key: cord-0832820-nn8l9de7
authors: Shankar, K.; Perumal, Eswaran; Díaz, Vicente García; Tiwari, Prayag; Gupta, Deepak; Saudagar, Abdul Khader Jilani; Muhammad, Khan
title: An optimal cascaded recurrent neural network for intelligent COVID-19 detection using Chest X-ray images
date: 2021-09-08
journal: Appl Soft Comput
DOI: 10.1016/j.asoc.2021.107878
sha: e4a63676547e8a3d245a34b4700cbab6a90eaecd
doc_id: 832820
cord_uid: nn8l9de7

In recent times,COVID-19, has a great impact on the healthcare sector and results in a wide range of respiratory illnesses. It is a type of Ribonucleic acid (RNA) virus, which affects humans as well as animals. Though several artificial intelligence-based COVID-19 diagnosis models have been presented in the literature, most of the works have not focused on the hyperparameter tuning process. Therefore, this paper proposes an intelligent COVID-19 diagnosis model using a barnacle mating optimization (BMO) algorithm with a cascaded recurrent neural network (CRNN) model, named BMO-CRNN. The proposed BMO-CRNN model aims to detect and classify the existence of COVID-19 from Chest X-ray images. Initially, image pre-processing is applied to enhance the quality of the image. Next, the CRNN model is applied as feature extraction and the hyperparameter tuning of CRNN takes place via the BMO algorithm to improve the classification performance. The BMO algorithm determines the optimal values of the CRNN hyperparameters namely learning rate, batch size, activation function, and epoch count. The application of CRNN and hyperparameter tuning using BMO algorithm shows the novelty of this work. A comprehensive simulation analysis is carried out to ensure the better performance of the BMO-CRNN model and the experimental outcome is investigated using several performance metrics. The simulation results portrayed that the BMO-CRNN model has showcased optimal performance with an average sensitivity of 97.01%, specificity of 98.15%, accuracy of 97.31%, and F-measure of 97.73% compared to state-of-the-art methods.

In recent times, Coronavirus Disease 2019 is referred to as COVID-19 which is considered as an epidemic disease from the end of December 2019 in the city of Wuhan, China. The name COVID-19 was dictated by the World Health Organization (WHO) as a novel and dangerous infection which comes under the class of Coronaviruses (CoV) and infectious viruses [1] . It results in the severe cases in critical care respiratory condition like severe acute respiratory syndrome (SARS-CoV) intend in shortness of breathing and death finally. Based on the survey reported by WHO the risk evaluation of COVID-19 is higher all over the world [2] . Additionally, the maximum number of people was positive for COVID-19 and people dies out of COVID-19. Some other lung disorders are Viral and Bacterial pneumonia, leading to high mortality.

Pneumonia diseases are caused due to the fungal infection of the lungs that is formed by pus and additional fluids filled in air sacs. However, bacterial pneumonia is highly critical; in particular, the small kids are affected easily due to the low immunity power [3] [4] [5] .

Coronaviruses. But the RT-PCR samples accomplished maximum false-negative levels for COVID-19 positive cases [6] . Then, the radiological investigations are performed by applying Chest X-Ray (CXR) as well as Computed Tomography (CT) scans to examine the health condition of patients. But the exposure to radiation results in high side effects. The CT scan provides a proficient model in screening, analysis, and progressive estimation of patients confirmed with COVID-19. Then, the clinical studies have reported positive CXR models paves the way for developing CT scans and reduce the medical overload of CT in 3 In computer-aided diagnosis (CAD) technologies are projected in the real-time solution for resolving the difficulties in chest X-rays, and to help the radiologists in predicting the diseases from low-contrast X-ray images [12] . The CAD models are integrated units of computer models with currently developed image processing approaches for performing interventional operations. Then, Artificial Intelligence is employed extensively for the advanced diagnostic function of CAD systems for different clinical applications like brain tumor classification or segmentation, reduced invasive aortic valve incorporation, and predicting pulmonary infections. Presently, Deep Learning (DL) models are part of a broader family of machine learning (ML) methods based on artificial neural network (ANN) with representation learning. It is applied for learning patterns and features from annotated data which is applicable for automated performance of a certain operation on earlier training like human sentiment classification as well as computer vision domains in surgery [44] [45] [46] .

Automatic investigations of COVID-19 and chest infections have been predicted under the application of clinical CT and X-ray imaging sectors. In recent times, [13] depicted that chest CT images and DL methodologies are significant in identifying and segmenting COVID-19 disease effectively. Therefore, it is concentrated on using CXR images as a primary device for COVID19 positive patients and pneumonia diseases. Automated classification of lung diseases in X-ray images has been projected for TB screening, prediction of lung intensity as well as severe pneumonia infections. However, examining the COVID19 disease in chest X-rays is still in progressive stage and examined in peer-review published articles. Some of the DL classification models like pre-trained InceptionV3 method were applied for detecting the COVID-19 [14] . Moreover, the pneumonia diseases are examines using CNNs approaches with better classification accuracy as projected in [15] . Drop-weight-related Bayesian CNNs have been employed for validating the maximum correlation of uncertainty by accomplishing supreme prediction accuracy in detecting the COVID-19 from given Xray images [16] .

In [17] , pneumonia X-ray images with the help of 3 diverse DL methods are employed for COVID -19 diagnosis. Under the application of the ResNet model, the dataset is classified into several labels like Age, Gender, and so on. Moreover, Multi-Layer Perceptron (MLP) classifier is applied and gained maximum accuracy. Yadav and Jadhav [18] processed a classification approach under the application of pneumonia details where SVM classifier, and InceptionV3, VGG-16 methodologies are DL modules. In this application, a dataset is classified into 3 types for enhancing the contrast as well as brightness zoom setting along with the augmentation model for an image with a dataset that has attained an optimal classification score. Abiyev and Ma'aitah [19] applied the Backpropagation Neural Network (BPNN), as well as Competitive Neural J o u r n a l P r e -p r o o f 4 Network, approaches for classifying the pneumonia information. By using pneumonia as well as healthy CXR images, some portions in a dataset are used as test data and related to the projected model with previous CNNs. Finally, better classification has been accomplished.

Stephen et al. [20] developed a DL technology for classifying the pneumonia data from scratch for data training. The input size is 200 x 200 pixels which have been employed for determining the viabilities of classification with a sigmoid function. Consequently, the optimal success rate is achieved in pneumonia.

Chouhan et al. [21] predicted the images of pneumonia by utilizing DL methods and 3 classes of the dataset.

Initially, pre-processing is performed for noise elimination. Followed by, the augmentation mechanism is applied for all images and employed a transfer learning for model training. Hence, effective classification accuracy has been accomplished. Islam et al. [22] developed a DL model using the integration of CNN and LSTM for COVID-19 detection and classification. Hussain et al. [23] employed DL with natural language processing tools to estimate the average sentiment, sentiment trend, and discussion topic related to COVID-19 vaccination. Melin et al. [24] proposed an ensemble neural network firefly algorithm for COVID-19 diagnosis.

Sujath et al. [25] projected a COVID-19 diagnosis model using linear regression, Multilayer perceptron, and Vector autoregression methods. A novel hybrid model is designed in [26] to predict COVID-19 using the integration of Statistical Neural Network (SNN) models and the Non-linear Autoregressive Neural Network (NAR-NN). Iwendi et al. [27] designed proposes a fine-tuned Random Forest with AdaBoost technique for predicting the health conditions of COVID-19 patients. A CAD model for COVID-19 is presented in [28] using the fusion of CNN with statistical and textural features. In Attallah et al. [29] , efficient CAD model is presented to detect COVID-19 using many CNNs and SVM models. El-bana et al. [30] presented a finetuned Inception v3 deep model to detect COVID-19 using multi-modal learning. Some other COVID-19 diagnosis models are available in the literature [31] [32] [33] . Though several methods are available in the literature, there is still a need to improve the COVID-19 diagnostic performance. Besides, only a few works have concentrated on the parameter optimization of the DL based feature extraction techniques.

In this paper, an intelligent COVID-19 diagnosis model using BMO algorithm with a CRNN model, called BMO-CRNN has been developed. The proposed BMO-CRNN model intends to identify and categorize the existence of COVID-19 from Chest X-ray images. Firstly, image pre-processing is applied to eliminate the noise. Also, the CRNN technique is applied as feature extraction and the hyperparameter tuning of CRNN takes place via the BMO algorithm. The CRNN explores the redundant and complementary information, and it involves two recurrent neural network (RNN) layers. The former one eliminates the redundant data and J o u r n a l P r e -p r o o f the latter one aims learning the complementary data. Lastly, the SoftMax (SM) layer is utilized for classification purposes to categorize COVID-19 or non-COVID-19. A comprehensive simulation analysis is done to ensure the goodness of the BMO-CRNN technique using the Chest X-ray (CXR) dataset. The key contributions of the paper are summarized as follows.

 An intelligent COVID-19 diagnostic method involving pre-processing, CRNN based feature extraction, and BMO based parameter optimization is proposed. To the best of our knowledge, the BMO-CRNN model has not been presented in the literature for this problem.

 The design of CRNN based feature extraction involves a series of RNNs, which necessitate a minimal number of parameters that are highly applicable for classification, and a limited number of training samples is needed.

 The parameter optimization of the CRNN model using the BMO algorithm using cross-validation helps to improve the diagnostic performance of the BMO-CRNN model for unseen data. The use of BMO algorithm for the parameter tuning of the CRNN model shows the novelty of the work.

 Detailed experiments are conducted to validate the performance of the BMO-CRNN on the benchmark CXR dataset under various settings compared to SOTA methods.

The remaining portions of the paper are planned as follows Section 2 discusses the proposed BMO-CRNN model. Section 3 performs the simulation process and discussion. At last, Section 4 concludes the paper.

The working principle involved in the BMO-CRNN model is depicted in Fig. 1 , which comprises preprocessing, parameter tuning, feature extraction, and classification. Once the input image is pre-processed, the BMO algorithm determines the parameters of CRNN namely learning rate, batch size, activation function, and epoch count. When the parameters are identified, the feature extraction process is carried out. Step 3: Substitute the hyperparameters of the CRNN respective to the points of barnacles. Partition the dataset into the training and testing part, afterward identify the subsequent series of training data by analysing the trend line. By comparing the predicted and actual data, the error can be determined.

Step 4: Reiterate the above process until the termination criteria are fulfilled. 

The input images from CXR images are pre-processed using the GF technique, which is a linear smoothing filter used for weight selection depending upon the structure of the Gaussian function. On the spatial or frequency domain, the GF technique is applied as an effective low pass filtering technique, particularly for noise removal. The 1-D Gaussian function of zero means can be denoted as follows.

The parameter involved in the Gaussian distribution computes the width of the Gaussian functions. In the case of processing images, 2D discrete Gaussian function of zero mean namely smooth filter [35] , and the respective function is expressed as follows. (2)

The BMO method is a new bio-inspired optimization mechanism used for resolving the optimization issues.

The BMO method is chosen over the other optimization algorithms due to the following reasons:

 BMO algorithm balances the trade-off between exploitation and exploration for producing a new offspring towards a globally optimum solution.

 BMO algorithm offers enhanced outcome over the other compared methods and attained global optimum, high exploration ability, and avoids local optima problem.

 It is highly flexible and efficient over other algorithms.

This model is evolved based on the barnacle's mating nature and related functions. Even though barnacles are hermaphrodites that carried male and female genitalia, the mating process is performed rarely, which applies extraordinarily long penises and finds a mate inside the striking distance [36] . The Barnacles ovaries are placed in a stalk and expand into the mantle, often expand into the thorax. Also, self-mating is carried out rarely in barnacles which do not leave the shells for mating. Barnacles can reproduce by sperm casting where male barnacle discharge the sperm to the water and the female consumes and fertilizes the eggs is named sperm casting or self-mating. The lifecycle of BMO is shown in Fig. 2 [37] .

The different steps involved in the BMO algorithm are listed below and the flowchart is shown in Fig. 3  9 capable to compute the sequential input using a recurrent hidden state with the activation of the former step. Thus, the system represents the dynamic temporal behavior. Consider the sequence data( 1 , 2 , … , ),

where implies the data in ℎ time step, the RNN update the recurrent hidden state ℎ using Eq. (3):

where denotes a non-linear function. Thus, RNN is composed of output ( 1 , 2 , … , ). Ultimately, the data classification is performed by an output . In the classical RNN method, the update rule of the recurrent hidden state in (1) is executed as given below:

where and refers the coefficient matrices for input and activation of recurrent hidden units. Assume

( 1 , 2 , … , )is a sequence probability that is degraded as:

( 1 , 2 , … , ) = ( 1 ) ⋯ ( | 1 , … , −1 ).

Followed by, the conditional probability distribution is developed using a recurrent network

where ℎ is gained from (3) and (6). The hyperspectral pixel is treated as sequential data and a recurrent network is applied for modelling spectral sequence.

As the Deep Learning (DL) family is significant, RNNs show the best outcome in ML as well as computer vision operations. To overcome these problems, a model is developed with a sophisticated recurrent unit.

Long short-term memory (LSTM) is defined as the class of recurrent hidden units which is suitable for learning long-term series [39] .

When compared with the LSTM unit, the gated recurrent unit (GRU) requires a smaller count of parameters that is highly applicable for classification, and a limited number of training samples is required [40] . Thus, GRU is selected as an essential component of RNN. The major components of GRU are 2 gating units which are employed for controlling the data flow within the unit. Rather than using the activation of hidden layer for band is expressed as,

where denotes the update gate which has been retrieved by J o u r n a l P r e -p r o o f

where means a sigmoid function, refers a weight value, and depicts the weight vector. Likewise, h is determined by,

where ⊙ refers an element-wise multiplication, and signifies the reset gate is obtained from

In particular, the data sequence is classified into sub-sequences = ( 1 , 2 , ⋯ , ), where it is composed of different class labels. Followed by, final sub-sequence , the length of alternate sub-sequences is = ( / ), which refers to the closer integers less than or equal to / . Therefore, the i-th sub-sequence , ∈ {1,2, ⋯ , }, it is composed of given bands,

Additionally, each sub-sequence is fed into the first layer of RNNs, which has a similar architecture and distributes parameters for reducing the parameters. In the case of sub-sequence , every sample is composed of output from GRU. The final feature representation for , is referred to be (1) ∈ ℜ 1 , where 1 implies the size of the hidden layer in first layer RNN. Afterward, the (1) , ∈ {1,2, ⋯ , } are combined for generating sequence = ( 1 (1) , 2 (1) , ⋯ , (1) ) where length is . These sequences are induced into second layer RNN for learning complementary details. Likewise, the first layer RNNs applies the result of GRU finally as the learned feature (2) . The classification result of can be attained by inducing the input (2) into the resultant layer with equally sized candidate classes . Hence, 2-layer RNNs have massive weight parameters. Finally, selected as a loss function and apply the BPTT model for optimization. Fig. 4 shows the structure of the cascaded RNN model [40] . 

After the completion of the feature extraction process using BMO-CRNN, SM based classification is performed to identify the existence of COVID-19. It will map the input vectors from the N-dimensional space into K classes, as given in Eq. (12) : labels, as defined below.

(θ k c) ( = 1,2, … K) (12) where θ k = [θ k1 θ k2 … θ kN ] represents the weight factors and K denotes the number of classes.

This section validates the efficiency of the BMO-CRNN model on COVID-19 diagnostic process. The proposed model is simulated utilizing Python 3.6.5 tool along with few packages. The details related to the dataset, evaluation metrics, and comparative results analysis are made in the succeeding sections.

The class. Fig. 5 shows some test images from the CXR dataset. J o u r n a l P r e -p r o o f

The classification results of the BMO-CRNN technique are examined using different evaluation parameters such as sensitivity, specificity, accuracy, and F-measure [41] .

Sensitivity determines the proportion of positive samples are which are properly identified, as defined in Eq.

(13):

where TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively.

Specificity determines the proportion of negative samples which are properly identified, as given in Eq. (14) .

F-measure defines the harmonic mean of precision and sensitivity, which can be equated using Eq. (15):

Finally, accuracy determines the closeness of the measurements to a particular value, as defined in Eq. (16):

The convergence rate analysis of the BMO algorithms in terms of best cost attained under several iteration counts in Fig. 6 . The figure displayed that the PSO algorithm has failed to showcase better convergence over the other methods. In addition, the GWO, ABC, and BFO algorithms have exhibited certainly increased convergence over PSO. However, the BMO algorithm has surpassed all the compared optimization algorithms and reached a minimum best score under varying iteration count. 

A detailed comparative study of the BMO-CRNN with other existing methods is performed in Table 3 . Fig.   11 demonstrates the analysis of the BMO-CRNN with existing models in terms of sensitivity and specificity. 

Coronavirus infections-more than just the common cold

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

CNN-based transfer learning-BiLSTM network: A novel approach for COVID-19 infection detection

Profiling early humoral response to diagnose novel coronavirus disease (COVID-19)

An optimized deep learning architecture for the diagnosis of COVID-19 disease based on gravitational search optimization

Prediction models for diagnosis and prognosis of covid-19 infection: systematic review and critical appraisal

A novel medical diagnosis model for COVID-19 infection detection based on deep features and Bayesian optimization

Can lung US help critical care clinicians in the early diagnosis of novel coronavirus (COVID-19) pneumonia?

Automated medical diagnosis of COVID-19 through EfficientNet convolutional neural network

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

A deep-learning based multimodal system for Covid-19 diagnosis using breathing sounds and chest X-ray images

AI-assisted CT imaging analysis for COVID-19 screening: Building and deploying a medical AI system in four weeks

Forecasting models for coronavirus disease (COVID19): a survey of the state-of-the-art

COVIDX-Net: a framework of deep learning classifiers to diagnose COVID-19 in X-ray images

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

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

Comparison of deep learning approaches for multi-label chest X-ray classification

Deep convolutional neural network based medical image classification for disease diagnosis

Deep convolutional neural networks for chest diseases detection

An efficient deep learning approach to pneumonia classification in healthcare

A novel transfer learning based approach for pneumonia detection in chest X-ray images

A combined deep CNN-LSTM network for the detection of novel coronavirus (COVID-19) using X-ray images

Artificial intesslligence-enabled analysis of UK and US public attitudes on Facebook and Twitter towards COVID-19 vaccinations

Optimization using the firefly algorithm of ensemble neural networks with type-2 fuzzy integration for COVID-19 time series prediction

A machine learning forecasting model for COVID-19 pandemic in India

Covid-19 mortality rate prediction for India using statistical neural network models. Frontiers in Public Health

COVID-19 patient health prediction using boosted random forest algorithm. Frontiers in public health

FUSI-CAD: Coronavirus (COVID-19) diagnosis based on the fusion of CNNs and handcrafted features

MULTI-DEEP: A novel CAD system for coronavirus (COVID-19) diagnosis from CT images using multiple convolution neural networks

A multi-task pipeline with specialized streams for classification and segmentation of infection manifestations in covid-19 scans

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

COVID-19 pneumonia diagnosis using a simple 2D deep learning framework with a single chest CT Image: Model Development and Validation

2020. Finding covid-19 from chest x-rays using deep learning on a small dataset

A new image denoising method based on Gaussian filter

Barnacles Mating Optimizer: A new bio-inspired algorithm for solving engineering optimization problems

Deep recurrent neural networks for hyperspectral image classification

Long short-term memory

Cascaded recurrent neural networks for hyperspectral image classification

Swarm intelligence based classification rule induction (CRI) framework for qualitative and quantitative approach: An credit risk analysis

Deep learning and evolutionary intelligence with fusion-based feature extraction for detection of COVID-19 from chest X-ray images

A novel hand-crafted with deep learning features based fusion model for COVID-19 diagnosis and classification using chest X-ray images

TermInformer: unsupervised term mining and analysis in biomedical literature

IoT-cloud based healthcare model for COVID-19 detection: an enhanced k-Nearest Neighbour classifier based approach

Res-CovNet: an internet of medical health things driven COVID-19 framework using transfer learning

The authors extend their appreciation to the Deputyship for Research & Innovation, Ministry of Education in Saudi Arabia for funding this research work through the project number 959.