key: cord-0716224-il51tzuy
authors: El-Dahshan, El-Sayed.A.; Bassiouni, Mahmoud.M.; Hagag, Ahmed; Chakrabortty, Ripon K.; Loh, Huiwen.; Acharya, Rajendra U.
title: RESCOVIDTCNnet: A residual neural network-based framework for COVID-19 detection using TCN and EWT with chest X-ray images
date: 2022-04-28
journal: Expert Syst Appl
DOI: 10.1016/j.eswa.2022.117410
sha: df2efe0d006527f6ef9b7f80b380fa19db700c87
doc_id: 716224
cord_uid: il51tzuy

Since the emergence of COVID-19, there has been an exponential surge in the number of casualties which increases the demand for numerous research works that can successfully detect the disease accurately in the early stage. This study provides some methods based on deep learning for the diagnosis of patients suffering from COVID disease, healthy controls, and pneumonia classes using chest X-rays. The methodology consists of four main phases: data acquisition, pre-processing, feature extraction, and classification. The chest X-rays images used in this study were obtained from various publicly available databases. In the pre-processing step, the images were filtered to improve the image quality, and empirical wavelet transform (EWT) was used to de-noise the chest X-ray images. Next, feature extraction via four deep learning models was attempted. The first two models are based on transfer learning models: Inception-V3 and Resnet-50. The third model is developed by combining the Resnet-50 with temporal convolutional neural network (TCN). The fourth model is our proposed model known as RESCOVIDTCNNet which combines EWT with Resnet-50 and TCN. Finally, the classification was performed by artificial neural network (ANN) and support vector machine (SVM). Our proposed RESCOVIDTCNNet has yielded an accuracy of 99.5% using five-fold cross-validation for 3-class classification. Our prototype has the potential to be used in underdeveloped countries where there is an acute shortage of radiologists to obtain the diagnosis immediately.

A new kind of viral pneumonia, coronavirus 2 (SARS-CoV-2) caused by severe acute respiratory syndrome was discovered in China [1] . The World Health Organization (WHO) has given an official name to the virus, COVID-19, in February 2020, and declared it a pandemic in March 2020. Globally, as of 15 th October 2021, there are over 239 million confirmed cases of COVID-19 and 4.8 million deaths [2] . Moreover, many individuals may have the COVID-19 infection but are asymptomatic. These individuals do not display symptoms of the COVID-19 disease, instead, they are the carrier of the virus and are capable of spreading it to people who are vulnerable to it [3] . Therefore, early diagnosis of the disease, even in the absence of symptoms, can control the spread of the virus and save the patient's life.

Several indicators can help to diagnose the patient's health status. The most common test for COVID-19 detection is the Reverse Transcription Polymerase Chain Reaction (RT-PCR) [4] .

However, RT-PCR is relatively low in sensitivity and time-consuming. Moreover, it is costly and requires specific materials and equipment, which are not easily accessible. Alternatively, various types of radiological imaging exist such as X-rays and the computed tomography (CT) to identify patients affected by pneumonia due to COVID-19 infection. It is reported that COVID-19 patients' lungs exhibit some visual features, such as markings and spots, that may distinguish COVID-19 positive cases from normal cases using radiological images [5] .

Unlike RT-PCR and CT, chest X-ray is cheaper, less time consuming, and it is readily available for COVID-19 screening. The X-ray has lower ionizing radiations than CT scans, which allows for multiple follow-ups on the effects of COVID-19 on lung tissue. Therefore, in this paper, chest X-ray images gathered from different databases have been used for detection of the COVID-19 disease. Sometimes noises appear in the X-ray images, which affects the diagnosis. To overcome this challenge, time-frequency analysis can be used to remove high-frequency components from the noisy images. Also, the image preprocessing step helps to further improve the model performance.

The danger of COVID-19 disease and its spread prompted researchers to develop many automatic diagnostic methods using X-ray images. Many traditional machine learning techniques have been presented in the literature for the early diagnosis of COVID-19 [6] [7] [8] [9] [10] [11] [12] .

The convolutional neural network (CNN) [13] , support vector machine (SVM) [14] , residual exemplar local binary pattern (ResExLBP), iterative ReliefF [7] , and Sobel filter [15] have been used. Moreover, most of the previous methods used deep learning networks to achieve good classification results. Ozturk et al. [16] present a deep learning structure based on a pretrained model known as DarkCovidNet and their proposed model is defined as classifier. The former models were used to diagnose X-ray lung images infected with COVID-19. This model used the Darknet-19 classifier [17] instead of building a model from scratch. Their proposed model was tested with 3 different classes such as COVID cases, normal, and pneumonia. The number of X-ray images are 1000 and they have obtained 98.08% and 87.02% accuracy rates for binary (i.e., normal and COVID-19) and three classes (i.e., normal, COVID-19, and Pneumonia) classifications, respectively. The proposed approach in [18, 19] used capsule networks to detect COVID-19 disease using X-ray images.

Recently, many transfer learning methods have been implemented for detection of COVID-19 using X-ray images as well [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] . The researchers in [23] used the Xception model to present an automated deep transfer learning method for the detection of COVID-19 infection from chest X-rays. They used 500 normal, 500 pneumonia, and 125 COVID-19 X-ray images in their study and achieved an classification accuracy of 97.4%. Ioannis et al. [20] reported an accuracy of 93.48% on test data by applying transfer learning models using a VGG-19 pretrained model for multi-classification diagnosis. In [25, 27] , multi-CNN comprising of several pre-trained CNNs is used to extract the features from chest X-ray images. Das et al. [30] developed an automated Covid-19 diagnostic method based on the VGG-16 pre-trained model. This method gave an accuracy of 97.67% on 2,905 chest X-rays images. Despite all these research works, the main research gaps in the automated detection of COVID-19 are as follows:

i. The accuracy of diagnosis still needs to be improved by reducing the misclassification.

ii. Most of the related deep learning models did not give enough attention to the preprocessing stage.

iii.

The architecture of the pre-trained models is not able to provide high optimal diagnostic results.

Therefore, in this study, we developed a diagnostic method for the automated detection of COVID-19 infection using chest X-rays using transfer learning method. The main contributions of this paper are given below:

 Investigation of various deep learning models for the automated detection of COVID- 19, pneumonia, and normal X-ray chest images.  Development of a novel model defined as RESCOVIDTCNNet using a combination of EWT, Resnet50, and temporal convolutional neural network (TCN).  Model is developed using more than 5000 X-ray chest images and reported the highest performance using the proposed model.  Application of empirical wavelet transform (EWT) to preprocess the chest X-ray images.

The novelty of the proposed RESCOVIDTCNNet model is in the combination of EWT with Resnet-50 and TCN. First, chest X-rays images are collected from different sources in the data acquisition step. Second, EWT is used in pre-processing to improve the resolution of the input images. Third, feature extraction is proposed based on two pre-trained models:

Inception-V3 and Resnet-50. Finally, artificial neural network (ANN) and SVM are applied in the classification step. The main reason for adding all these methods is to improve the classification accuracy. The combination between EWT and Resnet-50 helps to improve the quality of the X-rays images. The combination of Resnet-50 and Inception-v3 pre-trained models contribute to the feature extraction step. TCN was further applied to the extracted features to obtain more salient features present in the X-ray images. Recently, several studies were adapted to diagnose COVID diseases based on chest X-ray images. Transfer learning approaches relying on (CNN) can be used for classification, feature extraction [7, 8, 32, 33] , and transfer learning [24, 25, [28] [29] [30] 34] . The summary of studies developed for the automated detection of COVID-19 using X-ray images is summarized in Table 1 . It can be realized From Table 1 that, there is still room for improving the classification performance of the model using huge diverse datasets. Hence, we have proposed a novel RESCOVIDTCNNet to classify three classes using X-ray images reaching the maximum diagnosis accuracy. 

The methodology consists of four main phases and these phases are essential for the construction of the proposed deep learning models as shown in Figure. 1. The first phase is the data acquisition phase, and in this step, four main datasets are selected containing X-ray chest images of COVID, normal, and pneumonia. In the second filtration phase, the images are pre-processed using empirical wavelet transform (EWT). Then in the feature extraction step, four main deep learning models: Inception-V3, Resnet50, Resnet50-TCN, and RESCOVIDTCNNet are used. Finally, these extracted features are classified using multilayer perceptron (MLP) and support vector machine (SVM) classifiers.

This section presents various lung X-ray images gathered from publicly available datasets required for COVID-19 diagnosis. We have taken care of two image selection criteria in this work: (i) balance the X-ray images number in each class and (ii) not to choose the same image from different datasets twice. The first dataset used in this work was obtained by Muhammed Talo [16] and available on the Kaggle portal [49] . The dataset in [16] was obtained from two different databases. The first database was collected by Joseph Paul Cohen [40] , called COVID-19 X-ray image, this database consists of 125 COVID-19 images. The second database was provided by Wang et al. [41] , called ChestX-ray8, and this database comprises of 500 normal and pneumonia X-ray images. In this study, we have used 125 COVID-19 images from the COVID-19 X-ray image database, 329 normal, and 325 pneumonia from the ChestX-ray8 database. The second dataset was obtained from Paul Mooney [44] which iscalled Chest X-Ray images and it consists of 1592 normal and 4273 pneumonia images. The number of images selected for this study from this database is 1343 normal and 1345 pneumonia images. The third dataset was obtained from Tawsifur Rahman [42, 43] and is available on the Kaggle portal [50] . This dataset is called COVQU. The main aim of this dataset is to obtain the images of the COVID-19 class. This dataset consists of 3616 COVID-19, 10192 normal, and 1345 viral pneumonia images. 

In this paper, we used EWT [51] to filter and pre-process the X-ray images. EWT is similar to wavelet transform, in which the transformation based on the EWT results with two main sets which are details and the approximation coefficients. Let's consider a signal; the detail coefficients will be defined based on the convolution operation of the signal with EWT to obtain , whereas the approximation coefficients will be known using the convolution of the signal using a scaling function defined as . The details and the approximation coefficients ∅ are defined in the following equations respectively.

The previous equations are considered to be the details and the approximation of the frequency components obtained from the EWT. Extension of EWT for the 2D image can be defined as 2D Littlewood-Paley EWT, 2D Curvelet EWT, 2D Empirical Ridgelet transform, and 2D tensor EWT. In this study, 2D Littlewood-Paley EWT is applied because it aims at the construction of little-Paley wavelet filters in separate scales defined by various concentric rings. It detects the scales first and then detects the angular section with each scale ring.

Let's consider to be an image, then the details and the approximation coefficients of 2D Littlewood-Paley EWT are defined using the following equations:

where and are the 2D pseudo-polar Fourier transform and its inverse. Then EWT is * 2 2

applied to each band of the image.

In this study, ResNet-50 and Inception-v3 pre-trained CNN architectures are employed to classify the X-ray chest images into three classes: Normal, COVID-19, and Pneumonia. with various convolution kernels to improve convergence performance. The fully connected layer is used to convert multi-scale feature vectors into a one-dimensional vector, followed by 1×1 Inception modules. Finally, the Softmax classifier generates one vector with three classes probability (i.e., Normal, COVID-19, and Pneumonia).

Due to their superior capability to capture temporal relationships in sequential data, recurrent neural networks (RNNs) were the preferred choice for sequence modeling tasks. The RNNs, such as long short-term memory (LSTM) [55] , convolutional LSTM (ConvLSTM) [56] ,

WaveNet [57] , and gated recurrent unit (GRU) [58] , can capture long-term relationships in sequences and have attained state-of-the-art results in sequence modeling tasks. The TCN model [59] is a modification of CNN model designed for sequence modeling tasks with causal constraints. In several sequence modeling tasks, the TCN outperformed the RNN and its derivatives LSTM [60] and GRU [61] . Furthermore, the architecture of TCN is more simple and straightforward, as compared to RNN and LSTM networks. In addition, the TCN, like standard CNNs, offer the advantages of convolution kernel sharing and parallel computation, which helps to minimize the calculation time. Also, it can take a sequence of variable length and map it to an output sequence of the same length using a 1D fullyconvolutional network (FCN) architecture [62] . In addition, the convolutions in the architecture of TCN are causal convolutions. Fig. 3 shows the typical architecture of TCN.

The residual network is used in the TCN to avoid the deterioration in CNN performance when the number of layers gets very large. The TCN layer consists of residual blocks, each of which has two layers of dilated causal convolution. In addition, weight normalization is applied to the convolutional filters with the rectified linear unit (ReLU) [63] as the activation function in a residual block. The ReLU activation function must be used because TCN has two dilated causal convolutions and non-linearities. For regularization, a dropout layer is introduced after each dilated convolution. In TCN, the input and outputs can have different widths. Therefore, a 1×1 convolution is used to account for discrepant input-output widths.

The dilated convolution adds certain weights to the convolution kernel while leaving the input data intact, increasing the size of the time series viewed by the network while keeping the amount of calculation relatively constant. Let an input sequence , the { 0 , 1 ,…, } ∈ ℝ dilated convolution operation on the -the element of the 1D sequence [64] is defined as

where indicates the filter with size , the dilation -coefficient is , and :{0, 1, …, -1}→ℝ accounts for the direction of the past. As a result, dilation is the same as using a set -• step size between each pair of adjacent filter taps. The extended convolution is simplified to a standard convolution with . The output of the top layer may reflect a wider variety of = 1 inputs with a bigger expansion, thereby increasing TCN's receptive field. TCN can extend the receptive field in two ways: by raising the dilation factor or by selecting bigger filter sizes . As an example Fig. 3 are defined in the following sequence . The corresponding output that is going to 0 , 1 ,…, be classified is defined as . The overall structure of the proposed 0 , 1 ,…, RESCOVIDTCNNet is shown in Fig. 4 and the pseudocode is provided in algorithm.1. TCN aims to satisfy two main constraints [65] : (i) output of the TCN network should be equal to the length of the input and (ii) it depends on the previous information obtained from the previous steps. TCN residual blocks are based on casual convolution in which each layer at an output time is calculated with the region no later than time step in the past layer. The causal convolutional layers have a problem of limited respective size and to store a large amount of information it is required to stack many layers. Therefore, dilated convolution is applied to permit an exponentially large receptive field. The dilated convolutional layers structure is shown in Fig. 4 (b) . When the dilated convolutions increase with a dilated factor in an exponent way, this leads to an increase in the size of the receptive field to cover a large number of inputs in the history.

The structure of the residual block is shown in Fig. 4 (c) . Finally, the RESCOVIDTCNNet model consists of 4 main blocks. Each block consists of two dilated causal convolutional layers, two weight normalization layers, two RELU activation functions. The result of each block is an input to the next block. The four blocks are executed and then the whole model is ended by two main fully connected layers and a classification layer.

This section presents the results of the proposed methodologies obtained in various subsections. The first subsection presents the setup of the experiments and the software required to perform the de-noising, feature extraction, and classification. The second subsection indicates the results of the filtration process using EWT. The third subsection indicates the training parameters required for the deep learning training models. Finally, the last subsection denotes the quantitative analysis of the performance measurements obtained from various deep learning models.

The presented deep learning models and the classifiers used in this work were implemented using MATLAB software programming language. The classification stage and the performance measurements were obtained using WEKA software. The experiments were performed on a PC with Intel COREi9-10232 @ 1.80 GHz 1.99 GHz, 32 GB RAM. The total number of X-ray images (5012) was divided into three parts: training (60%), validation (20%), and test (20%). five-fold cross-validation (CV) strategy was used to assess the model performance. The classification was performed using two MLP and SVM classifiers to choose the best performing classifier.

EWT showed an efficient performance on the original X-ray chest images. It improved the performance of the original image to reach the SNR of 15.43 dB. In addition to this, to clarify the performance of the EWT, salt and pepper noise with variance of 0.05 is added to the original image. The variance is the amount of salt and pepper noise added to the X-ray lung images. The reconstructed image showed better quality and it removed the added noise and yielded the SNR of 19.322 dB. Figure 5 shows the performance of EWT (a) shows the result of the EWT on the original X-ray image, while (b) shows the results of the EWT on the salt and pepper X-ray image. 

The parameters considered during hyperparameter tuning of the deep learning models are shown in Table 3 . These parameters include the optimizer, momentum, learn rate (drop factor, schedule, and drop period), initially learn rate, L2 regularization, gradient threshold method and value, maximum epochs, and mini-batch size. The optimal parameters used for the deep learning models are determined experimentally [66, 67] .

The optimizer used during the training is the stochastic gradient descent with momentum (SGDM) and Adam's. SGDM operates on a random selection of data examples. The momentum value used in our study for the SGDM is 0.9 which helps to update the step of the previous iteration to the current iteration. The learning rate has three main options which are the schedule, period, and factor. The schedule is defined to be 'piecewise' in which the learning rate updates every certain epoch by a certain learning rate factor, while the learn rate drop period is considered to be the number of epochs for dropping the learning rate. The mini-batch size is a parameter that obtains a subset of a training set, and it can be used to evaluate the gradient of the loss function to update its weights. Another type of parameter is specified for the progress of the training process and its iterations such as the verbose and verbose frequency. The verbose is defined by 1, and the verbose frequency is defined by 50.

Another two parameters are required for the training of the models are the gradient threshold and the gradient threshold method. The gradient threshold is a method that is used to clip the values of the gradient that exceeded the gradient threshold. The gradient threshold specified in our experiment is Inf and the threshold method is the L2norm. The main aim of this parameter is to force the weights to be small, but it does not make them equal to zero, and non-sparse solution. The RESCOVIDTCNNet is trained with the same parameters as that of

Resnet50-TCN except that the X-ray images are fed to EWT before feeding to the deep learning model. 

To determine the performance of the proposed methods, quantitative and qualitative analysis is applied. They are briefly explained below.

In were drawn for each fold to provide a visual representation of the performance [64] . Table 4 shows the performance obtained for five-folds of the cross-validation obtained for Inception v3 and Resnet50 deep learning models obtained using MLP and SVM classifiers. It can be seen that the performance of the MLP classifier using Inception v3 model is higher than the SVM for 1, 2, 4, and 5 folds, while the SVM classifier performed higher than MLP during fold 3. On the other hand, it can be noted that MLP with Resnet50 performed higher than the SVM for 1, 3, 4, and 5 folds, whereas the performance of SVM is higher than MLP for fold2 only. Table 5 shows the performance measures obtained using five-folds with MLP and SVM classifiers for Resnet50-TCN and our proposed RESCOVIDTCNNet. It can be seen that the performance of MLP classifier with Resnet50 model yielded better results than the SVM for folds 1, 2, and 3, while the SVM classifier performed better than MLP for folds 4 and 5. Fig. 7 (a) and (b) indicate the accuracy and the precision obtained using five-fold cross-using validation for various deep learning models, (c) and (d) shows the recall and F1-score obtained using five-fold cross-validation strategy for various deep learning models. On the other hand, it can be noted from Table 7 that MLP performance with Resnet50 is the same as that of SVM for folds 1 and 3, whereas the performance of MLP is higher than SVM for folds 2, 3, and 5. Recall and the F1-score of the deep learning models, respectively. It can be seen that the proposed model has the highest performance than the rest of the models. Finally, the confusion matrix of the five folds is represented in Fig. 9 . The confusion matrices obtained for each of the four deep learning models are shown in Figs. 9 (a, b, c, and d) . The green cells in each matrix represent the number of correctly classified instances during the five-fold cross-validation strategy for each of the three classes, while the red cells represent the misclassified instances. It can be noted from the confusion matrices that the Inception V3

showed the lowest accuracy performance based on MLP, while the proposed RESCOVIDTCNNet achieved the highest accuracy performance with less number of misclassified instances in each class. 

This part presents the details of the folds and accuracy distribution in each fold of the deep learning models. The main aim of the qualitative analysis is to provide a good understanding of the results obtained by the deep learning models. Therefore, three different diagrams are illustrated to present the accuracies and the detailed results of the deep learning models. The first diagram is the Taylor diagram which is designed to indicate several appropriate representations of the systems [68] . It provides a statistical summary of how the models or the system match each other in terms of their correlation and the standard deviation. Figure   10 shows the Taylor diagram designed for the deep learning models. It can be seen from the figure that, the highest prediction performance can be realized from the RESCOVIDTCNNet as it has a correlation coefficient of 0.999 and a standard deviation of 0.1, while Inception-V3

has the lowest performance with a correlation coefficient of 0.971 and a standard deviation of 19 . The second diagram presented is the box plot diagram which provides a better indication of data spread. In our study, it is used to represent how the accuracies of the folds are spread out using the deep learning models with classifiers. This diagram is drawn for the deep learning model to represent the results of the five-fold cross-validation. In other words, Fig. 11 (a) shows the results of the deep learning models obtained from the five-fold cross-validation using ANN (MLP) classifier, while Fig. 11 (b) is the results obtained with the SVM classifier. The third diagram is the spider plot which presents a visual tool that can be used to organize the data logically ( Figure 12 ). It is also defined as a tool that organizes concepts based on space, color, and images. This diagram is drawn in a form of a diamond shape to represent the performance of the deep learning models. Figure 12 

The complexity of four deep learning models depends on several factors: processing or the learning time taken during the training of each of the deep learning models. In addition to this, the number of layers and filters in each layer can result in variances in the complexity between different deep learning models. In our study, the complexity is evaluated based on the learning time, features extraction from the fully connected layer, and classification of test data in each fold. We have employed five-fold cross-validation, whereby the training time will be multiplied by 5. 

This paper presents various deep learning methodologies used for the classification of three classes: COVID-19, normal, and pneumonia using X-ray images. The methodology consists of four main stages: data acquisition, pre-processing, feature extraction, and classification.

During the data acquisition phase, 5012 X-ray chest images (1670 COVID, 1672 normal, and 1670 pneumonia). In the pre-processing stage, the images are filtered using empirical wavelet transform (EWT).

It is important to know that the X-ray images do not contain noises, but the EWT improves the quality and hence enhances the classification accuracy of the proposed model. The next stage is the feature extraction and in this stage, two main pre-trained models InceptionV3 and Resnet50 are used. It was observed that the highest classification accuracy was obtained when fused with the temporal convolutional neural network (TCN). The Resnet50 showed a higher performance than Inception V3, therefore, it was chosen to fuse with TCN (Resnet50-TCN), also known as RESCOVIDTCNNet. As a result, the classification accuracy has improved from 97% to 99%. Finally, the last classification stage involves two main classifiers, namely ANN (MLP) and SVM, and the highest classification performance was achieved using MLP.

Advantages of the proposed model:

1. Enhanced the quality of the X-ray images using EWT algorithm to reach higher SNR.

2. Improved the diagnosis performance of the novel RESCOVIDTCNNet using transfer learning models.

3. Achieved the highest classification performance in comparison with the other related works developed using Chest X-ray images.

The limitation of our work is that the number of X-ray images used to develop the model is not large enough. Also, our proposed model misclassified 30 images from all folds.

In this work, deep learning models are presented based on the CNN framework to recognize the symptoms of COVID, normal, and pneumonia automatically. The proposed methodology consists of four main phases: data acquisition, pre-processing, feature extraction, and classification. In the first phase, the database belonging to healthy, pneumonia, and COVID chest X-ray images of both genders and different age groups are used. In the second phase, the X-ray images are filtered using EWT which yielded an efficient SNR value on the enhancement of the original X-ray images. In the third phase, two transfer learning models: The main limitation of this work is that, the number of images used in each class is not huge.

Hence, in the future, we plan to validate our generated model using more images taken from diverse races and age groups. Also, we intend to test our proposed system to detect other lung-related diseases like chronic obstructive pulmonary disease (COPD), asthma, and lung cancer.

Outbreak of pneumonia of unknown etiology in Wuhan, China: The mystery and the miracle

WHO situation report

Asymptomatic transmission during the coronavirus disease 2019 pandemic and implications for public health strategies

A new coronavirus associated with human respiratory disease in China

CovH2SD: A COVID-19 detection approach based on Harris Hawks Optimization and stacked deep learning

Improving the performance of CNN to predict the likelihood of COVID-19 using chest X-ray images with preprocessing algorithms

A deep-learningbased framework for severity assessment of COVID-19 with CT images

Fusion of convolution neural network, support vector machine and Sobel filter for accurate detection of COVID-19 patients using X-ray images

Automatic method for classifying COVID-19 patients based on chest X-ray images, using deep features and PSO-optimized XGBoost

Applications of artificial intelligence in COVID-19 pandemic: A comprehensive review

COVID-19 Detection from X-ray Images using Multi-Kernel-Size Spatial-Channel Attention Network

Learning distinctive filters for COVID-19 detection from chest X-ray using shuffled residual CNN

Deep learning. nature

Support-vector networks

A 3x3 isotropic gradient operator for image processing, a talk at the Stanford Artificial Project in

Automated detection of COVID-19 cases using deep neural networks with X-ray images

YOLO9000: better, faster, stronger

Convolutional capsnet: A novel artificial neural network approach to detect COVID-19 disease from X-ray images using capsule networks

Covid-caps: A capsule network-based framework for identification of covid-19 cases from x-ray images

Covid-19: automatic detection from xray images utilizing transfer learning with convolutional neural networks

Classification of COVID-19 from Chest X-ray images using Deep Convolutional Neural Network

Explainable deep learning for pulmonary disease and coronavirus COVID-19 detection from X-rays

Automated deep transfer learning-based approach for detection of COVID-19 infection in chest X-rays

Covid-19 screening on chest x-ray images using deep learning based anomaly detection

Computer-aided detection of COVID-19 from X-ray images using multi-CNN and Bayesnet classifier

A deep learning approach to detect Covid-19 coronavirus with X-ray images

Deep learning approaches for COVID-19 detection based on chest X-ray images

Hybrid ensemble model for differential diagnosis between COVID-19 and common viral pneumonia by chest X-ray radiograph

DenseCapsNet: Detection of COVID-19 from X-ray images using a capsule neural network

TLCoV-An automated Covid-19 screening model using Transfer Learning from chest X-ray images

Coronavirus disease (COVID-19) detection using X-ray images and enhanced DenseNet

Deep-covid: Predicting covid-19 from chest x-ray images using deep transfer learning

DeepCoroNet: A deep LSTM approach for automated detection of COVID-19 cases from chest X-ray images

A new composite approach for COVID-19 detection in X-ray images using deep features

Covid-net: A tailored deep convolutional neural network design for detection of covid-19 cases from chest x-ray images

Covidx-net: A framework of deep learning classifiers to diagnose covid-19 in x-ray images

Automatic detection of coronavirus disease (covid-19) using x-ray images and deep convolutional neural networks

Estimating uncertainty and interpretability in deep learning for coronavirus (COVID-19) detection

A deep learning framework for coronavirus disease (covid-19) detection in x-ray images

COVID-19 Image Data Collection

Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases

Can AI help in screening viral and COVID-19 pneumonia

Exploring the effect of image enhancement techniques on COVID-19 detection using chest X-ray images

Chest X-Ray Images (Pneumonia)

Identifying medical diagnoses and treatable diseases by image-based deep learning

Corona hack: chest X-ray-dataset Available

Study of different deep learning approach with explainable ai for screening patients with COVID-19 symptoms: Using ct scan and chest x-ray image dataset

COVID-19 detection from chest X-ray images using CNNs models: Further evidence from Deep Transfer Learning

COVID-19 Dataset

COVID-19 Radiography Database

Empirical wavelet transform

Deep residual learning for image recognition

Imagenet: A large-scale hierarchical image database

Rethinking the inception architecture for computer vision

Long short-term memory

Convolutional LSTM network: A machine learning approach for precipitation nowcasting

Wavenet: A generative model for raw audio

Learning phrase representations using RNN encoder-decoder for statistical machine translation

An empirical evaluation of generic convolutional and recurrent networks for sequence modeling

Benchmarking of LSTM networks

A deep bidirectional GRU network model for biometric electrocardiogram classification based on recurrent neural networks

Fully convolutional networks for semantic segmentation

Rectified linear units improve restricted boltzmann machines

Temporal convolutional networks for the advance prediction of ENSO

STCN: Stochastic temporal convolutional networks

A Self-Activated CNN Approach for Multi-Class Chest-Related COVID-19 Detection

Multi-classification of brain tumor images using deep neural network

Taylor diagram primer

Dear Prof. Dr. Binshan Lin Editor-in-Chief of Expert Systems With Applications "ESWA" journal.We are submitting the manuscript entitled "RESCOVIDTCNnet: A residual neural network-based framework for COVID-19 detection using TCN and EWT with chest X-ray image" to ESWA possible publication. There is no conflict of interest in this work.