A Comprehensive Survey on COVID-19 Detection and Classification Using Chest-X-Ray Images

This section gives the detailed analysis of various methods for identifying the COVID-19 by using machine learning, deep learning and transfer learning models. This survey is concentrated on purely on artificial intelligence approaches for feature extraction, segmentation and classification operations. Further, this survey also focused on feature selection operations using evolutionary, swarm intelligence, natural selection, meta-heuristics- based bio-optimization methods.

3.1 Survey based on COVID-19 dataset

This section gives the detailed survey on COVID-19 detection and classification methods using two different datasets such CT and CXR. The challenges presented in artificial intelligence approaches with respect to these datasets are studied.

3.1.1 Survey on COVID-19 classification using CT images

A chest CT scan is a diagnostic procedure that does not involve the use of any invasive procedures in order to obtain a clear image of the chest region of a patient. It generates x-rays in a more ideal shape, which results in photographs of the chest that are more specific. It creates images that contain bones, fats, muscles, and organs, giving doctors a better look, which is vital when attempting to make accurate diagnosis [43]. There are several variations of the chest CT scan, the most common of which are the excessive-decision chest CT test and the spiral chest CT test. In a single revolution of the x-ray tube, the high-definition chest CT scan can provide more information than just a single slice, often known as a snapshot. The spiral chest CT test usefulness consists of a table that moves in a continuous motion along a passage that resembles a tunnel while an x-ray tube travels in a spiral pattern. The fact that a three-dimensional image of the lungs may be produced using the spiral CT is one of the test's primary benefits.

GGO, consolidation, reticulation/thickened interlobular septa, nodules, and lesion distribution are some important CT functions (left, proper or bilateral lungs). The most noticeable CT functions [44] seen in COVID-19 pneumonia are areas of floor-glass opacification that are bilateral and subpleural, as well as consolidation that affects the lower lobes. During the intermediate period, which lasts for four to fourteen days after the onset of symptoms, the crazy-paving pattern and a potentially visible Halo signal become important functions. The lung CT results of patients diagnosed with COVID-19 are depicted in Figure 5. Figure 5(a) is an axial CT photograph taken within the first week of treatment of a patient who is 37 years old. It demonstrates peripheral bilateral GGO as well as a few small-scattered pulmonary nodules (arrows). Figure 5(b) is an axial CT photograph taken of a patient who is 52 years old during the second week after their initial presentation. This scan demonstrates that the patient has diffuse GGO, consolidation, and bronchiectasis (arrowhead)

5.png

Figure 5. Samples COVID-19 affected CT images adapted from [44]

The first category of labour to be covered in this discussion focuses on diagnosis methods as a segmentation issue. A CT image dataset containing 46,096 photographs of both healthy and inflammatory patients that has been proposed by Chen et al. [45] is comprised of these images. These photographs have been categorised by qualified radiologists. It turned into amassed from 106 individuals hospitalised with 51 confirmed cases of COVID-19 pneumonia and 55 patients serving as controls. Only deep learning models were employed for the segmentation process in this work in order for it to successfully identify the diseased region in CT images of patients who were either healthy or infected.

It came to be mostly dependent on the UNet++ semantic segmentation version [46], which is utilised to extract valid regions contained inside the images. It employed 289 randomly selected CT images and compared them to another 600 CT images that were also randomly selected. The professional model that was developed as a result of this study was implemented at the Renmin Hospital of Wuhan University in Wuhan, Hubei province, China, in order to assist in the assessment of current COVID-19 cases. It was also made available for free download on the internet, making it possible to conduct a speedy investigation of current incidents in several other areas. An artificial intelligence framework that is cloud-based and entirely free to use has been developed in order to provide assistance with the detection of COVID-19 pneumonia around the world. The second type of artwork viewed COVID-19 to be a binary type problem. Li et al. [47] proposed the use of transfer reading on the RESNET50 in order to extract visual functions from volumetric chest CT scans. This was done using the COVNet algorithm. Using the U-Net model, lung segmentation has developed into being a pre-processing task that can be accomplished. The problem of making a diagnosis became additionally approached as a three-class category problem: separating healthy patients from persons with different kinds of pneumonia and people with COVID-19. The research conducted by Song et al. [40] employed data collected from 88 patients diagnosed with COVID-19, 101 patients diagnosed with pneumonia caused by bacteria, and 86 healthy individuals as controls. It developed a neural network known as DRE-Net (Relation Extraction neural network), which changed into basically built on ResNet50. On this network, the Feature Pyramid Network (FPN) [48] and the Attention module have been included so as to symbolise finer-grained components of the images. The various CT image-based COVID-19 detection methods are outlined in Table 1, together with segmentation, class styles, datasets, and performance evaluations. Sensitivity (SEN), specificity (SPE), accuracy (ACC), and area beneath curve are the overall performance parameters that are provided here (AUC).

3.1.2 Survey on COVID-19 classification using CXR images

The identification of the GGO area, which can be difficult with CT pictures since it is concentrated in a small space, is one of the challenges presented in the COVID-19 detection process. In addition, the traditional procedures did not slice and extract the features from CT photographs any more. In addition, there is a possibility that there is not availability of more than one well-known dataset for CT snapshots. The use of CT imaging comes with a number of drawbacks, one of which is the requirement for an excessively high impacted individual dose and superior value. When it comes to detecting chest pathology, digital CXR radiography is the imaging modality that has a cheaper cost and a wider availability. Because of this, automatic diagnosis of COVID-19 skills in CXR will make it a particularly valuable diagnostic tool [57] in the fight against the disease. Digital X-ray imaging computer-aided assessment is used for the diagnosis of first-rate disorders such as osteoporosis, the majority of malignancies, and cardiac illness.

Assessment enhancement, on the other hand, is utilised as a pre-processing step because it is extremely difficult to differentiate soft tissue with a terrible assessment in X-ray imagery. Lung segmentation of CXRs is a step that is necessary and vital for you to take in order to detect lung nodules, and the literature suggests a number of different segmentation procedures [58]. Consolidation has been identified in COVID-19 infected patients based on the results of CXR exams. In one study that was conducted in Hong Kong [59], three different patients received daily CXR examinations, and all of them demonstrated progress in the lung consolidation over the course of three to four days. Improvement can be shown on subsequent CXR exams as time passes after the first diagnosis. The patient with the 0.33 showed no significant changes during the course of 8 days. However, an identical test showed that the floor glass opacities within the right decrease lobe peripheral on the CT are not visible at the chest radiograph, which were taken one hour after the primary observe. This chest radiograph became taken after the first observe. Despite this, CXR is still recommended in addition to CT for more in-depth radiological investigation. Several different automatic strategies relating to CXR were been forward. In the following section, we will go over the most important work, and Table 2 will provide a more organised explanation of the many methods used. Both of these sections will take place simultaneously.

6.png

Figure 6. Samples COVID-19 affected CXR images adapted from the research [59]

Figure 6 depicts chest radiographs taken of an elderly male patient affected by the disease who was originally from Wuhan, China and later travelled to Hong Kong, China. These three chest radiographs were determined based on the daily chest radiographs that were obtained on the patient who is being discussed here. On day four, the consolidation that began on day zero in the proper lower region continued, and fresh consolidative modifications were made in the proper midzone outer border and perihilar place. This exchange in the midzone makes an improvement on the day 7 movie, which includes high GGO territory.

To this day, quite a few computer vision models for X-ray COVID-19 that are mostly based on deep learning have been proposed. The model COVID-Net [60], which was proposed using the resource of Darwin AI in Canada, is one of the most significant advances that has been made. This work provides a network structure for the detection of COVID-19 cases based on chest X-rays by combining human-driven principled network layout prototype with system-pushed design exploration.

The COVIDX-Net was proposed by Hemdan et al. [27] based fully largely on seven particular architectures of DCNNs; particularly VGG19, DenseNet201, InceptionV3, ResNetV2, InceptionResNetV2, Xception, etc. Brilliant version mixing led to F1- values of 0.89 and zero respectively. 91 for both typical and COVID-19 occurrences. In a similar fashion, Asmaa Abbas and colleagues [61] proposed a Decompose, Transfer, and Compose (DeTraC) approach for the category of COVID-19 chest X-ray photos. In order to carry out the diagnosis, the authors utilised CNN capabilities based on pre-professional models that were run on ImageNet and ResNet. Ghoshal et al. [32] introduced the Uncertainty-Aware COVID-19 Classification and Referral version with the recommended Drop weights in keeping with Bayesian Convolutional Neural Networks. This model became used to classify and refer to COVID-19 records (BCNN). In order to make the COVID-19 detection as large as possible, several kinds of prediction uncertainty in deep mastering were used into succeeding paintings. One of its far epistemic or model uncertainty bills for the model parameters uncertainty because it no longer takes into account all of the factors of the data or the lack of schooling data one of its far epistemic or model uncertainty bills for the model parameters in Table 2 outlines a few more CXR picture enabled COVID-19 detection and class methods, including segmentation, feature extraction, and sophistication styles. These methods are presented in chronological order. Table 2 presents a comparison of the general performance of a variety of techniques with regard to a few different datasets.

3.2 Survey on CXR segmentation

This phase provides a targeted evaluation of several CXR picture segmentation approaches for localising the COVID-19 area, which enables the COVID-19 to be classified into the appropriate category. U-Net-primarily based CNN model for lung picture segmentation was carried out, and the type was completed with the aid of the worthwhile resource of method by making use of three different CNN architectures (ResNet, Inception, and VGG). However, this approach has difficulties dealing with the computational complexity. In addition, DEFU-Net has been established for the purpose of CXR photo segmentation. DEFU-Net is a modified version of U-Net that was developed through the use of the utilisation of the utilisation of the utilisation of twin encoder fusion (DEF)-based U-Net. In this instance, DEF is given an advantageous beneficial resource by virtue of the blending of Inception-CNN and Densely-CNN. The network is made deeper with the purpose of making it easier to retrieve contextual characteristics thanks to the densely connected recurrent pathway. Locating the inception blocks with dilatation allows for an increase in the diameter of the network as well as an improvement in the depiction of abilities. The inception blocks have the ability to gather information about the surrounding space both globally and locally from many receptive fields. At the same time, the two pathways are combined through the use of summing capabilities, which helps to maintain the contextual and geographic statistics necessary for element interpretation. This multi-learning-scale version is currently profiting from the excellent developers of the Chest X-ray dataset. However, this method struggles with having a poor level of type accuracy. After that, the FractalCovNet model became an advanced method for segmenting CT and CXR images; nowadays, this is a frequent practise thanks to the utilisation of U-Net and fractal blocks. The same FractalCovNet structure can also be utilised for the enhancement of CT and CXR pictures.

DeepSDM, which utilised rich details from the Signed Distance Map, was introduced by the authors in. (SDM). Through the utilisation of the multi-project methodology, the binary segmentation masks and SDM are determined upon simultaneously. In addition, the SDM regression assignment uses a boundary-primarily based simply definitely weighting technique, which compels the model to focus additional attention on pneumothorax and its contour. In addition, for joint segmentation and class operations, the COVID Smart Data based absolutely truly Network (COVID-SDNet) is proposed. However, using this strategy resulted in a lower F1 score, which in turn produced a reduction in the tool's sensitivity. The authors of proposed a technique known as the conditional generative adverse network (C-GAN) for the purpose of getting lung picture segmentation. After that, binary sturdy invariant scalable key-factors, also known as BRISK, are utilised so that the functions can be extracted from the segmented final effects. In conclusion, a DNN that is completely based on the VGG-19 model is employed for sophistication. Using a channel and spatial hobby module, the authors of suggested a three-terminal interest (TTA) model in order to robotically spotlight the objective area and improve the general overall performance of lung segmentation. At the same time, the 3-terminal hobby mechanism makes use of the advanced semantics of high-scale capabilities in order to enhance the area and recognition capabilities of the eye mechanism, reduce the noise caused by the ancient beyond, and emphasise the capabilities associated with intention. In addition, the self-hobby deep neural community is an advanced version of TTA. It has been developed with the execution of the proposed interest modules to U-Net for segmentation of lung regions and finished experiments even as converting the locations of the eye modules within the baseline network. Our Inception Residual Recurrent CNN version was proposed by the authors of for the purpose of COVID-19 identification and accurate segmentation was carried out using it. In addition, hybrid slime mould contained whale optimization set of regulations. Optimization is applied in order to decrease the losses caused by segmentation styles.

3.3 Survey on feature extraction methods

This section gives the detailed analysis of low level, statistical, texture feature extraction methods by using deep learning and machine learning models. Further, this section also deals with the high-level feature extraction methods. In authors used hybrid features by combining CNN features, Scale Invariant Feature Transform (SIFT), and GIST features. Further, Long Short-Term Memory (LSTM) was used to perform the classification operation. This method suffering with the high complexities. In authors used the wiener filtering for preprocessing the CXR image. Then, features are extracted by using local binary patterns (LBP), gray level run length matrix (GLRM), gray level co-occurrence matrix (GLCM) descriptors. Further, classification operation is performed by using ANN, with the salp swarm algorithm (SSA) based optimal feature selection. But this method suffers with the low efficiency. Further, deep learning models such as VGG19, VGG16, ResNet101, ResNet50, ResNet18 models are used for deep feature extraction methods are used. Further, the classification operation is performed by using SVM model with multiple kernels like Gaussian, Cubic, Quadratic and Linear. But these kernels resulted in low classification performance.

In authors used the Deep Neural Network (DNN) based Bi-LSTM network for deep feature extraction and classification. But this method resulted in considerable low performance with high computational complexity. Further, ResNET50v2 model is used to extract the deep features from the CXR images. Then, features are fused by using analysis of variance and features are selected by mutual information feature selection (MIFS) method. Then, KNN and SVM based machine learning models are used to classify the three classes. But, the MIFS selected the low seismic features with reduced probabilities because it extracted the DWT based low features from CXR images.

In the histogram-oriented gradient (HOG), CNN based joint feature extraction. Initially, modified anisotropic diffusion filtering is used to preprocess the CXR images. Then, the HOG based texture features and CNN based high level features are extracted from the dataset. Finally, VGGNet classifier is used, but this method is suffering with the higher training complexities.

In attribute filters, Gabor filter banks, and morphological operators are utilized for feature extraction. Contextual features reduced by convolutional filters (CFRCF) are used to optimize the deep features, which extracts the hidden relationship between features through nonlinear sub-statistical, global features. Finally, SVM classifier is used for classification. In authors utilized the Bag of Deep Visual Words (BoDVW) for feature extraction process. Further, shallow learner-based CNN is used for classification. Further, U-Net based capsule network is used for deep feature extraction. Then, DenseCapsNet, DenseNet121, and ResNet50 models are used for classification.

3.4 Survey on optimization techniques

This section gives the detailed analysis of various feature selection methods by using evolutionary and natural selection-based feature selection methods. These methods resulted in better selection of features as compared to deep learning features selections. In authors used different feature selection methods such as Advanced Squirrel Search Optimization Algorithm (ASSOA), which selected the CNN extracted features. Then, Multi-layer Perceptron based ResNet-50 is used for classification. But, the ASSOA is high computational complexities. Further, quasi-reflected slime mold algorithm (QRSMA) based meta-heuristic optimization utilized for feature selection, which is and improved version of conventional SMA. In authors developed the modified whale optimization algorithm with population reduction (mWOAPR) approach for efficient segmentation of COVID-19 region, which is an enhanced version of standard WOA.

In used the Nystrom sampling and coalition game theory for feature reduction process and selects the best features, which selected InceptionV3, Xception, and VGG16 extracted features. Further, fractional-order and marine predators’ algorithm (FO-MPA) is utilized for most similar feature selection, which grabbed Inception extracted features. In developed the MobileNetV3 is used for feature extraction, swarm based aquila optimizer is used for feature selection. This method effectively reduces the number of features, but it suffers with the loss of important features during the feature selection.

In authors used the grey wolf optimizer (GWO) for feature selection, and developed the optimized CNN (OptCoNet) for CXR based COVID-19 classification. But this method is capable of extracting only low-level features. In authors implemented a hybrid network, where deep features are extracted by using ResNet50, Inception-v3, and VGG19 models. Then, particle swarm optimization (PSO) with extreme gradient boosting is used to select the optimal features [65]. But this method suffers with the high computational complexity. Further, genetic algorithm (GA) based approach is used for selection of best features. Then, MLP based CNN is used for the classification process. In authors used three optimization methods such as Guided Whale Optimization Algorithm (GWOA), PSO and Stochastic Fractal Search (SFS). Initially, features are extracted by AlexNet models, and best features are selected by using SFS algorithm. Then, SVM model is used for classification process, and the network loss values are optimized using GWOA algorithm. Further, overall performance is improved by using PSO approach.

Authors implemented Fuzzy C-Mean (FCM) clustering for segmentation of COVID-19 region. Further, optimal features are selected by using Information Gain (IG) optimization and classification is performed using back propagation-based ANN. In authors used WOA for optimization of GAN based hyperparameters. In authors used ensemble of ResNet50-Error correcting model for feature extraction and GWO for optimization for optimal feature selection. Further, MLP based CNN is optimized by using WOA and BAT algorithms. This method suffering with the high computational complexity. Table 3 presents summary of few artificial intelligence approaches, with respect to the feature extraction, feature selection and classification models. Further, Table 3 also compares the performance of various approaches on different datasets.