key: cord-0722600-g3xr1ay8
authors: Akbar, Sumaiya Begum; Thanupillai, Kalaiselvi; Sundararaj, Suganthi
title: Combining the advantages of AlexNet convolutional deep neural network optimized with anopheles search algorithm based feature extraction and random forest classifier for COVID‐19 classification
date: 2022-04-10
journal: Concurr Comput
DOI: 10.1002/cpe.6958
sha: 6b232da395a24aeada77fac2a687b06fec9e19aa
doc_id: 722600
cord_uid: g3xr1ay8

In this article, COVID‐19 detection and classification framework based on anopheles search optimized AlexNet convolutional deep neural network for random forest classifier is implemented. Here, the COVID‐19 dataset is taken from Joseph Paul Cohen database. Then, the input images are preprocessed with the help of fuzzy gray level difference histogram equalization technique (FGLHE) and fuzzy stacking technique for color enhancement and noise elimination in the input images. The FGLHE technique and fuzzy stacking technique are combined together and forms into stacked dataset image. This stacked dataset are trained with AlexNet convolutional deep neural network model and the feature packages acquired via the models are processed by the anopheles search algorithm. Subsequently, the efficient features are combined and delivered to random forest (RF) classifier. The proposed approach is implemented in MATLAB. The proposed ADCNN‐ASA‐RFC provides 91.66%, 69.13%, 34.86%, and 70.13% higher accuracy, 79.13%, 60.33%, and 63.34% higher specificity and 77.13%, 58.45%, 25.86%, and 55.33%, higher sensitivity compared with existing algorithms. At last, the simulation outcomes demonstrate that the proposed system can be able to find the optimal solutions efficiently and accurately with COVID‐19 diagnosis.

learning strategies, AI is employed in numerous applications, like image detection, data classification, and segmentation. X-ray imaging is better than CT, because lesser ionizing radiation, fast data acquisition, accessibility on intensive care units (ICUs), and portability. There is no consensus yet in the combined use of chest CT and CXR in the treatment of COVID-19 pneumonia. The role of machine learning algorithms is to recognize objects. The machine learning algorithm has following advantages, such as it easily identifies trends and patterns, handles multidimensional and multivariety data, does not need human intervention (automation). The random forest classifier is used as the machine learning algorithm. 29, 30 Here, random forest classifier is used to accurately classify the COVID-19 chest X-ray images in to COVID-19, pneumonia, and normal.

The novelty of this article is to detect the type of disease affected in chest X-ray images and classify the images based on normal, COVID-19, pneumonia with high accuracy less computational time and error rate.

The key contributions of this article are abridged as follows,

• Here, COVID-19 detection and classification framework based on anopheles search optimized AlexNet convolutional deep neural network for random forest classifier is implemented.

• An original dataset is preprocessed with the help of FGLHE technique and fuzzy stacking technique for color enhancement and noise elimination in the input images.

• FGLHE technique and fuzzy stacking technique are combined together and forms into stacked dataset image.

• This stacked dataset are trained with the AlexNet convolutional deep neural network model and therefore the feature packages acquired via the models are processed by the anopheles search algorithm.

• Subsequently, the efficient features are combined and delivered to random forest (RF) classifier. 30 • The proposed approach is implemented on MATLAB.

• The proposed COVID-19 detection and classification framework based on anopheles search optimized AlexNet convolutional deep neural network for random forest classifier performance analysis like accuracy, sensitivity, specificity, precision, and F-score are compared with existing algorithms, like mobile net deep convolutional in social mimic optimization for support vector machine (MNDL-SMO-SVM) and res net deep learning in state of art algorithm for binary classifier (RSDL-STA-BC), computed tomography, and chest X-rays for COVID-19 (HDDNs) Detection, 27 novel process for COVID-19 detect with artificial intelligence under chest X-ray imageries (ResNet) 28 respectively.

The remainder of this article is organized as: Section 2 portrays the recent research works, Section 3 describes the proposed COVID-19 detection and classification framework based on anopheles search optimized AlexNet convolutional deep neural network for random forest classifier, Section 4 demonstrates the experimental results, Section 5 presents the conclusion of this article.

Among the recent research works related with COVID-19, a few research works are reviewed here, Jelodar et al., 22 have presented the automated removal of COVID-19 -consideration as social networks and language process scheme supported by topic modeling to discover numerous problems associated with COVID-19 as public opinion. Additionally, the experiments showed that the sample came with an accuracy of 81.15%, which was next precision of many recognized machine learning methods for COVID-19.

Rajaraman et al., 23 have presented the use of repetitive pruned sets of deep learning models to notice the pulmonary expression of COVID-19

with chest X-rays. The knowledge learned was well designed for enhancing performance and generalizations within associated task of classifying chest radiographs as usual, viewing bacterial pneumonia or COVID-19 virus as abnormal. The most effective performance models were pruned again and again for reducing the problem and enhance memory effectiveness.

Wang et al., 24 have suggested a completely unique noise resistant framework to discover noisy labels. They initially advertised a noise robust data loss which was generalization of data loss for segmentation and mean absolute error (MAE) loss. The experimental outcomes demonstrated that: (1) noise resistant matrix loss outstrips existing noise resistant loss functions; (2) COPLE-Net reaches greater execution to next-generation image segmentation networks.

Albahri et al., 25 have suggested the artificial intelligence (AI) techniques used on corona virus disease detection and classification in 2019.

A comprehensive methodology to evaluate and compare AI processes utilized at entire COVID-19 medical image classification tasks. The experimental outcomes shows that no relevant study has assessed and compared the AI processes utilized in COVID-19 medical image classification tasks.

Abdel-Basset et al., 26 have presented to extract quickly from chest X-ray images the similar small regions that have recognizing features of COVID-19. IMPA performance was verified through entry levels among 10 and 100 on nine chest X-ray images compared with five algorithms:

Equilibrium Optimizer, whale optimization, sine cosine, Harris-Hawks and Slap Swarm. The outcomes show that the hybrid method outperforms other methods for variety of measurements.

Irfan et al., 27 have presented the contribution of hybrid deep neural networks (HDNN), chest radiographs, computed tomography for COVID-19

diagnosis. The aim was to improve the HDNN, computed tomography and X-ray image. It was categorized as three classes: normal, pneumonia, and COVID-19. Firstly, computed tomography along chest X-ray imageries, named "hybrid imageries" (with 1080 × 1080 resolution) were composed as dissimilar sources, involving Git Hub, the COVID-19 X-ray database, Kaggle, data collection of COVID-19 images and actual med COVID-19 chest X-ray dataset.

Almalki et al., 28 have presented an innovative model for the diagnosis of COVID-19 with artificial intelligence on chest X-ray images. The presented algorithm takes various starting residual blocks which was adapted to the information with feature maps of different depths at dissimilar scales. The presented efficient deep learning blocks recycled dissimilar regularization procedures to diminish overfitting based on small COVID-19 data set. Multiscale features were removed at dissimilar levels from presented deep learning model integrated into several machine learning models to authenticate that combination of deep and machine learning models.

Iwendi et al., 29 have presented a classification of people with COVID-19 with adaptive neuro-fuzzy inference system. The aim was to regulate such attributes under early detection of Corona virus disease with adaptive neuro-fuzzy inference system (ANFIS). The presented work calculates the accuracy of different machine learning and chooses the better classifier for COVID-19 identification. The COVID19 dataset was categorized with support vector machine because it accomplished 100% accuracy between the entire classifiers. Additionally, the ANFIS was executed on such classified dataset, resulting at 80% risk prediction for COVID-19.

Iwendi et al., 30 have presented a prediction of COVID-19 patient health with powered random forest algorithm. Where, a refined random forest model powered by AdaBoost algorithm was deemed. The presented model utilizes that geographic, travel, health, and demographic data of COVID-19 patient. The model has an accuracy of 94% and F1 score of 0.86 on data set utilized. Data analysis exposes a positive correlation among the gender of patients and deaths denote that most patients are between 20 and 70 years old.

From the Literature survey, [22] [23] [24] [25] [26] [27] [28] [29] [30] several methods, such as deep learning and machine methods are used to detect and classify the corona virus, it is observed that all the presented works deals with COVID-19 detection and classification with accuracy but faces lot of difficulties, such as noisy in nature, computational time is increased, more expensive to acquire, the assessment and reference process of COVID-19 AI classification occurs multicomplex attribute problem. The histograms of enhanced images spread the full range of gray-scale means there is no saturation and washed-out on output images. But, this proposed method overcomes all this issues and provides more accuracy.

Here, the proposed method of COVID-19 detection and classification framework using anopheles search optimized AlexNet convolutional deep neural network is implemented for random forest classifier. The fuzzy gray level difference histogram equalization technique and focus stacking technique are used in preprocessing section for enhancing the input image. Then the output of preprocessing are combined together as stacked dataset, which is trained with AlexNet convolutional deep neural network and the trained output into random forest classifier. This random forest classifier classifies as normal, COVID-19 and pneumonia. The overall block diagram for proposed ACDNN-ASA-RFC is shown in Figure 1 . From Figure 1 , the input COVID-19 chest X-ray images are given to the training and testing phase, then the imageries are preprocessed to eliminate the noises using fuzzy gray level difference histogram equalization technique and then the images are staked using the fuzzy stacking technique. After that the image features are extracted using the AlexNet convolutional deep neural network model and therefore the feature packages acquired via the models are processed by the anopheles search algorithm. Then F I G U R E 1 Block diagram for COVID-19 detection and classification using ACDNN-ASA-RFC for chest X-ray image the images are classified random forest (RF) classifier. Random forest (RF) classifier accurately classify the images as normal, pneumonia, and normal.

The detail discussion regarding the COVID-19 detection and classification framework using anopheles search optimized AlexNet convolutional deep neural network for random forest classifier is given below in the following section.

In this article, the three classes of data sets are available for COVID-19 classification such classes are COVID-19 normal, and pneumonia. Here, COVID-19 images are taken from Joseph Paul Cohen database. In which, 70% data set is utilized as training and 30% as testing data in the experimental analysis. This data set consists of total 458 chest images, from this 295 images are detected as COVID-19, 98 images are detected as pneumonia, and 65 images are detected as normal.

In this section the preprocessing is employed for separating the redundant content like noise, lights, and background on provided chest X-ray image. The preprocessing step is used for standardizing the input. The dataset is renovated by FGLHE technique and fuzzy stacking procedure.

Gray level differences are blurred to cope with the uncertainties that exist under the input image. A fuzzy gray level clipping range is intended for controlling negligible contrast development. Fuzzy gray level difference histogram balancing techniques are featured to improve the contrast in the MR clinical image and not to lose their naturalness. 31 In this case, the fuzzy gray level is initially computed for eliminating the uncertainties under the histogram image, and then followed using clipping process to manage the extreme expansion rate.

The size of input image is m × n,

The pixel intensity of an input image placed at (a, b) is denoted from I (a, b) = k s .

Here, s ∈ {0, 1, . … … , L − 1}, L implies count of intensity levels.

The following binary space partitioning (BSP) is calculated on R region through L × L dimension wherever the L = 5. The BSP operate from a description of the spatial structure, and it is essential to handle noise intensity in homogeneity and lighting variations on input image. The entire pattern of comparison goes to the R region connected with parameter T r = 0.3.

BSP P (i n , i c ) described that ratio among intensity of central pixels gray level i c and neighborhood intensity pixels i n is given by,

Gray level difference among intensity of central pixel i c through the linked coordinates (a, b) and the intensity of the neighborhood pixel i n by the linked coordinates (m, n) on region R is computed as

where N is represented as the Gaussian field.

The Gaussian membership function uses the described gray level differences vaguely,

here σ implies Gaussian function that may be extended and fitted an exacting weight for assigning the usual interval of gray level differences. This method supports to get that fuzzy gray level differences histogram. 

In this section, image stacking mixes multiple images taken or recreates them at changed focal lengths. This is a common way to enhance the standard of photographs within the dataset. This method aims to remove noise as initial image by combining a minimum of two images under same row and separating image.

implies marginal Fourier transform of reconstructed target function f(x, y, z) regarding variable (y, z) and then target function

here the constant amplitude factor in the Fourier integral is deserted. Comparing Equation (6) find the reconstruction of K y,z (

where,

Basically, it can be implemented beyond one-dimensional integral by adding the obtainable discrete values F w and Fourier transform digital implementation or their inverse. Thus, the target function is expressed according to the discrete form of reconstruction at (x = x n , y, z)

During this article, MATLAB library is employed to stack system. 32 The initial dataset is stacked on the recreated dataset by Fuzzy system. The convolutional is the base layer of CNN. In this layer, the input image passes through the filter. The out coming filter values has a characteristic map. This layer uses few cores that slip and finish the pattern to remove the lower and higher level features. The kernel implies 3 × 3 or 5 × 5 shaped matrix that needs to be transferred an input pattern matrix. The convolutional layer output is expressed as:

where X ; j implies j feature map on layer l, W l−1 j implies j kernels on layer l-1, Y l−1 a denotes the feature map on layer l-1, b l j implies bias j feature map on layer l, N indicates total number of features on layer l-1, (*) refers vector convolution mode. The second layer after convolutional layer is pooling layer. The pooling layer is frequently used to present maps produced to drop that number of feature maps. The max pooling process chooses only maximal value with the specified array size on every feature map, resulting at decreased output neurons. It is linked with the completely connected layer after the global average pooling layer. The major objective of this layer is to avoid overfitting and network divergence.

The completely connected layer is the significant layer of CNN. This layer acts as multilayered perception. The rectified linear unit (ReLU) 33 trigger function is generally utilized in completely connected layer, though the Soft Max trigger function is utilized to forecast output images on final layer of the completely connected layer. The mathematic calculation of two activation functions is below:

here, x i and m indicates input data and number of classes, correspondingly. Additionally, it normalizes the dimensional features of the images with the use of computing variance as well as mean value of the input image samples. In this, two parameters , are introduced for recovering the feature distribution function (t) using parameter training. The calculation formula is mentioned as Equation (2) and Equation (3),

where , denotes the sample mean value and variance value. Υ Denotes the constant value and i represents the total number of samples. Also, , represent the parameter used for feature extraction. By using AlexNet various COVID-19 features are extracted. While detecting the images some errors may occurs that will reduce the accuracy. For reducing the error rate, the parameters of the AlexNet is optimized using the Anopheles search algorithm.

Anopheles search algorithm is a new meta-heuristic algorithm for optimizing the errors in the networks. Here anopheles search algorithm is used to optimize the parameters of the AlexNet such as , for improving accuracy and to lessen that error rate during classification process. In this is recycled to upsurge the accuracy and is used to lessen that error rate. Anopheles search algorithm (ASA) 31 is a process evolved in every problem creates a solution with AS algorithm that moves to the optimal solution. Every solution decides the variance on optimal value by likening the local value acquired on final repetition.

Step by step procedure of ASA Here, optimizations are takes place based on the anopheles mosquitoes that are close to three tents. Some anopheles mosquitoes are positioned that not carriers near three tents. At initial tent, a malaria-infected child with parasitism gets ready to transmission. At second, a malaria-infected child with parasitism not ready to transmission. At third, a healthy child is existed. The mosquito sniffs the smell of children's body through the air. The outcomes demonstrate that the mosquitoes have inspired to the smell of the boy who is in the first tent twice as much as other children. By this process, the Anopheles is used to optimize the parameters of the AlexNet for increasing accuracy and to reduce the error rate. Figure 2 portrays the flow chart of ASA-AlexNet.

Step 1: Initialization.

The initial anopheles is initialized for optimizing the parameters of the AlexNet. When the term P in ASA expresses the population, the parameter B articulates the better optimal value, and parameter I implies number of iterations. Decision variables consist of lower bound lb and upper bound ub value. The following equations are used to implement the AS algorithm.

In Equation (11), an inverse distance is taken among X i and point Y i , then C represents optimal point on solution space according to the location X i may be consequent as below,

Step 2: Random generation.

Mosquitoes are distributed randomly under search space. Then the distance calculation equation is given in Equation (18) Dist

If b = 0.5, the O will enlarge. The closer value of b to their upper one of the Anopheles movement maximized. If b = 0, then the distance and optimality are employed for computing anopheles movement.

Step 3: Fitness function.

This is used to attain that objective function for increasing accuracy and lessen the error rate. The fitness function is expressed in Equation (19),

where is represented as the accuracy and is represented as the error rate.

Step 4: Updation of ASA for increasing accuracy and decreasing error rate.

In this optimization, the parameters are determined by obtaining the perceived odor density of each mosquito and moved toward the optimum point depending on the accomplished value. Then the accuracy is increased by reducing the error rate and the maximization equation is given in Equation (20) F I G U R E 2 Flow chart for AlexNet-ASA

where Maximixe(Accuracy) (X i ) , is represented as the objective function, X i refers set of decision variables, , which is given in Equation (21), (22) ,

Then the error rate minimized equation is given in Equation (23) Minimize(Error rate)

Equations (18) and (21) are known as the optimization equation for achieving the objective function. The parameters are optimized based on the odors of anopheles.

Step 5: Termination.

Finally, the objective function is utilized for maximizing the accuracy by minimizing the computational time with fault. At last, the best values are selected from the AlexNet classifier through ASA mechanism, which are effectively classified the input chest X-ray imageries as COVID-19 and pneumonia images.

In this section, random forest classifier 19 is employed to the COVID 19 classification. The random forest classifier is a stochastic search system to solve optimal solutions in huge and complex search spaces. It is effectively recycled for a popular evolutionary algorithm (EA) classification. Every individual is an input data encoding named chromosomes. The search for a better solution is guided through an objective function named fitness function. High fitness function designated solutions have more capacity to create innovative solutions to lower fitness value ones, while weak fitness function ones will be phased out.

where F is the fitness function, C i indicates the number of classified instance and T s the training samples.

The random forest classifier (RFC) is most effective cooperative learning procedure proven to the most popular and powerful systems under model recognition and machine learning for categorizing high dimensions and skewed issues.

Assume the learning set

Consider m is the vectors, N ∈ X here X is set of numeric or symbolic observations and M ∈ Y where Y refers class label set. A classifier denotes mapping X → Y. A novel input vector is categorized by every individual tree of random forest.

This section describes about simulation result detection, 27 new approach for COVID-19 diagnosis with artificial intelligence on chest X-ray imageries (ResNet) 28 algorithms respectively.

In this article, the three classes of data sets are available for COVID-19 classification such classes are COVID-19 normal, and pneumonia. In this the COVID-19 dataset is taken from the Joseph Paul Cohen database. In the experimental investigation, 70% dataset is recycled as training, 30% is recycled as testing data. This data set consists of total 458 chest images, from this 295 images are detected as COVID-19, 98 images are detected as pneumonia, and 65 images are detected as normal.

The simulation parameters of proposed method are verified on Table 1 . 

The measure of the overall effectiveness of classification system is named as accuracy

For identifying patterns of a negative class is computed to measure the ability of the classifier

For identifying patterns of a positive class is computed to measure the ability of the classifier

F-measure creates the use of recall and precision is described as

Positive predictive value

Negative predictive value

Although specimens belonging to the selected class are properly recognized via classifier, such specimens are placed in TP codes. Other models belonging to correctly recognized opposite classes are within the TN codes within the confusing matrix. 

In this article, COVID-19 detection and classification framework based on anopheles search optimized AlexNet convolutional deep neural network for random forest classifier is successfully implemented. The proposed work is simulated in MATLAB. The proposed ADCNN-ASA-RFC provides higher precision 86.66%, 66.33%, and 70.66% and higher F-score 80.33%, 53.33%, and 59.66% compared with existing algorithms like mobile net deep convolutional in social mimic optimization for support vector machine (MNDL-SMO-SVM) and res net deep learning in state of art algorithm for binary classifier (RSDL-STA-BC). The performance of proposed model is evaluated through expert radiologists and it is prepared through a greater database. This model is recycled on remote locations in countries exaggerated by COVID-19 to overwhelm the shortage of radiologists. In this way, this method is applicable in the clinical use.

This ADCNN-ASA-RFC is appropriate in the applications of real time. The ADCNN-ASA-RFC is also able to classify the patients based on COVID-19 by collecting real time data from various health care services, such as medical diagnose center, hospital. Not only that, this method is utilized to proficiently classify the input dataset image as COVID-19, normal, pneumonia. Furthermore, this method assesses the accuracy of various deep learning and machine strategies employed to detect the COVID-19 at beginning stage. This method proves better prediction along classification tasks when compared with other classifier methods. These tasks are applied in COVID-19 for disease probability prediction, screening, diagnosis, treatment guidance, and complication management. Several methods have been applied to fulfill these clinical purposes using their own specific capabilities. Hence, this application represents a promising tool to aid the stratification of COVID-19

patients.

Future work will focus on generating a pipeline that connects chest X-ray scanning computer vision models through such sorts of healthcare and demographic data process models. These models will be incorporated to requests that help the evolution of mobile healthcare. It can deliver a step toward semiautonomous diagnostic system and give rapid detection for regions exaggerated by COVID-19.

Data sharing is not applicable to this article as no new data were created or analyzed in this study.

https://orcid.org/0000-0002-0931-5709

Kalaiselvi Thanupillai https://orcid.org/0000-0002-9423-0223

Transmission of 2019-nCoV infection from an asymptomatic contact in Germany

Estimated effectiveness of symptom and risk screening to prevent the spread of COVID-19

Critical review of COVID-2019 in the world

Wuhan seafood market may not be source of novel virus spreading globally

COVID-19: what has been learned and to be learned about the novel coronavirus disease

The deadly Corona virus (COVID-19)

COVID-19: emergence, spread, possible treatments, and global burden. Front Public Health

Virology, epidemiology, pathogenesis, and control of COVID-19

Genome sequencing and phylogenetic reconstruction reveal a potential fourth rhinovirus species and its worldwide distribution

Ideal position and size selection of unified power flow controllers (UPFCs) to upgrade the dynamic stability of systems: an antlion optimiser and invasive weed optimisation algorithm

A multi-objective hybrid algorithm for planning electrical distribution system

Trusted secure geographic routing protocol: outsider attack detection in mobile ad hoc networks by adopting trusted secure geographic routing protocol

Survey on software defect prediction techniques

Clinical recommendations for in-hospital airway management during aerosol-transmitting procedures in the setting of a viral pandemic

COVID-19 detection in chest X-ray through random forest classifier using a hybridization of deep CNN and DWT optimized features

COVID-19 driven advances in automation and artificial intelligence risk exacerbating economic inequality

AI" stand for increasing inequality in the era of COVID-19 healthcare?

Using AI ethically to tackle COVID-19

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

Pulse coupled neural networks for image processing -A review

Diagnosis of leaf disease using enhanced convolutional neural network

Deep sentiment classification and topic discovery on novel coronavirus or COVID-19 online discussions: NLP using LSTM recurrent neural network approach

Iteratively pruned deep learning ensembles for COVID-19 detection in chest X-rays

A noise-robust framework for automatic segmentation of COVID-19 pneumonia lesions from CT images

Systematic review of artificial intelligence techniques in the detection and classification of COVID-19 medical images in terms of evaluation and benchmarking: taxonomy analysis, challenges, future solutions and methodological aspects

A hybrid COVID-19 detection model using an improved marine predators algorithm and a ranking-based diversity reduction strategy

Role of hybrid deep neural networks (HDNNs), computed tomography, and chest X-rays for the detection of COVID-19

A novel method for COVID-19 diagnosis using artificial intelligence in chest X-ray images

Classification of COVID-19 individuals using adaptive neuro-fuzzy inference system

COVID-19 patient health prediction using boosted random forest algorithm. Front Public Health

A novel reformed histogram equalization based medical image contrast enhancement using krill herd optimization

Blurring-effect-free CNN for optimization of structural edges in focus stacking

Metaheuristic anopheles search algorithm

COVID-19 detection using deep learning models to exploit social mimic optimization and structured chest X-ray images using fuzzy color and stacking approaches

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

Combining the advantages of AlexNet convolutional deep neural network optimized with anopheles search algorithm based feature extraction and random forest classifier for COVID-19 classification