COVID-19 detection using deep convolutional neural networks and binary-differential-algorithm-based feature selection on X-ray images

The new Coronavirus is spreading rapidly, and it has taken the lives of many people so far. The virus has destructive effects on the human lung, and early detection is very important. Deep Convolution neural networks are such powerful tools in classifying images. Therefore, in this paper, a hybrid approach based on a deep network is presented. Feature vectors were extracted by applying a deep convolution neural network on the images, and useful features were selected by the binary differential meta-heuristic algorithm. These optimized features were given to the SVM classifier. A database consisting of three categories of images such as COVID-19, pneumonia, and healthy included in 1092 X-ray samples was considered. The proposed method achieved an accuracy of 99.43%, a sensitivity of 99.16%, and a specificity of 99.57%. Our results demonstrate that the suggested approach is better than recent studies on COVID-19 detection with X-ray images.


INTRODUCTION
COVID-19's rapid spread has resulted in the death of numerous people worldwide. Muscle aches, cough, and fever are all symptoms of the virus, which can be detected through clinical trials and radiographic imaging.
Medical imaging is critical for disease diagnosis, and disease X-rays and computed tomography (CT) scans can be used in the deep network to aid in the disease's diagnosis.
The process of classifying and diagnosing disease from an image using a neural network is divided into four steps: feature extraction, optimal feature selection, network training, and model performance test. The feature extraction step is divided into two types. In the first type, image processing techniques, algorithms, and filters extract the features. Among the features extracted from the images are the tissue shapes and textures used to classify patients. In the second type, the original images and their actual output class are fed into the convolution network as input data, and the features are extracted automatically in the final flattened layer following the network training process and weight adjustment.
Certain features extracted from the deep network may have a detrimental effect on classification accuracy [1].
As a result, effective feature selection methods are critical. There are three distinct types of feature selection methods. The filter method uses features' intrinsic properties and statistical indicators such as the fisher score, information gain, chi-square, and correlation coefficient. The wrapper method employs a learning algorithm that searches the feature space for a subset of features that optimize the classification accuracy. To this end, wrapper approaches employ metaheuristic methods for selecting feature subsets and performing cross-validation. Finally, the hybrid method employs both filter and wrapper methods [2]. Metaheuristic methods outperform other feature selection methods in applications where many features are required.
Classification performance is improved by analyzing extracted features from images and selecting the optimal features [3]. Numerous feature selection (FS) studies have been published in the field of medical imaging, including Robustness-Driven FS (RDFS) for lung CT images [4], Shearlet transform FS from brain MRI images [5], principal component analysis for lung X-Ray images [6], genetic algorithm (GA) for lung nodules [7], bat algorithm (BA) versus particle swarm optimization (PSO) in lung X-ray images, and the flower pollination algorithm (FPA) from lung images [8].
The studies above propose that machine vision combined with metaheuristic algorithms can classify patients using lung images. On the other hand, existing diagnostic methods for the COVID-19 virus using X-ray images require a large amount of memory, ample time, and a large number of features. As a result, an intelligent system appears necessary to assist doctors and treatment staff in accurately and rapidly classifying COVID-19 patients in reducing disease-related mortality. This research aims to develop an efficient procedure utilizing artificial intelligence methods to assist doctors and patients in accurately predicting COVID-19. The research is novel in that it employs a binary differential evolution algorithm to design a deep learning structure based on feature selection for COVID-19 diagnosis. The contributions of the study include: 1-Using a deep convolutional neural network without a pre-trained network to design an intelligent system based on lung X-ray images and extracting features with the least amount of memory required to create and train the network.
2-Selecting the optimal features of the differential metaheuristic method that improves performance indexes.
3-Increasing classification accuracy for multi-class problems, including patients with COVID-19, pneumonia, and the healthy group.
The study is structured as follows: Section 2 reviews related works. Section 3 presents the proposed methodology and model for COVID-19 detection using deep convolution and binary differential algorithms. Section 4 contains the experimental results, and section 5 discusses the method and compares this with prior works. Finally, the study concludes.

RELATED WORKS
Hemdan, Shouman, and Karar used deep learning models to infer the positive or negative status of COVID and reported that the VGG19 model performed better with an accuracy of 90% on 25 COVID infected and 25 non-COVID images [9]. Toaçar, Ergen, and Cömert incorporated 295 COVID images, 98 pneumonia images, and 65 normal images into MobileNet and SqueezeNet [10]. They extracted features from trained Net models and then used the SMO algorithm to select the features, with an overall accuracy of 99.27% reported for the SVM classifier.
Zhang, Xie, Li, Shen, and Xia investigated an 18-layer ResNet model for 100 COVID and 1431 pneumonia images and reported an accuracy of 95.18% [11]. Apostolopoulos & Mpesiana pre-trained VGG19 on 224 COVID, 700 pneumonia, and 504 normal images, where the results demonstrated a 98.75% accuracy [12]. The authors of [13] evaluated the DarkNet with 17 convolutional layers using 127 COVID, 500 pneumonia, and 500 normal images and reported an accuracy of 98.08%. In [14], the performance of CNN was improved via preprocessing image algorithms, resulting in a model with 94.5% accuracy.
The authors of [1] developed a COVID-19 classification method based on two datasets that combined a CNN named Inception, a pre-trained Imagnet as a feature extractor, the Marine Predators Algorithm as a feature selector, and a KNN as a classifier.
The first dataset contained 200 positive COVID-19 images and 1675 negative images, whereas the second dataset contained 219 positive COVID-19 images and 1341 negative images. Accuracy was reported as 98.7% for dataset 1 and 99.6% for dataset 2. Canayaz validated a COVID-19 diagnosis model by combining VGG19, ResNet, AlexNet, and GoogLeNet with two metaheuristic algorithms titled "binary particle swarm optimization" and "binary gray wolf optimization." The highest overall classification accuracy was 99.38% after binary gray wolf optimization was used to select features from 1092 X-ray images from the COVID-19, pneumonia, and healthy category records [15].
One of the previous works' limitations is their reliance on pre-trained deep networks, which require a large amount of memory. Additionally, many input features plus a lengthy detection time are further drawbacks of these methods. In this study, a deep learning approach based on feature selection is proposed via the binary differential evolution algorithm to overcome these limitations and improve COVID-19 detection.

Deep convolution
Convolutional neural networks are used in machine learning as a feature extractor and classification method.
The input to a convolutional network is the original data, such as images. The network extracts the features where IMG denotes the input image with height =H, width=W dimensions, and tc is the number of image channels, C is the filter matrix with c1*c2 dimensions, and bs is a bias value for each filter C, i=0…H, j=0…W Following convolution, the unwanted values are removed using the ReLu layer, and the input is then reduced using the pooling layer. The effective input vector is then passed to the fully connected layer, which functions similarly to the MLP. In the final section of the deep convolution layers, Softmax [17], classification layers perform classification using ADAM (adaptive moment optimizer) [18], the lost function is shown below (equation where M denotes the sample size, denotes the actual class for the mth sample, ̂ denotes the predicted output class for the mth input data, and Γ denotes the regularization coefficient.
ADAM is a gradient-based optimization algorithm that uses the exponential moving average of the gradient and the square of the gradient to update the neural network weights and solve deep network issues effectively.
The deep neural network comprises numerous layers, each with its own set of learning parameters, namely weights and biases. Applying the optimal feature selection algorithm to the ADAM optimizer increases the optimization's speed and accuracy.

Binary Differential Evolution
Differential evolution (DE) [19] is a heuristic evolutionary method for minimizing the continuous problem.
The concept of binary differential evolution (BDE) [20] is extended to address issues of feature selection. It is composed of three distinct builders, including mutation, crossover, and selection. Initially, dimensions D are used to generate the initial population, where D is the number of features to optimize. For the mutation operation, three random vectors 1 , 2 , 3 are selected for vector such that 1 ≠ 2 ≠ 3 ≠ ., k is a population vector arrangement.
If the dth dimensions of vectors 1 and 2 are equal, the dth feature of the difference vector (Equation 3) is zero; otherwise, it has the same value as vector 1 .
Afterward, the mutation and crossover operations are executed, as shown in Equations 4 and 5.
where W represents the try vector, CR (0, 1) represents the crossover amount, and (0, 1) represents a random number. In the selection procedure, if the fitness value of the try vector is greater than that of the current vector , it will be replaced. Otherwise, the current vector is stored for the next generation.

Description of data
Canayaz developed a COVID-19 X-ray data set that included three subgroups of patients, including those with COVID-19, those with pneumonia, and those who were healthy [15]. By combining data from this database, a total of 364 images for each of the three categories were obtained as a balanced dataset [21][22][23]. The total number of images is equal to the number of classes multiplied by the number of class instances = (3*394)1092, with a 224 x 224 dimension. The same data is used in this study to predict COVID-19 disease using a convolutional neural network and to select optimal features using the binary differential metaheuristic algorithm. Figure 2 illustrates a representative sample of three output classifications: COVID-19, pneumonia, and healthy.
Geometric mean= √Sensitivity * Specificity 2 (9)  Convolutional networks are used to transform data into feature vectors. Given that some network features may degrade the model's performance [1], after extracting 400 features in the first fully connected layer, the binary differential metaheuristic algorithm was used to select the optimal feature subset and eliminate unnecessary features. The binary differential algorithm's parameters were population = 20, iteration = 100 (figure 5), and a crossover rate of 1. The amount (1-(geometric mean)) of the SVM classifier [26] was regarded as the population's fitness values ( Figure 5). Following the binary differential algorithm's execution, 340 optimal features were selected.

Performance comparison
The conventional validation (CV) method, which employs random sampling, is one of the training and testing protocols used to determine the model's accuracy and validate the estimation results. According to the CV method, 70% of data was used for training, 15% for validation, and 15% for testing [27,28]. The proposed method was applied to the data, and 100 runs were performed to prevent overfitting [29]. The optimally selected features from the differential algorithm and the initial extracted features from the deep convolutional neural network were entered into the SVM classifier.
The confusion matrix for the SVM classifier's original and optimized features is demonstrated in Table 1 using training, testing, validation, and total data. Transfer learning models are trained to classify 1,000 different types of object images and must be retrained to classify specific issues such as COVID-19 detection. Although the learning process is prompt in models like ResNet and SqueezeNet, they require preprocessing the input image, sizing the data set, and setting multiple parameters. The upper layers extract color and edge features, while the deeper layers extract complex features.
Process time increases as the number of layers in transfer learning models increases. The trained transfer learning model's feature map and activation layers must be customized for the specific COVID-19 problem, which requires a large amount of memory. After fine-tuning the pre-trained model's principal component analysis (PCA), the optimal feature can be selected using heuristic methods, automated encoders, or variance-based selectors. Finally, ensemble methods, such as a combination of SVMs or other classifiers, can be used to predict COVID-19 disease diagnosis accuracy. Using semi-supervised self-learning methods may result in acceptable accuracy and reduced labeling time.
In future work, a different feature selection algorithm and the application of additional learners may produce improved results. Along with the images, the parameters derived from clinical trials can create a new model with a novel combination of features for diagnosing the disease and possibly predicting mortality as a result.

CONCLUSION
The number of people infected with COVID-19 has risen rapidly. Machine vision techniques and artificial intelligence are critical in diagnosing and treating disease. The purpose of this paper was to propose a method for the "COVID-19" problem via a set of lung images that included three categories of pneumonia, COVID-19, and healthy.
A deep convolutional neural network consisting of 11 layers was applied to extract the features. The binary differential metaheuristic method was used to select relevant features and eliminate unrelated features. Lung Xray images were classified using an SVM classifier based on these optimal features. This study demonstrated that the accuracy indicator and the number of relevant extracted features outperformed previous methods using the same data. Based on a deep neural network and a metaheuristic feature selection algorithm, the proposed model can be used in various other medical applications.