mathematical Methods in Medicine, 2021.

26.

Nguyen

, Kay

, Tan

, Yan

, Seng Ng

, Iyengar

, Peshock

, Jiang

, Deep Learning–Based COVID-19 Pneumonia Classification Using Chest CT Images: Model Generalizability, Frontiers in Artificial Intelligence, 2021.

27.

Chaddad

, Hassan

, Desrosiers

, Deep Radiomic Analysis for Predicting Coronavirus Disease in Computerized Tomography and X-Ray Images, IEEE Transactions on Neural Networks and Learning Systems 33(1) (2022),pp. 3–11.

28.

Sakib

, Tazrin

, Fouda

M.M.

, Fadlullah

Z.M.

, Guizani

, DL-CRC: Deep Learning-Based Chest Radiograph Classification for COVID-19 Detection: A Novel Approach, IEEE Access 8 (2020),pp. 171575–171589.

29.

Castiglione

, Vijayakumar

, Nappi

, Sadiq

, Umer

, COVID-19: Automatic Detection of the Novel Coronavirus Disease From CT Images Using an Optimized Convolutional Neural Network, IEEE Transactions on Industrial Informatics 17(9) (2021),pp. 6480–6488.

30.

Shah

P.M.

, Ullah

, Shah

, Gani

, Maple

, Wang

, Shahid

M.A.

, Ul Islam

, Deep GRU-CNN Model for COVID-19 Detection From Chest X-Rays Data, IEEE Access 10 (2022), 35094–35105.

31.

Abbas

, Abdelsamea

M.M.

, Gaber

M.M.

, 4S-DT: Self-Supervised Super Sample Decomposition for Transfer Learning With Application to COVID-19 Detection, IEEE Transactions on Neural Networks and Learning Systems 32(7) (2021),pp. 2798–2808.

32.

El-Sayed El-Kenawy

, Mirjalili

, Ibrahim

, Alrahmawy

, El-Said

, Zaki

R.M.

, Eid

M.M.

, Advanced Meta-Heuristics, Convolutional Neural Networks, and Feature Selectors for Efficient COVID-19 X-Ray Chest Image Classification, IEEE Access 9 (2021), 36019–36037.

33.

Poomani

, Sutha

, Soundar Kathavarayan

, Wiener filter based deep convolutional network approach for classification of satellite images, Journal of Ambient Intelligence and Humanized Computing 12(1), 2021.

34.

Ramadan

, Aly

, CU-Net: A New Improved Multi-Input Color U-Net Model for Skin Lesion Semantic Segmentation, IEEE Access 10 (2022), 15539–15564.

35.

Sushma

, Raghavendra

C.K.

, Prashanth

, CNN based U-Net with Modified Skip Connections for Colon Polyp Segmentation, 2021 5th International Conference on Computing Methodologies and Communication (ICCMC), (2021), pp. 1762–1766.

36.

Mirjalili

, Moth-Flame Optimization Algorithm: A Novel Nature-inspired Heuristic Paradigm, Knowledge-Based Systems, 2015.

37.

Theerthagiri

, Jacob

, Ruby

A.U.

, Yendapalli

, Prediction of COVID-19 Possibilities using KNN Classification Algorithm, Available Online, 2020.

38.

Gampala

, Rathan

, S

C.N.

, Shajin

F.H.

, Rajesh

, Diagnosis of COVID-19 patients by adapting hyper parametertuned deep belief network using hosted cuckoo optimization algorithm, Electromagn Biol Med 41(3) (2022),pp. 257–271.

39.

Aslan

M.F.

, Sabanci

, Durdu

, Unlersen

M.F.

, COVID-19 diagnosis using state-of-the-art CNN architecture features and Bayesian Optimization, Comput Biol Med 142, 2022.

40.

Dhiman

, Vinoth

, Kumar, A. Kaur and A. Sharma, DON: Deep Learning and Optimization-Based Framework for Detection of Novel Coronavirus Disease Using X-ray Images, Interdiscip Sci 13(2) (2021),pp. 260–272.

41.

Mirjalili

, Lewis

, The Whale Optimization Algorithm, Advances in Engineering Software 95 (2016),pp. 51–67.

42.

Mirjalili

, Mirjalili

S.M.

, Lewisa

, Grey Wolf Optimizer, Advances in Engineering Software 69 (2014),pp. 46–61.

43.

Haennah

J.H.J.

, Christopher

C.S.

, King

G.R.G.

, Prediction of the COVID disease using lung CT images by Deep Learning algorithm: DETS-optimized Resnet 101 classifier, Front. Med 10, 2023.

44.

Jensha Haennah

J.H.

, Seldev Christopher

Dr. , Gnana King

G.R.

Dr. , A Novel Deep Learning Algorithm for COVID-19 Diagnosis with Feature Fusion Vector, In press.