A Healthcare System for COVID19 Classification Using Multi-Type Classical Features Selection

1  Introduction

The research on coronavirus is going on from the last two decade [1], however it got much intention by the researchers of Science & Engineering domain in general and Medical Sciences in particular after its spread in Wuhan city, Hubei Province in mainland China [2]. Angiotensin-converting enzyme 2 (ACE2) is a receptor for SARS-CoV, the novel coronavirus that shows the random symptoms like fever, pain, nasal congestion, running nose and acute respiratory syndrome that in some serious situations leads towards death. The major human coronavirus category is SARS-CoV and the novel coronavirus SARS-CoV−2 that causes coronavirus disease in 2019 [3]. Once it has been declared an epidemic, the health sectors of even first world countries were not equipped with necessary medical instruments required to cope with the influx of patients in any specified area [4]. To date, the confirm corona virus cases are 4,253,802, worldwide. The total deaths are 287,250 and United State (USA) is a highly affected with this virus. The need of computer technology arises that provide the medical specialist an intelligent input about the incoming patient so the resources of the hospitals can be intelligently utilized.

Keeping in view the great success of artificial intelligence methods in medical imagery a number of scientists undergoes exploitation to these tools in recent Covid-19 situation. In this regard the most recent literature survey is presented. A predictive performance deep learning models based on neural network has been exploited to distinguished Covid-19 from community acquired pneumonia on chest [5] by extracting the visual features from volumetric chest CT exams. The real time multi-step forecasting of Covid-19 to estimate its spread and its life span across the china is proposed in [6] to help public health planning and policymaking. Identification of Covid-19 cases quickly to reduce spread in susceptible population a machine learning based method is proposed [7]. The applicability of AI methods for detecting infected people using chest radiography images is an alternate quick method of accurate screening proposed in [8] using a deep convolution neural network, although the method required large dataset for better accuracy and reliability of the method. However, currently researchers facing the problem of the availability of Chest CT Scans. In this view, we believe that the employing of classical features like texture and local points are more effective to deal with less number of images. Therefore, we collected 16 patients data which publicly available. The data consists of both normal and confirmed COVID19 Pneumonia CT Scans. A few sample images are shown in Fig. 1.

Figure 1: Sample Scans of healthy and Positive COVID19

This motivates the authors to study some alternate technology based mechanism that can classify the COVID-19 infected people in the categories of only infected, mild and savior-infected patients that need immediate intention by the medical specialists [9]. A stack of the dataset obtained from Chest CT Scan for various patients motivate the authors to incorporate method based of artificial intelligence to perform classification and predictions with a high level of accuracy. As AI inspired methods are a powerful tool so as the increase in dataset from private and public hospital in future would even strengthen the study. Farooq et al. [10] presented a convolutional neural network (CNN) based method for discriminating COVID19 patients based CT scans. They used publically available data for fine tuning of pre-trained ResNet-50 CNN architecture. For training a model, they used different input sizes like 224×224×3, 128×128×3, and 229×229×3. The presented model achieves a state-of-the0art prediction accuracy on selected dataset. Afshar et al. [11] trained a CapsuleNet based CNN framework using Chest X-ray images for the identification of COVID19 patients. The main functionality of described method is to easily train on small available dataset and gives sufficient accuracy. Pereira et al. [12] introduced a classification scheme for COVID19 pneumonia and healthy patients using chest X-ray (CXR). Initially, they introduced an approach to tackle the problem of imbalance data. Then, they extract texture features using pre-trained models and performed fusion. For evaluation, they used Rydls-20 dataset and achieved notable performance. Ozturk et al. [13] showed the importance of AI for diseases detection. Inspired by AI, they presented an automated system based on you only look once (YOLO). For training a model, they used Chest X-ray images. Then, two types classification problem is performed-binary class and multi class, where the attained accuracy was sufficient which supports the proposed technique. Amyar et al. [14] presented a deep learning architecture for COVID19 lesion segmentation and classification. They used Chest CT scans for the evaluation of presented technique. In the segmentation phase, they used encode and decoder. Then, features are extracted from CNN model and perform classification using multi-layer perceptron. The results support the performance of presented technique. Luz et al. [15] also presented a quick deep learning technique for identification of COVID infected patients. For this purpose they used Chest X-ray Scans. They inspired by the slow testing procedure of COVID19 patients and tried to support the health care system. Rahimzadeh et al. [16] described a fused network based deep neural network architecture for classification of COVID19 X-rays. They used Xception and Resnet and trained on collected data. Then, they extract features from both and fused their information. They conclude that the fusion of multiple models information gives better results. Kassani et al. [17] described a computer aided diagnosis (CAD) system for identification of COVID19 patients using Chest CT and X-ray images. They used many CNN models and get the important features all of them. Later, they fused all collected features and passed to multi-class support vector machine (SVM) for final identification. In addition, a critical analysis is also conducted for examine the automated techniques which are using for COVID infection identification from X-ray images [18].

The above techniques are summarizes in this text. These studies mostly based on deep learning, features fusion, and reduction. However, it is noticed that the deep learning models required a huge amount of data for training; however, currently sufficient data is not available. Therefore, it is required a data augmentation which is not a perfect approach and not assurance of better performance. The second step which is followed by above researchers is fusion of features. However, the fusion of features also added little redundant information which is resolved through selection of robust features or reduction techniques. The reduction techniques are not provides better results because it is high chances that few important features are discarded but in the selection process only those features are selected whose meet the defined criteria. In this work, we focused on classical features that are fused using proposed Max-Correlation maximization approach. Later, features are optimized using PSO along with extreme learning machine (ELM) fitness function.

2  Proposed Methodology

In this work, we proposed a new method for prediction of COVID19 from Chest CT scans based on the fusion of multi-type features and particle swarm optimization–extreme learning machine (PSO-ELM) based features optimization. The proposed method is shown in Fig. 2. In this Figure, it is shown that proposed method consists of four principle steps-multiple feature extraction using classical techniques(i.e., DCT, DWT, and DRLBP), Max-Covariance Maximization approach for fusion of all these features, optimization of features using PSO along ELM fitness function, and finally optimized features embed in SVDD classifier for final classification.

Figure 2: Proposed multi-type features fusion and optimization using PSO-ELM for prediction of COVID19 patients using CT Scans

2.1 Database of CT Images

A total of 7000 images are collected from Radiopaedia [19] which includes 4000 COVID19 positive CT scans and rest of them are healthy. From this website, 58 patients CT scans are available for research purpose. The confirmed COVID19 positive cases are tested by RT-PCR testing kit. The confirmed cases are also includes the travel history of patients. From the total CT images, we used 2000 COVID19 Positive and 1500 healthy scans for training the classifier using proposed method and rest of them for testing. A few sample images are shown in Fig. 1. However, from these scans, it is not easy to get 100% accuracy due to complex texture problems. Also, the most of the region in the image is healthy, so the prediction of COVID19 and healthy scans using machine learning is depends on the nature of extracted features.

2.2 Features Extraction

In this modern era, demand of real time solutions has focused the research community to develop highly efficient and accurate algorithms. Extraction of features from image dataset for the purpose of classification plays a vital role in the development of such systems. These features must be highly uncorrelated and have significant variations. Many feature extraction techniques are found in the literature and their significance has been proven in solving many problems [20]. However, it has been observed that features performing better in one domain may not be equally effective in other domains. It leads us to use the combination of better performing features. In this manuscript we have used Discrete Wavelet Transform (DWT), Discrete Cosine Transforms (DCT), Dominant Rotated Local binary Pattern (DRLBP) and Shannon Entropy along Kurtosis features. Significance, usability and detail of each of these feature extraction techniques have been discussed in the following subsections.

2.2.1 Discrete Wavelet Transform (DWT)

Wavelets have been effectively used for extraction of features from image dataset for more than 30 years. It has the property of representing image at multi-resolutions without loss of information. Moreover, it captures local information present in spatial domain and represents it in frequency domain. DWT transforms the input image into four sub-bands at first level of decomposition [21]. The next-level decomposition is done using the same way as presented in [21]. Other sub-bands can also be used for decomposition but they possess different properties as HH sub-bands are insensitive to illumination variations. We have used approximation sub-band as they carry much information.

Consider, ω(u,v) is an original 2D image of dimension R×C, where (R,C)∈R. The DWT of ω(u,v) is computed using following mathematical formulation:

Dw(j,r,c)=1RC∑u=0R−1∑v=0C−1ω(u,v)ψj,r,c(u,v)(1)

Df(j1,r,c)=1RC∑u=0R−1∑v=0C−1ω(u,v)φij,r,c(u,v)(2)

Here, Dw(.) represent scaling function, Df(.) is wavelet function, i∈[H,V,D], where H denotes horizontal, Vdenotes vertical, and D represents diagonal sub-bands, respectively. The other notations like r denotes row pixels, c denotes column pixels, and j represent decomposition function. The 2D transform scaling function ψ and wavelet transforms φ are computed for the measurement of intensity variations in different directions i.e., horizontal, vertical and diagonal. These directions are formulated as follows:

ψu,v=ψ(u)ψ(v)(3)

φu,vH=φ(u)φ(v)(4)

φu,vV=φ(u)φ(v)(5)

φu,vD=φ(u)φ(v)(6)

Detailed of the scaled and transformed basis functions are calculated using the following formulation:

ψj,r,c(u,v)=2j/2ψ(2ju−r,2jv−c)(7)

φj,r,c(u,v)=2j/2φi(2ju−r,2jv−c)(8)

where, i=[HVD]. Visually, the DCT features are computed for 2D image is shown in Fig. 3. In this figure L and H are low and high pass filters, respectively. As shown in this figure, four outputs are obtained and dimension of each direction is 131×131 because length of R=C=256. After that we combined these directions in one array A, where the size of A is 524×130. However, our target is to get a one-dimensional array, so we applied variance and get a vector of 1×524. Mathematically, the variance process is defined as follows:

V(i)=1(R−1)(C−1)∑u=1R−1∑v=1C−1|Puv−μ|2(9)

Figure 3: Computation of DWT features from Chest CT image

2.2.2 Discrete Cosine Transform (DCT)

DCT is one of the old and most widely used techniques for extraction of features [22]. It has high information packing capability and thus is used for compression of images. This property enables us to extract significant information from images. Image data after applying DCT is transformed into frequency domain where data with high frequency and low frequency can be separated. Low frequency features have high variation and contain useful information suitable for classification problems. DCT is computed of an image ω(u,v) of size R×C using the following formulation: Here r and c are the corresponding rows and columns of the input image and point (u,v) contain pixel value intensities and point (x,y) are the corresponding output ranging x={0,1… N−1} and y={0,1… M−1}.

DCT(x,y)=α(x)α(y)∑u=0R−1∑c=0C−1Cos[πx2R(2u+1)]Cos[πy2C(2v+1)]ω(u,v)(10)

α(x),α(y)={1Rfor(x,y)=02Rfor(x,y)≠0(11)

As a output, a resultant vector of dimension 256×256 is obtained, but we required a single dimensional vector; therefore, we applied a variance formulation to convert matrix into a single vector. After applying variance, we obtained a vector of dimension 1×256.

2.2.3 Dominant Rotated Local Binary Pattern (DRLBP)

Local Binary Pattern (LBP) has been used as an efficient feature descriptor by many researchers working in the field of computer vision [23]. It captures local information but suffers degradation in case of rotation of image dataset. Variants of LBP descriptor have also been proposed but rotation invariant feature descriptors lack local orientation information and results in deterioration in accuracy because of less of discriminative features. Dominant Rotated Local Binary Pattern (DRLBP) proposed by Mehta et al. [24] preserves local information along with rotation invariance. This technique not only captures local distribution of pattern but also holds information regarding types of patterns. These characteristics make it a more powerful descriptor. The DRLBP computes the feature in a local circular region by taking the difference of the central pixel with respect to its neighbors as presented in equation below:

DRLBPN=∑n=0N−1s(gn−gm).2mod(n−D,N),   s(gn−gm)={ 1 gm>=gn0 gm<gn  (12)

Here gm and gn are the gray level values of the middle pixel and its neighboring values respectively and n represents the index of the neighboring pixel. Value of n varies from 1 to N − 1. We have used these features by taking only one as radius of the circular neighborhood so N is 9 in our case. D is considered as a pixel whose value is maximum among those neighboring pixels as given in the given equation, where the dimension of output vector is 1×59.

D=argmaxn∈(0,1,…,N)(gn−gm)(13)

2.2.4 Shannon Entropy Along Kurtosis Features

The statistical features based on Shannon Entropy along Kurtosis are described in this section with detailed mathematical analysis. In the initial phase, we are calculating the entropy features of each image in the form of row and column pixel. As we have image ω(u,v) of dimension 256×256 which depicts that 256 rows and 256 columns. First, we compute the kurtosis of image, which return the texture features. Mathematically, kurtosis is described as:

Kr(i)=E(ω(u,v)−μ)σ4,σ=1(R−1)(C−1)∑u=1R−1∑v=1C−1|ω(u,v)−μ|2(14)

where, E(.) is expected value, μ is mean value of image, and σ is a standard deviation of original image. This formulation return a one-dimensional vector of dimension 1×256 and for N images, its length is N×256. After that, entropy is computed of kurtosis vector Kr(i). A single value is return by entropy which we multiplied with kurtosis vector. The main purpose is to give a chance of maximum feature before setting a threshold function. As shown in Fig. 4, the histogram, QQ plot, and probability plots are added. From these plots, it is shown that after the multiplication, the maximum features occurrence is between 2 and 8, so by following this, we set a threshold function as follows:

Kr(i)={kr(i)~for(x1≤Kr(i)≤x2)∈Kr(i)⏞IgnoreOtherfeaturesValues(15)

where, kr(i)~ denotes final entropy along kurtosis vector, Kr(i)⏞ denotes updated kurtosis vector after multiplication, and (x1,x2) depicts lower and upper bound, respectively. Hence, the final dimension of feature vector is N×240 for this work.

Figure 4: Representation of Entropy along Kurtosis features

2.3 Maximum Covariance Based Features Fusion

In the last few years, features fusion shows a huge performance in the area of medical imaging [25]. The main purpose of features fusion is to get the important information of one problem from several places to get more detailed information [26,27]. Many fusion techniques are presented in the literature and most of them are used in medical imaging [28,29]. In this work, we implemented a Maximum Covariance based technique for features fusion. This technique is working in parallel process. Initially, make the equal length of each vector based on their mean padding. After that, calculate the Maximum Covariance among pair of features. The positive maximum covariance value based features are shifted in a fused vector. Mathematically, this process is formulated as follows:

As we have four feature vectors names-DWT, DCT, DRLBP, and Shanon Entropy along Kurtosis, denoted by Δ1[r×k], Δ2[r1×k], Δ3[r3×k], and Δ4[r4×k], respectively. The length of each feature vector is N×524, N×256, N×59, and N×240, respectively. Based on the mean padding, make the equal length of each vector at the very first step, where final length of each vector is N×524. We consider two feature vectors Δ1[r×k] and Δ2[r1×k], where k represents the sample images that are used for feature extraction, r and r1 represent features for v1 and v2, respectively. Let x1 be an arbitrary column vector for r feature matrix (v1) and x2 for r1 feature matrix (v2). Then data projection for 1×k row vector is defined as:

a1=x1TΔ1,a2=x2TΔ2(16)

Through MCA, perform the optimal feature patterns x1 and x2 that maximize their covariance as follows:

C=Cov[a1,a2](17)

=Cov[x1TΔ1,x2TΔ2](18)

=1k−1[x1TΔ1(x2TΔ2)T](19)

C=x1TCv1v2x2(20)

Cv1v2=1k−1Δ1Δ2T(21)

where, Cv1v2 represent covariance matrix among v1 and v2 whose ith and jth feature is the covariance of v1i(t) with v2j(t). This process is continued for all features of v1i∈Δ1 and v2j∈Δ2. Based on maximum covariance, a feature pair is selected and share in a fused vector denoted by ξ(k). Similar, this process is also performed for other feature vectors and in the end, a final fused vector is obtained of dimension N×K. The visual representation of this fusion process is shown in Fig. 5. However, all features are not relevant for final prediction, so it is essential to remove few of them. Therefore, we implemented improved PSO algorithm. The detail of features selection is given in Section 2.4.

Figure 5: Representation of features after Maximum Covariance based Fusion

2.4 PSO Based Feature Optimization

PSO is an efficient and one of the most widely used optimization problem solving technique [30]. It was proposed and developed by James Kennedy and Russell Eberhart in 1995. Like GA, it is a population-based algorithm developed using the concept of evolutionary theory inspired by social behavior of bird flocking. Individuals in the population are called particles combined to form a swarm. Each particle in the swarm is identified by its position and velocity in a multi-dimensional search space. All the particles moves to discover their best position in the search space by keeping record of their previous best positions achieved and try to continue improvement in finding the global best position. Global best position is governed by fitness function. PSO tries to optimize the fitness function until the stopping criteria is met. PSO starts with the initialization of random population of n-individuals known as particles. In this work, we have fused feature vector of dimension N×K, number of selection is 500, total populations 100, and max-iterations 500. Each particle in the population is a solution in the n-dimensional search space n×d having d decision variables. After initialization of parameters, fitness is calculated using ELM classifier [31]. Based on fitness function, if position of kth feature is 1, then taken this feature as a useful feature and if position of kth position feature is 0, then it not chosen. Mathematically, position and velocity of particle k is defined as follows:

Vk=vk1,vk2,….vkd(22)

Xk=xk1,xk2,….xkd(23)

Here, value of k=1,2,3,…n. ELM is employing is a fitness function to calculate the fitness of each particle. Initial best fitness of every particle is also its best so far and is denoted as Pbest,k and its best fitness function is represented by fpbest. The global best position is represented by gbest,k. The velocity and position of each particle are updated using the following equations:

Vkt+1=wVkt+c1r1(pbest,k−Xkt)+c2r2(gbest−Xkt)(24)

Xkt+1=Xkt+Vkt+1(25)

Here w is weight inertia parameter and is used to control the impact of previous velocity in the new direction. It creates a balance between exploration and exploitation. Small value of w decelerates the particle, whereas large values adds diversity and causes exploration to occur. c1 and c2 controls the cognitive and social parameters. Velocity equation consists of three components namely momentum, cognitive and social. Momentum part prevents particles from abrupt changes and cognitive component quantifies the performance relative to previously achieved performances. Social component contributes in achieving the best position determined by the swarm. A brief description of features selection using PSO is given in Algorithm 1. The visual representation of feature space after selection of best features is shown in Fig. 6. These selected features are finally passed in SVDD classifier for final prediction [32]. The decision function of SVDD is defined as:

g(x)=sign(||δ(x)−α||2−R2)(26)

={1x∈HealthyScan−1x∈COVID19Scan(27)

where, g(x) represent decision function, δ(x) is mapping of feature space, α is a center, and R2 represents distance among α of hypersphere and any of the support vectors on the boundary. This function returns 1 for healthy scan and −1 for COVID19 positive samples.

Figure 6: Representation of features after applying PSO for feature selection

3  Results and Discussion

The detailed experimental results are presented in this section with detailed quantitative measures and graphical plots. For detailed analysis of proposed results, two steps are followed-fusion of features and most discriminant selected features. The totals of 40 patients Chest CT Scans are utilized to evaluate the proposed method. Several performance measures are considers like sensitivity rate, specificity rate, precision of COVID samples, F1-Score, AUC, error rate in terms of FNR, and Accuracy. These measures are calculated for only testing data. As in this article, we consider the 50% images for training the proposed scheme and rest of them for testing. A detailed training process proposed method as well as testing process is shown in Fig. 7. In this figure, it is shown that the fused features are optimized through PSO and trained SVDD model. After that, testing step is performed and same features are extracted and fused. Later, PSO is applied and optimize features that are passed in trained model. The next step is most important called features matching and in the output, a predicted labeled output is attained. The proposed method is implemented in MATLAB2020b using Personal Desktop Computer of 16GB RAM, 1 T Hard, 256 SSD, and 8 GB of Graphics Card.

Figure 7: Training and testing process of proposed system

3.1 Quantitative Results of Features Fusion Step

This section summarizes the proposed features fusion results. Before discussion of the efficiency of fusion step, we talk about separate features efficiency. We embed original features in PSO and computed accuracy on based classifier. On the base classifier, the computed accuracies are 76.4%, 72.9%, 68.3%, and 76.77% for DWT, DCT, DRLBP, and Entropy-along-Kurtosis. However, tis accuracy is not enough; therefore, we implemented a fusion approach based on maximum covariance. The results of this process are given in Tab. 1. In this table, it is shown that SVDD is employing as a base classifier and other classifiers are used for further analysis of proposed method. The maximum reported accuracy is 88.6%, whereas the other calculated measures are sensitivity (88.5%), specificity (91.0%), precision (88.5%), F1-Score (89.73%), area under the curve (AUC) is 0.99, and 11.5% FN rate, respectively. Fig. 8 confirms the accuracy of proposed method. In this figure, it is illustrated that COVID19 scans are 86% correctly predicted, whereas healthy scans are 91% appropriately predicted. The error rate of each class is 14% and 9%. Also, the receiver operating characteristic (ROC) plots are also added and illustrated in Fig. 9. From this figure, we can verify the AUC of proposed method using fusion approach. These measures are also calculated on other selected classifiers like Fine Tree, linear discriminant analysis (LDA), Gaussian naïve bayes (GNB), kernel naïve bayes (KNB), linear support vector machine (LSVM), cubic SVM (CSVM), fine K-nearest neighbor (FKNN), and Ensemble tree (E-Tree), where the attained accuracies on each of them are 86.7%, 88.3%, 88.5%, 81.1%, 87.7%, 87.2%, 86.8%, and 86.2%, respectively. The sensitivity and specificity rates of these classifiers are (86.5%, 84.0%), (88.5%, 90%), (88.55%, 90.0%), (81.0%, 67.0%), (87.5%, 90%), (87.0%, 88%), (87%, 79%), and (86.0%, 90%), respectively. The F1 score, false negative rate (FNR), AUC, and precision rate are plotted in Fig. 10. From this figure, it is shown that the efficiency of KNB is low as compared to other classifiers. Moreover, it is also shown that the proposed fusion process gives tremendous accuracy as compared to single feature vector.

Figure 8: Confusion matrix of proposed features fusion step

Figure 9: ROC plots of proposed method after fusion of multiple features

Figure 10: Visual representation of calculated performance measures for all classifiers

3.2 Quantitative Results of Features Selection

This section summarizes the results of proposed features selection using Improved PSO algorithm. Tab. 2 reviews the quantitative results. The highest accuracy after employing this step is 93.1% which is increased 4.6% as compared to fusion steps. The sensitivity and specificity of proposed method are 93.2% and 92% which are verified through Fig. 11, in terms of confusion matrix. In this figure, it is shown that the correctly predicted COVID19 scans are 94% and healthy scans 92%. However, the error rate is 6% and 8%. If we compared this figure with Fig. 8, then it is perceived that the prediction rate of the selection step is increased as compared to fusion step. ROC plots are also illustrated in Fig. 12, which supports the attained AUC. The comparison of proposed accuracy is conducted with few other classification techniques like Fine Tree, LDA, GNB, and name a few more and achieved accuracies of 92%, 92.1%, 92.8%, 92.6%, 92.9%, 91.8%, 90.8%, and 92.3%, respectively. Also, the calculated values of other measures like sensitivity and specificity are (92%, 90%), (92%, 90%), (92.55%, 88%), (92.4%, 96%), (93%, 95%), (92%, 91%), (91%, 89%), and (92.5%, 90%), respectively. These values are calculated for top-to-bottom classifiers except SVDD (proposed). Moreover, F1-score, FNR, precision, and error rates are plotted in Fig. 13, that supports the authenticity of features selection performance. If we compared these measures with Tab. 1 and Figs. 11‐13 then it can be show that the performance of selection method is improved for all classifiers. From results, we can easily say that the performance of an automated system is always based on the quality of features. If features are relevant to the problem, then results are improved, otherwise degrades.

Figure 11: Confusion matrix of SVDD using proposed framework

Figure 12: ROC plots of proposed method after features selection

3.3 Discussion

This section presented the detailed discussion of presented method in terms of calculated performance metrics and visual plots. As described in the introduction section that we are dealing with Chest CT scans for prediction of COVID19 positive and healthy images. These sample scans are showing in Fig. 1. The main problem is classifying these scans using computer vision. For this purpose, we proposed a, integrated model based on multiple features fusion and improved PSO based selection. Fig. 2 shows the main architecture diagram. In this figure, it is showing that multiple features are extracted and fused in one matrix using Maximum Covariance approach. The fusion process improves the image information for correct prediction; however, it is chance to add many irrelevant features as visual representation is showing in Fig. 5. The quantitative results are presented in Tab. 1, which shows the maximum accuracy is 88.6%. Figs. 8, 9 and 10 also support the proposed accuracy. Later, improved PSO is applied on fused feature matric and selects the most informative features. Visually, these features are showing in Fig. 6. From this figure, it is shown that the number of features is minimized after applying improved PSO. Tab. 2 shows the quantitative results of results of proposed selection method with the accuracy of 93.1%. Also, Figs. 11, 12 and 13 support this accuracy. This accuracy is improved as compared to fusion step. For comparison, we compute the accuracy of each feature type separately and then compare with fusion and selection process. Fig. 14 shows the feature based comparison. In this figure, it is observed that the accuracy of original extracted features like DWT, DCT, DRLBP, and Shanon entropy along Kurtosis (SEaK) is 79.8%, 83.2%, 80.6%, and 82.1%, respectively. However, after the fusion process, 6% accuracy is increases. Further, the selection process, also improves the features which effect show in the accuracy. This last step, selects only informative features that improves the accuracy of correct prediction. A few predicted results by presented system are shows in Fig. 15. These results are obtained in the testing process. Overall, proposed system give significant performance for prediction of COVID19 scans.

Figure 13: Visual representation of calculated performance measures after features selection for all classifiers

Figure 14: Features based comparison

Figure 15: Prediction of COVId19 positive and healthy Chest CT Scans using proposed method

4  Conclusion

Using CT images, this paper presents an automated method for predicting COVID19 Positive and Healthy Scan. The steps are as follows: database preparation, multiple feature extraction, feature fusion, selection of most informative features, and prediction. Four types of features are computed along Kurtosis such as DWT, DCT, DRLBP, and Shanon entropy (SEaK). Because these features alone did not provide enough accuracy, we fused them all together using the Maximum Covariance method. The fused feature matrix boosts prediction accuracy. However, it includes several redundant and inappropriate features. For this purpose, an improved PSO algorithm is used, yielding a unique feature vector. Finally, the SVDD classifier is used to make the final prediction. The outcome demonstrates the improved accuracy of the presented method. Based on the results, we conclude that the fusion process is important to increase the information against one image for correct prediction but it is not guaranteed that all features are important. Moreover, it is also concluded that the selection process improves the feature representation and prediction accuracy. In the future, most recent optimization techniques shall opt for better accuracy and minimum computational time [33–36].

Funding Statement: This research was supported by Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (P0012724, The Competency Development Program for Industry Specialist) and the Soonchunhyang University Research Fund.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the present study.