key: cord-0861764-f096ndan
authors: Jalali, Seyed Mohammad Jafar; Ahmadian, Milad; Ahmadian, Sajad; Hedjam, Rachid; Khosravi, Abbas; Nahavandi, Saeid
title: X-ray image based COVID-19 detection using evolutionary deep learning approach
date: 2022-03-30
journal: Expert Syst Appl
DOI: 10.1016/j.eswa.2022.116942
sha: 124abb78ba82db639f63886ca077fe31b2c838fc
doc_id: 861764
cord_uid: f096ndan

Radiological methodologies, such as chest x-rays and CT, are widely employed to help diagnose and monitor COVID-19 disease. COVID-19 displays certain radiological patterns easily detectable by X-rays of the chest. Therefore, radiologists can investigate these patterns for detecting coronavirus disease. However, this task is time-consuming and needs lots of trial and error. One of the main solutions to resolve this issue is to apply intelligent techniques such as deep learning (DL) models to automatically analyze the chest X-rays. Nevertheless, fine-tuning of architecture and hyperparameters of DL models is a complex and time-consuming procedure. In this paper, we propose an effective method to detect COVID-19 disease by applying convolutional neural network (CNN) to the chest X-ray images. To improve the accuracy of the proposed method, the last Softmax CNN layer is replaced with a [Formula: see text]-nearest neighbors (KNN) classifier which takes into account the agreement of the neighborhood labeling. Moreover, we develop a novel evolutionary algorithm by improving the basic version of competitive swarm optimizer. To this end, three powerful evolutionary operators: Cauchy Mutation (CM), Evolutionary Boundary Constraint Handling (EBCH), and tent chaotic map are incorporated into the search process of the proposed evolutionary algorithm to speed up its convergence and make an excellent balance between exploration and exploitation phases. Then, the proposed evolutionary algorithm is used to automatically achieve the optimal values of CNN’s hyperparameters leading to a significant improvement in the classification accuracy of the proposed method. Comprehensive comparative results reveal that compared with current models in the literature, the proposed method performs significantly more efficient.

At the end of 2019, in Wuhan, China, the first outbreak of novel coronavirus disease occurred. The COVID-19 virus attacks and mutates rapidly in the lungs of a confirmed patient. The tainted lungs are inflamed and overflowing with fluid in such a situation. When we un-5 dertake CT-Scan or X-ray images of an infected person, the results reveal dark spots in the lungs named Ground Glass Opacity (Parekh et al. 2020 ).

The spreading level of COVID-19 disease is significantly higher than its forecasting or detection speed according to its highly contagious existence (Ismael & Şengür 2020) . In the failure of an intelligent diagnostic tool, it 10 is important to diagnose suspected COVID-19 patients quickly and accurately (Panwar, Gupta, Siddiqui, Morales-Menendez, Bhardwaj, & Singh 2020 , Chakraborty & Mali 2020 . The current techniques available to diagnose this spreading pandemic are slightly accurate and time-consuming (Shi et al. 2020 , Alamoodi et al. 2020 ). There are generally three major ognized as time-consuming approach. CT-Scan can detect inflammation of the lungs in terms of location, shape, and scale. CXR offers a more ac-20 curate way of diagnosis for COVID-19 and a clear image of air sacs. Thus, we have chosen CXR images of the lungs for the experiments conducted in this work (Panwar, Gupta, Siddiqui, Morales-Menendez, Bhardwaj, & Singh 2020 , Pereira et al. 2020 ).

Medical institutions produce and collect large volumes of data that provide 25 immensely helpful patterns and information, which go way further than what traditional analytical methods can handle. It should be noted that the RT-PCR test is one of the expensive and time-consuming approaches for the identification of suspects of COVID-19 (Kashir & Yaqinuddin 2020 2021). Therefore, a better approach could be found in which a combination of deep learning classifiers and clinical images provides rapid and reliable identification of the COVID-19 virus throughout CXR analysis of pulmonary images. Deep Learning is a machine learning category that facilitates computer systems to learn important data features automati-35 cally from a set of data gathered for different applications such as machine vision (Yadav & Jadhav 2019 , Pouyanfar et al. 2018 ) and recommender systems (Tahmasebi et al. 2021 , Ahmadian et al. 2014 , Moradi et al. 2016 , Ahmadian, Meghdadi, & Afsharchi 2018a , Rahmani et al. 2019 , Ahmadian, Meghdadi, & Afsharchi 2018b , Ahmadian, 40 Joorabloo, et al. 2018 , Ahmadian, Ahmadian, & Jalili 2021 . Convolutional neural network (CNN) is one of the popular deep learning models which has been widely used in different research area such as medial applications. CNN can take input in 2D or 3D images into consideration and make proper utilization of spatial and features information (Yuan et al. the chaotic salp swarm algorithm is utilized to optimize the feature matrix. Then, the EfficientNet-B0 model is used as a deep learning approach to detect COVID-19 cases. Karasu et al. (Karasu et al. 2017) proposed Hu et al. (Hu et al. 2022 ) proposed a novel framework for COVID-19 detection by integrating classification and regression tasks in a multi-task 70 multi-modality support vector machine approach. Accordingly, a feature extraction method is applied to the chest CT images to obtain appropriate features for each segment of input image. Then, an effective model is considered based on data augmentation of over-sampling approach to address class imbalance problem. Finally, the proposed support vector ma-75 chine model is utilized to detect COVID-19 cases. In (Deb et al. 2022 Although the deep neural network models have gained considerable accu-85 racy in image classification applications, the design of these models has been mainly done manually, which is a time-consuming and complex process. It is worth mentioning that deep neural networks contain several hyperparameters which determining their values is a challenging task. Also, the efficiency of these models is strongly dependent on the values used 90 for their hyperparameters. Therefore, developing effective methods to automatically determine the values of these hyperparameters is a critical issue to improve the performance of deep neural networks (Jalali, Khosravi, Alizadehsani, et al. 2019 , Ahmadian & Khanteymoori 2015 , Jalali, Ahmadian, et al. 2019 , Mousavirad et al. 2020 2020, Khodayar et al. 2021 , Jalali, Osorio, et al. 2021 . Such methods should automatically provide the optimal structure of deep neural networks leading to obtain the most accuracy in the shortest time. Deep Neuroevolution (DNE) is a remarkably effective and feasible approach to obtain automatically the architectures of deep neural networks using pow-100 erful evolutionary algorithms. The aim of this approach is to select the best values for hyperparameters leading to promote accuracy, minimize the network over-fitting, and promote consistency. However, evolutionary algorithms mainly suffer from low convergence speed and tapping into local optima. Moreover, making a balance between the exploration and 105 exploitation phases of these algorithms plays a critical role in enhancing their ability of search process (Jalali et al. 2022 , Ahmadian, Jalali, Raziani, & Chalechale 2021 J o u r n a l P r e -p r o o f Journal Pre-proof Hedjam, et al. 2020 2020).

In this paper, an effective image classification method is proposed based on deep CNN model to detect COVID-19 disease from chest X-ray images.

In order to improve the classification accuracy of the proposed method, the last Softmax CNN layer is replaced with a KNN classifier to take 115 into account the agreement of the neighborhood labeling. Moreover, an evolutionary algorithm is developed by improving the original version of competitive swarm optimizer (CSO) model (Cheng & Jin 2014) . To this end, three powerful evolutionary operators: Cauchy Mutation (CM), Evolutionary Boundary Constraint Handling (EBCH), and tent chaotic map 120 are incorporated into the search process of CSO model to improve its search capabilities, increase its convergence speed, and make an excellent balance between the exploration and exploitation phases. Then, the improved evolutionary algorithm is employed to derive the optimal architecture of deep CNN model for classifying X-ray images whether or not 125 they are affected by COVID-19 disease. Therefore, the proposed COVID-19 detection method takes much shorter time in designing of deep neural network compared to manual methods, and at the same time achieves higher classification accuracy. In summary, the principal contributions of this work are as follows:

• A novel image classification approach is developed by employing deep CNN model to identify patients whether or not they are infected with COVID-19 disease according to the chest X-ray images. In the proposed method, the last Softmax CNN layer is replaced with a KNN classifier to improve the classification accuracy by taking into account the agreement of the 135 neighborhood labeling.

• An effective evolutionary algorithm is proposed by considering three powerful evolutionary operators including Cauchy mutation, evolutionary boundary constraint handling, and tent chaotic map in the search process of the 7 J o u r n a l P r e -p r o o f Journal Pre-proof original version of competitive swarm optimizer model. The main advan-140 tage of the proposed evolutionary algorithm is to make a balance between the exploration and exploitation phases which results in increasing convergence speed and reducing the probability of falling into local optima.

• The proposed evolutionary algorithm is applied to automatically achieve the optimal values of hyperparameters of CNN model leading to a sig-145 nificant improvement in the accuracy of classification method. Therefore, different from most of the previous methods, the proposed method does not need to perform a manual trial and error process to obtain the appropriate values of CNN's hyperparameters.

• Extensive experiments are carried out to assess the effectiveness of the 150 proposed DNE algorithm. The results demonstrate that the proposed DNE method can detect COVID/NON-COVID cases more accurate and reliable than the other state-of-the-art models.

The remainder of this paper is structured accordingly. The literature review is discussed in Section 2. Section 3 outlines the entire proposed DNE 155 classification method for COVID-19 images. Section 4 explains the experimental procedures. Finally, Section 5 sets out the conclusion remarks and future directions.

Although the COVID-19 disease is a very new issue in the world, an enor- images. To this end, a graph diffusion model is developed using an optimization methodology to make pseudo-labels of input data for the inputs 220 of deep neural network model. In (Bhattacharyya et al. 2022) , the chest X-ray images are utilized as the input data to make an effective approach for diagnosing COVID-19 disease. In particular, a segmentation method is developed based on the conditional generative adversarial network to extract the lung area from the chest X-ray images. Then, a deep neural By investigating the approaches reviewed in this section, it can be concluded that all of them utilize a static design of deep neural networks for developing their classification models. In other words, these approaches adjust the hyperparameters of deep neural networks manually without considering any heuristic. Whilst, the performance of deep neural net-245 works is significantly dependent on the values selected for their hyperparameters. In addition, adjusting the values of these hyperparameters in a manual way is very time-consuming and might cause to reduce the efficiency of deep neural networks. Different from the previous works, we develop an automatic strategy in this paper to tune the hyperparame-250 ters of deep neural networks using an enhanced evolutionary algorithm. Therefore, the main advantage of the proposed method is to automatically determine the optimal values of the hyperparameters of deep neural networks leading to improve the performance of the proposed method in identifying COVID-19 cases. 

Deep learning is a technique using in machine learning, in which a deep architecture designs several linear and non-linear processor modules to represent high-level functionality in a given dataset. A wide range of ap- Let X with R A×B is used to determine the input in the convolutional layer and the input data size is defined as A and B. The output for the convolutional layer is then stated as follows:

where C n represents the mth output of convolutional layer, and n refers the widely-used activation function which is defined in the following form:

The pooling technique allows in reducing the algorithm complexity with ter performing the traversing operation, the updated feature maps will eventually be extracted. This is the entire process of the CNNs in general. image to a vector of real numbers, we used Euclidean distance feature in KNN as the most widely accepted distance mechanism. The Euclidean distance of KNN algorithm is formulated as following:

where x i and x j are the m−dimensional vectors of the i th and j th samples, and the index r indicates the r th real-valued feature of each sample. 

In order to solve large-scale optimization problems, (Cheng & Jin 2014) introduced a robust and efficient particle swarm optimization (PSO) variant, known as competitive swarm optimizer (CSO). In CSO, the particles benefit from randomly chosen opponents and not from the strongest global 340 or the individual position. The population of swarms is split randomly into two groups in each iteration and competitions between the particles in any group take place on a pair principle. The winner particle is directly transferred during every competition toward the next iteration, while the losing particle updates its position and velocity with a knowledge of the winning 345 particle based on the following equations:

where t represents the iteration counter, G t 1 , G t 2 , G t 3 represent three [0, 1] range vectors that are generated randomly, while x t w and x t l are respectively the winner and loser particles.x t represents the mean swarm position in iteration t, and the effect ofx t is handled by λ. 

where t shows the scale element. The probability distribution function of 365 the Cauchy component is shown by:

The mutation operator disrupts the population of the search agents and enables them break of the local minima. The deployment of CM in CSO is defined as following:

where W j represents the matrix of weights, x ij is the jth position of the 370 ith search agent and population size of the search agents is denoted by P .

The Cauchy mutation operator is applied as follows:

J o u r n a l P r e -p r o o f Journal Pre-proof

where A is a random vector generated by Cauchy.

-Second modification:

In CSO algorithm, per each iteration, the position of search agents is up- conditions. The EBCH updating process is given by the following formula:

where o i represents the out-of-bound search agent, ξ and ϱ are random 385 variables in [0,1] interval, ub i and lb i denote to the upper and lower bounds of the search space, respectively, and x b i is considered as the ith best search agent.

-Third modification:

In order to improve CSO convergence speed, λ parameter plays a key 390 role. In the iterative cycle of CSO, the evolutionary chaotic maps are used to strengthen the ability of utilizing and keeping the core search step harmonized. This concept motivates us to fine tune the λ parameter using a powerful chaotic map named tent map. The formula of tent map is given as following: 

where C represent a continues value in the [0, 1] interval for exploration 425 purpose in the continuous search space area, ζ is an operator responsible to map from C to [1, n + 1], and ψ is another mapping operator from ζ to [1, 2, 3, ..., n]. Each integer value belonging to the solution's continuous dimension can thereby be computed using the following formula:

Using Eqs. (4) and (5) Design P CNN models based on S;

Calculate the fitness (accuracy) Fi of each particle Xi with Eq.(17) based on T r;

gbest ← Xi; Xi of the best fitness;

for i=1 to ⌊P ⌋ (half of population) do

Select randomly two particles, Xk and Xm;

if Fk > Fm then

Xw = Xk; #Xw is the winner particle

Xl = Xm; #Xl is the loser particle Add Xw into new population;

Remove Xk and Xm from the population;

i= i + 1;

for i=1 to ⌊P ⌋ (half of population) do

Update velocity of loser as shown in Eq. (4);

Update the parameter λ by tent chaotic map using Eq. (11);

Update position of loser using Eq. (5);

Apply CM operator using Eq. (9);

Apply the EBCH operator using Eq. (10);

Calculate the fitness of new loser, Fl;

Move new loser into new population;

Update gbest if there is better solution;

i= i + 1;

end for

Pass new population to next iteration; 

The proposed DNE framework uses eleven parameters that should be learned automatically in a deep evolutionary scheme. With respect to pa- • F1-score (F 1) is the harmonic average of two other metrics precision (P) and recall (R) defined as follows:

• Accuracy is defined as the rate of the corrected classified cases:

• AUC is described as the area under the ROC (receiver operating characteristic) curve showing the performance of classification/detection model.

It is defined as follows: The convergence speed of the EA models is compared and the results are 560 shown in Fig. 6 . It can be seen from this figure that the convergence speed of the proposed method is better than other EA models. Moreover, the proposed method obtains the best accuracy among other EA models in different iterations. we can see from this figure, the confusion matrix of the proposed method presents better result than the other models. It is worth noting that a confusion matrix with more true predicted samples can be considered as a better result. Therefore, it can be concluded that the proposed method based on KNN algorithm obtains better performance than the other clas-585 sification algorithms. Fig. 8 shows the box plots of different classification algorithms based on the accuracy metric. It can be seen from this figure that the proposed method is the best performer among all classification algorithms as it obtains the best average accuracy. Therefore, it can be concluded that the KNN algorithm can provide better classification strategy 590 in comparison to other models to classify the input COVID/Non-COVID images. In this section, we analyze the run-time of our proposed algorithm and other benchmarking methods for COVID-19 diagnosis. The time consuming results for all algorithms in terms of optimization, training, and testing times are tabulated in Table 6 . As can be seen from this table, in terms of optimization time consumption, our proposed algorithm has the short- Table 2 where it is shown that the proposed method outperforms other compared methods. The main reason to obtain such results is to consider the three evolutionary operators in the proposed evolutionary algorithm which leads to a significant improvement in the accuracy of classification Table 3 where we can see that the KNN classifier performs better than other models. This is due to the fact that the KNN classifier takes into account the agreement Table 4 where it can be seen that the proposed method is better than other models. Therefore, we can conclude that considering the KNN classifier 695 in the last layer of CNN model and also using the proposed evolutionary algorithm to optimize the hyperparameters of CNN model significantly improve the classification accuracy.

In addition, to prove that the experimental results are statistically significant, we performed a ranking statistical test called the Friedman ranking 700 test to theoretically evaluate and analyze the performance of the proposed model and all benchmark algorithms. The results of this statistical test reported in Table 5 prove that the experimental results are statistically significant. Also, the run time of the proposed method is compared with other methods and the results are shown in Table 6 based on optimiza-705 tion time, training time, and test time. These results demonstrate that the proposed classification method performs in a shorter time with respect to other compared models. Therefore, it can be concluded that the proposed method not only achieves more accurate classification results, it also performs faster than other compared methods.

Coronavirus disease (COVID-19) has become a very challenging issue in the world and made enormous problems in different countries. One of the main approaches to detect COVID-19 cases is to investigate chest Xray images by physicians. Due to the high performance of deep neural 715 networks in interpreting images, several COVID-19 detection approaches have already been developed. However, these approaches mainly ignore the tuning of hyperparameters of deep neural networks optimally and utilize a static way to manually set the values of these hyperparameters. To address this challenge, we proposed a novel image classification method in 720 this paper to detect COVID-19 cases based on the chest X-rays images.

Specifically, we utilized CNN model as a deep neural network to make a trained classification model based on the input samples. Moreover, to optimize the hyperparameters of the CNN model, we used an improved evolutionary algorithm to find optimal values leading to enhance the clas- where the results showed that the proposed method can perform in a faster way than the other models. As the future works, other evolutionary algorithms can be employed to obtain better results for the CNN model.

Moreover, other deep neural network models can be developed to make a

A deep learning based trustand tag-aware recommender system

A novel deep neuroevolution-based image classification method to diagnose coronavirus disease (covid-19)

An efficient 750 cardiovascular disease detection model based on multilayer perceptron and moth-flame optimization. Expert Systems

Alleviating data sparsity problem in time-aware recommender systems using a reliable 755 rating profile enrichment approach

A temporal clustering approach for social recommender systems

Ieee/acm international conference on advances in social networks analysis 760 and mining (asonam)

A social recommender system based on reliable implicit relationships. Knowledge-Based Systems

Training back propagation neural 765 networks using asexual reproduction optimization

Incorporating reliable virtual ratings into social recommendation systems

A social recommendation method based on an adaptive neighbor selection mechanism. Information Processing and Management

An improved model of trust-aware recommender systems using reliability measurements

Sentiment analysis and its applications in fighting covid-19 and infectious diseases: A systematic review

Survival exploration strategies for harris hawks optimizer

X-ray covid-19 detection based on scatterwavelet transform and dense deep neural 785 network

Augmented covid-19 x-ray images dataset

Recognition of covid-19 disease from x-ray images by hybrid model consisting of 2d curvelet transform, chaotic salp 790 swarm algorithm and deep learning technique

Medical image analysis using convolutional neural networks: a review

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

Deep learning and its role in covid-19 medical imaging

A deep learning based approach for automatic detection of covid-19 cases using chest x-ray images

Explainable 805 deep learning for pulmonary disease and coronavirus covid-19 detection from x-rays

Sufmofpa: A superpixel and meta-heuristic based fuzzy image segmentation approach to explicate covid-19 radiological images

A competitive swarm optimizer for large scale optimization

A multi model ensemble based deep convolution neural network structure for detection of covid19

An improved quantuminspired differential evolution algorithm for deep belief network

Covid-19 disease identification from chest ct images using empirical 820 wavelet transformation and transfer learning

Harris hawks optimization: Algorithm and applications

Multi-task multi-modality svm for early covid-19 diagnosis using chest ct data

Explainable deep graph diffusion pseudo-labelling for identi-830 fying covid-19 on chest x-rays

Deep learning approaches for covid-19 detection based on chest x-ray images

An oppositional-cauchy based gsk evolutionary algorithm with a novel deep ensemble reinforcement learning strategy for covid-19 diagnosis

Automated deep cnn-lstm architecture design for solar irradiance 840 forecasting

Evolving artificial neural networks using butterfly optimization algorithm for data classification

Towards novel deep neuroevolution models: chaotic levy grasshopper optimization for short-term wind speed forecasting

Neuroevolution-based autonomous robot navigation: a comparative study

A novel evolutionary-based deep convolutional neural network model for intelligent load forecasting

Autonomous robot navigation using moth-flame-based 860 neuroevolution

Optimal autonomous driving through deep imitation learning and neuroevolution

A new ensemble reinforcement learning strategy for solar irradiance forecasting using deep optimized convolutional neural network models

Parsimonious evolutionary-based model development for detecting artery disease

Autonomous robot navigation system using the evolutionary multiverse optimizer algorithm

A new hybrid deep neural architectural search based ensemble reinforcement learning strategy for wind power forecasting

Histogram based vehicle license plate recognition with knn method

Loop mediated isothermal amplification (lamp) assays as a rapid diagnostic for covid-19

Deep learning 890 for pattern recognition of photovoltaic energy generation

Comparing image classification methods: K-nearest-neighbor and support-vector-machines

Sarsnet: Covid-19 detection from chest x-rays by combining graph convolutional network and convolutional neural network

Deep transfer learning-based covid-19 prediction using chest x-rays. medRxiv

A least squares support vector machine model optimized by moth-flame optimization algorithm for annual power load forecasting

A transfer learning method for pneumonia classification and visualization

Visualizing deep convolutional neural networks using natural pre-images

The whale optimization algorithm

A trust-aware recommender algorithm based on users overlapping community structure

Neural network training using a biogeographybased learning strategy

Deep neural network to detect covid-19: one architecture for both ct scans and chest x-rays

dforml (knn)-pseaac: Detecting formy-925 lation sites from protein sequences using k-nearest neighbor algorithm via chou's 5-step rule and pseudo components

Automated detection of covid-19 cases using deep neural net-930 works with x-ray images

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

Application of deep learning for fast detection of covid-19 in x-rays using ncovnet

Review of the chest ct differential diagnosis of ground-glass opacities in the covid era

Covid-19 identification in chest x-ray images on flat and hierarchical classification scenarios

A survey on deep learning: Algorithms, techniques, and applications. ACM Computing Surveys (CSUR)

Optimising control and prediction horizons of a model predictive control-based motion 950 cueing algorithm using butterfly optimization algorithm

Lglmf: local geographical based logistic matrix factorization model for poi recommendation

Grasshopper optimisation algorithm: theory and application

Lung infection quantification of covid-19 in ct images with deep learning

Developing an efficient deep neural network for automatic detection of covid-19 using chest x-ray images

Review of artificial intelligence techniques in imaging data acquisition, segmentation and diagnosis for covid-19

A hybrid recommendation system based on profile expansion technique to alleviate cold start problem

Intelligent skin cancer diagnosis using improved particle swarm optimization and deep learning models

Deep learning on image denoising: An overview

Covid-net: a tailored deep convolutional neural network design for detection of covid-19 cases from chest x-ray 980 images

Deep convolutional neural network based medical image classification for disease diagnosis

Novel 985 bio-inspired memetic salp swarm algorithm and application to mppt for pv systems considering partial shading condition

Deep learning-based decision-tree classifier for covid-19 diagnosis from 990 chest x-ray imaging

Hyperspectral image denoising employing a spatial-spectral deep residual convolutional neural network

A hybrid grasshopper optimization algorithm with bat algorithm for global optimization

Deep learning-based detection for covid-19 from chest ct using weak label. 1000 medRxiv

Improvement of generalization ability of deep cnn via implicit regularization in two-stage training process

Multi-core sine cosine optimization: Methods and inclusive analysis

better representation of the input images and provide better performance 740 for the detection of coronavirus disease.