key: cord-0928116-4pauo2qo
authors: Sujathakrishamoorthy; Mohan, Surapaneni Krishna; Priya, Veeraraghavan Vishnu; Gayathri, R.; Lorate Shiny, M.
title: Artificial intelligence–based solutions for early identification and classification of COVID-19 and acute respiratory distress syndrome
date: 2021-05-21
journal: Data Science for COVID-19
DOI: 10.1016/b978-0-12-824536-1.00024-1
sha: 1a6e9f69658da72115c1dea9fd0b3c8d700d1736
doc_id: 928116
cord_uid: 4pauo2qo

COVID-19 has spread all over the globe; the initial case was detected at the end of 2019. The identification of disease at an early stage is needed to provide proper medication and isolate patients to preventing the spread of virus. This chapter focuses on the application of an artificial intelligence–based enhanced kernel support vector machine (E-KSVM) approach to detect COVID-19 and acute respiratory distress syndrome (ARDS). KSVM is enhanced by the use of the particle swarm optimization algorithm to tuning the parameters of KSVM. First, preprocessing takes place to remove unwanted details and noise. This is followed by the Hough transform to extract useful features from the image. Finally, the E-KSVM model is applied to classify images into normal, COVID-19, and ARDS. An extensive set of experimentations takes place on a chest X-ray dataset and ensures that the E-KSVM model has the ability to detect the disease effectively. The simulation outcome indicates that the E-KSVM model attains a maximum sensitivity of 72.34%, specificity of 75.20%, accuracy of 74.01%, and F score of 73.94% with a minimum computation time of 8.039s.

used an arithmetic approach to investigate and detect the infection of the SARS epidemic. The reproduction values for two various groups, Hong Kong and Toronto, were 1.2 and 1.32, respectively. This chapter presents an enhanced kernel support vector machine (E-KSVM) to detect COVID-19 and acute respiratory distress syndrome (ARDS). The E-KSVM model operates on three levels: preprocessing, feature extraction, and classification. Once the images are preprocessed, Hough transform (HT) is applied as a feature extractor and the E-KSVM model is employed as a classifier. The KSVM is enhanced by the use of the particle swarm optimization (PSO) algorithm to tune the parameters of KSVM. An extensive set of experimentation takes place on a chest X-ray dataset and the classification of images takes place for three categories: normal, COVID-19, and ARDS. 

The images gathered from a database might have irregular data as well as background noise. Here, the preprocessing phase is mostly applied to eliminate noise from computed tomography (CT) images and develop noise-free images applicable for future computations. Various morphological tasks have been processed to produce a mask. Dilation and hole filling are core tasks in some references from a binary edge map of an image under the application of a gradient driven threshold approach:

where GT is the gradient threshold that applies Otsu's adaptive framework. The binary image is dilated with the help of a diamond structuring component. Then, the mask is improved with an actual image. The two predefined images depict the mask produced by a gradient model and a label removed image. The key objective of a dilation task in the binary image is to introduce a higher lung region to compute in the further stage. The major step in future is to avoid the pectoral image from CT images. Such muscles are composed with nearer intensity measures compared with tumor intensity. Thus, the muscle had to be rejected from an image to attain the effective feature extraction with the application of a maximization approach, also termed segmentation models. As a result, it classifies data values according to the higher likelihood condition. The major constraint for maximization determination is:

The application of equation maximization provides a feasible way to compute the maximum likelihood function. To eliminate the pectoral muscle and maintain the remaining lung region, four intensity class segmentations were processed on CT images under the application of maximum estimation. 

Hough transform is defined as the feature extraction method applied in digital signal processing to evaluate shape parameters from corresponding boundary points. Hence, HT has been used to detect random shapes. Normal parameterization is provided by:

HT is referred as a tolerant space in edges and is free from noise, which is a derivation of arbitrary transform. It provides points from a diverse angle. Canny edge prediction is applied in the preprocessed image before employing the HT. In addition, it projects an optimal edge prediction filter to split the edges with the application of the first derivative of a Gaussian. These operations are used to decrease the preprocessing duration and offer a reliable data source that withstands geometrical as well as ecological modifications to calculate HT. It applies measures for every edge point (X, Y) for the image estimated from the given function. In addition, nonanalytical space is estimated in Eq. (33.5) with a particular batch of boundary points. For shape q, it is named P in Eq. (33.6):

For each X B , r is calculated and it stored as function 4. The value for r for each pixel X of gradient function 4ðX Þ in an image is calculated in Eq. (33.8) and is stored in the accumulator:

AðX þ rÞ (33. 8) Under the application of an accumulator, HT results in preprocessed CT images. Then, some features are selected from the transform image. Only the effective features should be selected; inefficient features might decrease the efficiency of the classification model. Here, the work intensity features were chosen and the application of these features results from the complexity of interpretation. It has the well-defined masses, speculated mass, illdefined mass, architectural distortion, asymmetry, and so on. The intensity features applied for mean, variance, entropy, and standard deviation are highly effective.

Many enhanced methods in SVM are deployed and the KSVM is a well-known and productive technique. Therefore, the merits of KSVM are that it can be processed from diverse scenarios for natural language processing, biomedical, and computer visibility; there are few tunable parameters; and training often employs quadratic optimization.

Thus, the main aim is to remove the combined local minimum depicted by alternate statistical learning models such as neural networks.

Let the p-dimensional training dataset of size N be:

where y n is either À1 or 1 implies the class 1 or 2. Each x n is a p-dimensional vector.

The maximum-margin hyperplane that divides class 1 from class 2 is task of SVM. In general, the hyperplane is described as provided in Eq. (33.10):

where , shows the dot product and w is a normal vector. There is a requirement to choose wandb to improve the margin from two parallel hyperplanes. Hence, two hyperplanes can be represented as:

This task has been assumed to be an optimization issue. It is the main aim is to enhance the distance from two parallel hyperplanes, which refers to removing data falling into a margin. With the application of easy numerical knowledge, the issue can be expressed as: s:t:y n ðwx n À bÞ ! 1; n ¼ 1; . N

Practically, there is a lack of hyperplanes that divides the methods effectively. To solve the issue, a soft margin approach was used that selects a hyperplane to divide accessible instances with the limitation of an increasing distance of closer samples. A positive slack variable x n is established to choose the degree of misclassifying sample x n . Then, the optimal hyperplane that isolates the data might be attained by using Eq. (33.14):

x n (33.14)

s:t: y n ðwx n À bÞ ! 1 À x n x n ! 0; n ¼ 1; . ; N where P represents the error penalty. Therefore, the optimization is developed as a trade-off for a big margin as well as tiny error penalty. The restraint optimization problem can be resolved under the application of a Lagrange multiplier:

x n À X N n¼1 a n ½y n ðwx n À bÞ

The min-max problem is hard to predict; thus, the dual form model is applied to deal with such a problem.

Eq. (33.14) might be equated as:

a m a n y m y n K ðx m ; x n Þ (33.16)

0 a C; X N n¼1 a n y n ¼ 0; n ¼ 1; . ; N A benefit of the dual form is the absence of slack variables, x n , along with constant C, which exists from the shortcomings of Lagrange multipliers.

Linear SVM is composed with limitations on the linear hyperplane, which is capable of dividing complex realistic data. To normalize in the form of nonlinear hyperplane, the main objective is applied to SVM. The final method is scientifically the same; each dot product is replaced with a nonlinear kernel function. The KSVM allocated the maximummargin hyperplane from a converted feature space. The transmission may be nonlinear, and converted space may be highly dimensional. Although the classifier is a hyperplane from a high-dimensional feature space, it can be nonlinear in an imaginative input space. For a single kernel, the least variable is required to develop a kernel as flexible and to modify it for practical data. It is applied with an radial basis function (RBF) kernel because it is efficient for producing a remarkable outcome. The kernel is expressed as: 

a a a b y a a b exp À kx a À x b k 2s 2 (33. 19) 

The PSO technique has been used to process parameter optimization. Also, it is a global optimization method accelerated from the behavior of bird flocking or fish schooling. It is a simple to model and provides rapid execution. PSO computes the searching task by using a set of particles maximized for each iteration. To derive the best solutions, all particles have to be shifted in the direction of preceding best ðp best Þ and best global ðg best Þ positions in the swarm:

st:fitnessðp m ðk à ÞÞ ¼ min

. ; t ½fitnessðp m ðkÞÞ;

where m is a particle index, P implies the particle number, k denotes a round index, t refers to the iteration value, and p is a position [30] . The extension of velocity and position of particles are carried out by applying Eqs. where V is the velocity. Inertia weight w used to manage global exploration and local exploitation. r 1 and r 2 are uniformly distributed with random parameters from the range of ð0; 1Þ c 1 and c 2 are acceleration coefficients. At this point, particle encoding is operated with variables C and s in Eq. (33.22).

This work employs the computation of fivefold cross-validation to attain the best tradeoff from processing complexity as well as reliability evaluation. The whole dataset is divided into five mutually unique subsets with same size, which have four subsets for training; the last subset is used to test the model. It has been iterated around five times; hence, each subset is applied in testing. The five fold (FF) of the PSO model to select the classifier accuracy is:

y s y s þ y m (33. 24) where y s and y m are the number of effective classification and misclassification, respectively. PSO is implemented for a higher FF function.

The dataset holds a set of 30 images, including 10 images each for the COVID-19 [31], ARDS, and normal categories. Some sample images in the three categories are shown in Fig. 33.2 . Table 33 .1 provides a comparative analysis of models. Fig. 33.3 depicts a comparative analysis with respect to sensitivity and specificity. The random tree (RT) method requires a lower sensitivity and specificity of 72.34% and 69.04% whereas the SVM approach attains a better sensitivity and specificity of 70.58% and 73.49%. However, the E-KSVM accuracy of 73.10%. Therefore, the proposed E-KSVM model has effective results and obtained a higher accuracy of 74.01%. Fig. 33 .5 shows a comparative analysis with respect to the F score. The RT model requires a lower F score of 66.77% and the SVM model accomplishes a slightly gradual F score of 72.76%. Hence, the projected E-KSVM approach has a productive outcome with the best F score of 73.94%. 

This chapter presents an E-KSVM model for the detection of COVID-19 and ARDS. The input image was preprocessed to eliminate unwanted details and noise. Then, the features in the preprocessed image were extracted by HT. Finally, the E-KSVM model was executed to classify feature vectors into appropriate classes: normal, COVID-19, and ARDS. A detailed experimental analysis was performed on a chest X-ray dataset and confirmed that the E-KSVM model has the ability to detect the disease effectively. The simulation outcome indicated that the E-KSVM model attained a maximum sensitivity of 72.34%, specificity of 75.20%, accuracy of 74.01%, and F score of 73.94% with a minimum computation time of 8.039s. In future work, the experimental outcome can be further increased using deep learning concepts. 

Artificial intelligence distinguishes COVID-19 from community acquired pneumonia on chest CT

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

Mapping the Landscape of Artificial Intelligence Applications against COVID-19

Rapid Ai Development Cycle for the Coronavirus (Covid-19) Pandemic: Initial Results for Automated Detection & Patient Monitoring Using Deep Learning Ct Image Analysis

A deep learning algorithm using CT images to screen for Corona Virus Disease (COVID-19)

Towards an artificial intelligence framework for data-driven prediction of coronavirus clinical severity

Artificial Intelligence (AI) applications for COVID-19 pandemic

A Novel Ai-Enabled Framework to Diagnose Coronavirus Covid 19 Using Smartphone Embedded Sensors: Design Study

Development and Evaluation of an AI System for COVID-19 Diagnosis. medRxiv

Estimating Uncertainty and Interpretability in Deep Learning for Coronavirus (COVID-19) Detection

Performance of radiologists in differentiating COVID-19 from viral pneumonia on chest CT

Artificial intelligence applied on chest X-ray can aid in the diagnosis of COVID-19 infection

Artificial intelligence in the battle against coronavirus (COVID-19): a survey and future research directions

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

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

Transmission scenarios for Middle East Respiratory Syndrome Coronavirus (MERS-CoV) and how to tell them apart

Genomic characterisation and epidemiology of 2019 novel coronavirus: implications for virus origins and receptor binding

Novel Coronavirus: Where We Are and what We Know

Clinical Characteristics of 2019 novel coronavirus infection in China, medRxiv (2020). Available vonline

Estimating the unreported number of novel coronavirus (2019-nCoV) cases in China in the first half of January 2020: a data-driven modelling analysis of the early outbreak

Estimation of the asymptomatic ratio of novel coronavirus infections (COVID-19), medRxiv

Estimation of the transmission risk of the 2019-nCoV and its implication for public health interventions

Novel coronavirus outbreak in Wuhan, China, 2020: intense surveillance is vital for preventing sustained transmission in new locations

Real-time estimation of the risk of death from novel coronavirus (COVID-19) infection: inference using exported cases

Ensemble forecast of human West Nile virus cases and mosquito infection rates

Comparison of four different time series methods to forecast hepatitis A virus infection

Forecasting seasonal outbreaks of influenza

Real-time influenza forecasts during the 2012e2013 season

Forecasting versus projection models in epidemiology: the case of the SARS epidemics

An MR brain images classifier system via particle swarm optimization and kernel support vector machine