The outburst of novel corona viruses aggregated worldwide and has undergone severe trials to manage medical sector all over the world. A radiologist uses x-rays and Computed Tomography (CT) scans to analyze images through which the existence of corona virus is found. Therefore, imaging and visualization systems contribute a dominant part in diagnosing process and thereby assist the medical experts to take necessary precautions and to overcome these rigorous conditions. In this research, a Multi-Objective Black Widow Optimization based Convolutional Neural Network (MBWO-CNN) method is proposed to diagnose and classify covid-19 data. The proposed method comprises of four stages, preprocess the covid-19 data, attribute selection, tune parameters, and classify covid-19 data. Initially, images are fed to preprocess and features are selected using Convolutional Neural Network (CNN). Next, Multi-objective Black Widow Optimization (MBWO) method is imparted to finely tune the hyper parameters of CNN. Lastly, Extreme Learning Machine Auto Encoder (ELM-AE) is used to check the existence of corona virus and further classification is done to classify the covid-19 data into respective classes. The suggested MBWO-CNN model was evaluated for effectiveness by undergoing experiments and the outcomes attained were matched with the outcome stationed by prevailing methods. The outcomes confirmed the astonishing results of the ELM-AE model to classify covid-19 data by achieving maximum accuracy of 97.53%. The efficacy of the proposed method is validated and observed that it has yielded outstanding outcomes and is best suitable to diagnose and classify covid-19 data.

Coronavirus disease (COVID-19) laid major effects all over the world’s economic growth since past two years [^{th} March-2020, the World Health Organization (WHO) proclaimed COVID-19 a pandemic. Germany, France, Vietnam, New Zealand, South Korea, Saudi Arabia, Taiwan, and other countries have successfully limited the spread of COVID-19. While certain countries, including Brazil, the United States and India have been battling the pandemic for a long time. Dating 25th of March to the 31^{st}June-2020, India was under the toughest lockdown in its history to stop the coronavirus’s spread. However, the Indian government began exposing lockdown all over the country in stages beginning from 1^{st} July 2020 till 31^{st} August-2020, because a developing country like India cannot afford to lose money due to phase by phase lockdown [^{st} July to 30^{th} July-2020, with the second phase starting from 1^{st} August to 31^{st} August-2020, and the third phase starting from 1^{st} September to 31^{st} October-2020 [

Using MBWO-CNN, a COVID-19 diagnostic and classification model is suggested in the current research work. COVID-19 is detected, classified and well analyzed using the suggested MBWO-CNN model. Parameter adjustment, classification, pre-processing and feature extraction are the four processes involved in this approach. The input images are first preprocessed, then CNN-based feature extraction is performed.

Numerous approaches for predicting the existence of COVID-19 have been proposed by the researchers. This section reviews some of the methods. Gifani et al. [

The MBWO-CNN method’s working procedure is depicted in

The MBWO-CNN model is used to extract the features image after it has been prepared. The use of the MBWO algorithm aids during the choice of CNN’s initial hyper parameters.

CNN is a Deep Learning (DL) model subcategory that signifies the greatest accomplishment from image processing. During the image classification phase, CNN is mostly used to examine visual pictures. CNN is the most favored since it combines both hierarchical architecture and excellent picture feature extraction. The layers are initially assembled in a three-dimensional style. The neurons in the applying layers aren’t the same as the entire secondary layer’s population of neurons. In the other words, the secondary layer also has a limited number of neurons. As a result, they are reduced to a single vector containing all possible values which are further combined in the depth dimension. The training and testing procedures of Deep Convolutional Neural Networks (DCNN) in the categorization of COVID-19 are shown in

Input, hidden layers and output make up CNNs. Convolution, pooling, Rectified Linear Unit (ReLU) and Fully Connected (FC) layers are usually found in the hidden layers.

The input for the convolution layer is a convolution task. It passes the data to the following layer.

Clustering results with a neuron present the succeeding layer are concatenated in the pooling layer.

Fully Connected layers form a single layer that connects all of the neurons, with additional neurons

appearing in successive layers. The neurons of the FC layer get input from all of the present layer’s

parts.

The CNN algorithm works by extracting features from images. There is no need to manually extract features. As a result, the features are unequipped, and information is gained during network training with a collection of images. The DL model is particularly efficient in computer vision operations to perform the training process. CNN’s use a large number of hidden layers to do feature prediction. The layer makes learned features more challenging [

CNN learns experience through issues with hyper parameter tuning, according to the literature. Batch size, activation functions, stride, momentum, hidden layer, kernel size, kernel type, padding, learning value and count of epochs are the hyper parameters. Some variables need to be fine-tuned. The following is an example of a multi-objective Fitness Function (FF).

α_{n} and α_{p}, respectively, denote the sensitivity and specificity qualities. The number of true positives that are accurately categorized is processed by sensitivity, which is a true positive rate. The confusion matrix is used to estimate sensitivity (α_{n}), which can be calculated statistically based on the literature.

False Negative (FN) and True Positive (TP) measurements are referred to as

α_{n} is chosen as a number between 0 and 100. α_{n} considers the value of ‘100’ to be realistic. The ratio of correctly categorized True Negatives (TN) is determined by specificity (α_{p}), which is evaluated as follows.

F1-score = combination of precision/Recall or =2

Accuracy = Correct Prediction/Total Prediction

False Positive (FP) and True Negative (TN) measurements are denoted by

α_{n} belongs to the range [0,100]. It’s possible to think about the phenomenon of α_{p} approaching 100.

The flowchart of the BWO method is shown in

(i) The Begin Population

The optimization challenge can be solved by employing problem score metrics to create a good framework for a problem-solving solution. The structure is referred to particle position and chromosome in Genetic Algorithm(GA) and particle swarm optimization (PSO) techniques, respectively. However, in black widow optimization, it is referred to as a “widow.” Currently, it is suitable in the shape of a black widow spider and is expected to face all difficulties. The measurements of black widow spiders are depicted by these black widow spiders’ variables that cause issues. Furthermore, the architecture is illustrated in order to resolve the standard functions as a collection.

A widow is described as such an array of 1*N that reflects a remedy for problem and is expressed as follows for N dimensional optimization problems.

The floating-point value is represented by the variable measurements (x1, x2…, x_{N}). The fitness of a widow is determined by calculating the Fitness Function (FF),F for a widow of a certain variable (x_{1}, x_{2….} , x_{N}). Hence,

Along with assistance of the current population of spiders, a proposed widow matrix of size Np*N is built in order to launch the optimization technique, where, Np is also represented floating-point value. After that, a set of parents chosen at random to determine pro-development through mating. Once the mating is complete, male black widows are consumed by females.

(ii) Procreate

Pairs begin to mate and give birth to the following generation since they are autonomous in nature. The web is where the mating takes place. Throughout each mating, approximately 1,000 eggs are laid. Some of the spider lings die for various reasons, but the healthy one survives. An array called ‘alpha’ is created for reproduction. This is done to make sure the widow array is correct produced using random offspring values when the specified function is used. The parents are represented by x_{1} and x_{2}, while the offspring are represented by O_{1} and O_{2}.

It is repeated for

(iii) Cannibalism

So far, three varieties of cannibalism have been observed. After mating, a female BW consumes the male, which is known as sexual cannibalism. Fitness tests are used to assess both female and male participants. Second, when a healthy spider ling consumes the vulnerable ones, this is known as sibling cannibalism. Cannibalism Rating (CR) is calculated based on the number of survivors [

(iv) Mutation

Individual mutations values are selected at random to create a population in this method. In the array, the chosen solutions are swapped out in random order and mutation rate is used to calculate it.

For the purpose of classification, the latest studies have used ELM-SA. It takes the features it has gathered and assess the likelihood of things being present in an image. Both activation and deactivation have the same function. In establishing non-linearity and reduce overfitting, a dropout layer is used [

where G(w, b, x) denotes the activation function of the hidden layer. The input weight matrix that connects the input and hidden layers is defined by w, the hidden layer’s bias weight is denoted by the letter b and the weight of output hidden layers is denoted by A = [A_{1},A_{2,}A_{3}…A_{m}].A = ELM makes use of d input neurons, n training instances, there are m output neurons and k hidden neurons (i.e., m classes) whereas

where K_{j} is the m-dimensional output vector required for the i-th training instance x_{i}, from the input layer to the j-th hidden neuron, the j-th weight vector is denoted by d-dimensional w_{j} and the bias of the j-th hidden neuron is described by b_{j}, (w_{j}, x_{i}) refers to the inner product of w_{j} and x_{i} in this technique. As an activation function, the sigmoid function s is used as a result, the equation can be used to define the result of the j-th hidden neuron below.

where the arithmetic expression exp(.) of exponents, and the steepness attribute is represented by M^{2}.This technique

where T∈R^{n*m} is the desired outcome, A ∈ R^{n*k}. L=

where the output weights are =[ β_{1},…,β_{h}] and l(x)=[l(x_{1)},………..l(x_{n})] the feature space for such an input sample x is l(x)=[l(x_{1)},………..l(x_{N})] , where

So for n input samples, we can get L =

L is the Moore-Penrose generalized inverse of matrix L^{+} and T =

In general,

where C denotes the regularization factor, implying that we not only estimate the least norm of but also minimize training error. We can get more robust approach in this way. The output weights can then be determined as follows:

Then, using the minimum norm least-squares solution, A is obtained:

ELM is represented as follows: where C is a regularization,

Through a kernel function, Kernel-based ELM (KELM) is now available. Suppose

where

The symbols x_{i} and x_{j} show the i-th and j-th training occurrences, respectively. After that, the consequences of KELM are represented as follows, swapping LL^{T} for B.

where F_{ELM}(x) denotes the KELM method results of the simulation, in addition l(x)L^{T} =

The hiding node count is estimated and fixed in KELM, which is a key aspect. It does not contain any arbitrary feature mappings. Furthermore, due to the presence of the kernel technique, processing time is limited to ELM. The ELM-AE technique’s framework is depicted in

Suppose X^{(i)}=_{k}, k = 1 to n, the i-th data representation. λ^{i}_{.}= ^{i} and T is modified by X^{i} according to

Where L^{i} represents the i-th hidden layer’s resulting matrix when X^{i} is used, and λ^{i} has been resolved as follows:

Furthermore,

where X^{^} stands for X^{i} final implications. X^{^} is used as a hidden layer, resulting in the estimate of final weight β^{^}, and it is performed as follow.

The classifiers results are obtained, in the classification of chest X-ray images using the MBWO-CNN model [

The findings of the MBWO-CNN model produced a classifier analysis are shown in

Number of runs | F-score | Sensitivity | Accuracy | Specificity | Precision |
---|---|---|---|---|---|

1 | 95.90 | 95.10 | 95.10 | 95.78 | 96.78 |

2 | 95.65 | 95.16 | 96.22 | 97.12 | 96.22 |

3 | 96.32 | 95.18 | 96.90 | 96.87 | 96.22 |

4 | 96.95 | 96.20 | 97.10 | 96.15 | 97.43 |

5 | 97.15 | 96.53 | 96.67 | 96.53 | 97.43 |

6 | 98.13 | 98.22 | 96.87 | 97.53 | 98.53 |

7 | 97.45 | 96.42 | 97.54 | 96.21 | 97.53 |

8 | 98.40 | 97.43 | 97.10 | 96.83 | 97.53 |

9 | 97.61 | 96.76 | 98.30 | 97.85 | 97.43 |

10 | 98.09 | 96.78 | 97.53 | 96.56 | 97.53 |

Average | 97.16 | 96.37 | 96.93 | 96.74 | 97.26 |

In this research work, we proposed an efficient MBWO-CNN method to diagnose and classify covid-19 data. Preprocessed input image was tuned and synchronized according to static dimensions and later attribute selection was carried out. Tuned hyper parameters assisted in the collection of preliminary hyperparameters of CNN model. To conclude, the attribute vectors are categorized and the images were classified into proper class labels ascovid-19 or non-covid-19. A thorough and fairer research was accomplished to evaluate the efficiency of the suggested MBWO-CNN method in detecting the conduct. The outcomes were examined further on some aspects. The suggested MBWO-CNN method attained high classification accuracy of 97.43%, sensitivity of 96.68%, and specificity of 97.25%. Consequently, this research work proved the efficacy of MBWO-CNN method and well-thought-out as an efficient model to diagnose and classify covid-19 data. Soon, this suggested method will be used in IoT and diagnostic tools associated with cloud in order to support and assist distant sufferers.