Deep SVDD and Transfer Learning for COVID-19 Diagnosis Using CT Images

The novel coronavirus disease (COVID-19), which appeared in Wuhan, China, is spreading rapidly worldwide. Health systems in many countries have collapsed as a result of this pandemic, and hundreds of thousands of people have died due to acute respiratory distress syndrome caused by this virus. As a result, diagnosing COVID-19 in the early stages of infection is critical in the fight against the disease because it saves the patient's life and prevents the disease from spreading. In this study, we proposed a novel approach based on transfer learning and deep support vector data description (DSVDD) to distinguish among COVID-19, non-COVID-19 pneumonia, and intact CT images. Our approach consists of three models, each of which can classify one specific category as normal and the other as anomalous. To our knowledge, this is the first study to use the one-class DSVDD and transfer learning to diagnose lung disease. For the proposed approach, we used two scenarios: one with pretrained VGG16 and one with ResNet50. The proposed models were trained using data gathered with the assistance of an expert radiologist from three internet-accessible sources in end-to-end fusion using three split data ratios. Based on training with 70%, 50%, and 30% of the data, the proposed VGG16 models achieved (0.8281, 0.9170, and 0.9294) for the F1 score, while the proposed ResNet50 models achieved (0.9109, 0.9188, and 0.9333).


Introduction
COVID-19 is the common name of the disease caused by the novel coronavirus (2019-nCoV). Te infected patients will mostly have mild to severe respiratory illnesses. Although most patients can recover without the need to take certain medications, some will get very ill and require medical treatment. People with chronic diseases such as cardiovascular disease, diabetes, chronic respiratory disease, or cancer, in addition to those over 65 years of age, are more likely to develop serious illnesses. People of diferent ages can become very ill or die due to the infection [1]. COVID-19 has impacted every aspect of our everyday lives since it spreads fast and causes acute respiratory distress syndrome [2]. Infected people in the initial stage have numerous typical symptoms, including headaches, fevers, coughs, and severe pain in most muscles of the body [3,4]. Many patients perished as a result verify its efcacy as a substitute for identifying infected individuals [10]. Tese studies reveal that most patients' CT images have consolidation spots and ground-glass opacities (GGO), indicating that CT scans may be used to verify the presence of COVID-19 disease and monitor therapeutic progression [10,11]. Te coronavirus spreads quickly between people or when in contact with an area where an infected person resides. To avoid erroneous diagnoses, a team of radiologists must review the CT [12].
As a result of the rapid development of vision techniques, numerous vision models have been used to detect anomalies in medical images and identify tumors. Researchers have proposed several techniques for segmenting lesions in the skin, brain, and other tissues. Tese systems utilized a variety of medical images, including CT scans, ultrasounds, X-rays, and magnetic resonance imaging (MIR) [13,14]. Tese techniques provide amazing results, sometimes superior to those of radiologists [15].
Several studies have used CT images and deep learning (DL) to identify COVID-19 and discover lung abnormalities [16,17]. DL techniques can extract the appropriate features, which, in turn, increases the efectiveness of the models they use. Tis is why almost all of the proposed approaches use them to detect coronavirus in CT scans. On the other hand, DL techniques require large amounts of data to train deep models. Terefore, the lack of data makes it difcult to train a deep model to detect COVID-19 using CT scans [18].
Despite the encouraging results of these studies in identifying COVID-19, additional enhancements are necessary to address some limitations. Te lack of balanced and sufcient data, intraclass variation, and interclass similarity represent the main obstacles to building a robust system. In addition, since Corona is a new virus, the researchers gathered CT images from several sources, so most of them are blurry, and there is also a high variation between them. As a result, if a model were developed to categorize CT scans as COVID-19 or healthy, all lung abnormalities would be labeled as COVID-19, which is not medically accurate.
To address the aforementioned challenges, a novel approach using deep SVDD and transfer learning is proposed to identify COVID-19. Two diferent pretrained models (VGG16 and ResNet50) are used separately with deep SVDD. Our approach is composed of three models. Each model can recognize a particular class. In general, and specifcally for COVID-19 diagnostics, we believe that we are the frst to use pretrained models with DSVDD. Te following is a summary of the research's contributions: (i) With the help of a radiology specialist, we constructed a new dataset including three classes (intact, COVID-19, pneumonia) of CT images from three available resources on the Internet.
(ii) We proposed a new approach based on the oneclass DSVDD consisting of three models. Each model is trained to fnd a closed hypersphere combining only one class of samples. As a result, when a new virus is discovered, we simply train a new model on the newly discovered virus data without afecting the other models.
(iii) Te basic deep network in the original DSVDD was replaced with pretrained VGG16 and ResNet50 networks to address the data shortage, and then we trained the models using each network separately.
Te remainder of this work is structured as follows: Section 2 represents the related work. A short overview of all the concepts and techniques utilized in our research is provided in Preliminaries Section 3. In Section 4, we describe the dataset used in this research. In Section 5, the proposed approach is explained in detail. Section 6 outlines the experiments and results in detail. We conclude this work in Section 7.

Related Work
Numerous studies have proposed DL techniques to identify COVID-19. Tese studies addressed the data defciency by enhancing data, transferring learning, or designing a new neural network structure. A three-branched deep neural network (DNN) based on the ResNet50 structure was proposed by Wu et al. [19]. Te network receives three diferent views of the CT images, one for each branch. Te extracted features for each branch are fused and passed to fully connected layers to determine whether they are COVID-19 or not. Using ten pretrained models; Ardakani et al. [20] identifed COVID-19 pneumonia and non-COVID pneumonia in the cropped lesion regions of CT images. A radiologist must crop the infected areas in the proposed framework. Amyar et al. [21] suggested a threetask DNN. Tis network starts with an encoder, which receives CT images and generates output that is used in three branches to classify, segment, and reconstruct the CT image. Te U-Net structure is employed to segment and reconstruct the CT image. Gozes et al. [22] used the U-Net to clip the lung parenchyma. Te region of interest was then classifed as intact or COVID-19-infected using pretrained ResNet50. Wu et al. [23] designed a DNN to identify COVID-19 and determine the infection area using CT scans. Te Res2Net is used as the base classifcation model after replacing the last dense layer with a two-channel layer representing the probability of COVID-19 and intact patients. Te segmentation model is made up of encoders and decoders based on VGG16 and the U-Net style. Yousefzadeh et al. [24] used the pretrained EfcientNetB3 model to build a system to classify all slices of a CT fle into normal, COVID-19, and non-COVID. Tis system used the features extracted from the trained model to train fully connected layers. Wang et al. [25] propose a two-stage framework. First, the lung lobes were extracted using 3D-UNET. In the second stage, two subnetworks are used to categorize the extracted lobes. While one subnetwork is trained to identify pneumonia, the other predicts the kind of pneumonia. Hu et al. [26] proposed a deep model to identify COVID-19 using CT scans. A trained 3D U-Net is used to segment the parenchyma of a normalized CT scan. Tis model is inspired by the VGG architecture, where all convolution blocks use a kernel of size 3 × 3. Because the COVID-19 lesion regions change depending on severity, multi-scale learning is utilized. Te multiscale features from multiple levels are combined to represent the network's fnal output. Li [27] proposed a framework for distinguishing COVID-19, communityacquired pneumonia, and nonpneumonia. It is a deep 3D model that extracts 2D and 3D features based on ResNet50. Tis model receives sequences of CT slices to generate the feature maps. Tese feature maps are combined using a max pooling layer and then passed to a fully connected layer, which uses the SoftMax activation function to generate probability scores for each class. Using CT scans, Wang et al. [28] presented a method for diagnosing COVID-19 using CT scans. First, manually obtain the ROI (region of interest). Te ROIs are sent to the preprocessing stage, where images are converted to grayscale and the smallest rectangle enclosing the parenchyma is extracted using thresholding and food fll techniques. Te method used the pretrained Inception model for the classifcation task. Using transfer learning and trained models GoogleNet, ResNet, and AlexNet, Zhou et al. [29] built the ensemble model (EDL_COVID). Tese models are initialized with the pretrained weights to act as feature extractors. Tis model used the relative majority vote algorithm and Softmax for the fnal outputs. Fan et al. [30] proposed a model with two branches: one that extracts local features using VGG19 and the other extracts global features using a transformer. Te extracted local and global features are fused using a bidirectional fusion model that connects the two branches. Te proposed model classifes the input images as intact, COVID-19, and non-COVID-19 pneumonia. We summarized the related work in Table 1.

Preliminaries
Tis section briefy describes the techniques used and illustrates some notations. [31]. SVDD is an OC-SVM-related technique that is used to solve one-class classifcation problems. In the one-class task, typically only positive class data are available, whereas the negative class data are either unavailable or insufcient to train a binary classifer. Te idea behind SVDD is to separate positive data by a hypersphere characterized by a radius R > 0 and a center C ∈ F. Most positive samples in the feature space F are grouped by this hypersphere. Let X ⊆ R d is the space of the inputs, φ κ (x): X ⟶ F κ is a transfer function with kernel Κ: X × X ⟶ [0, ∞), F κ is the space of the dot product.

Support Vector Data Description (SVDD)
Given positive training samples D � x 1 , x 2 , . . . . . . . . . . . . . . . ., x n }, x i ∈ X, , i � 1, 2, . . . . . . . . . ..n, n refers to the sample count. Following is a formalization of the SVDD problem: To allow the soft margin, slake variable ξ is used, υ ∈ (0, 1] controls the trade-of between penalties ξ i and the volume of the sphere. When a point x meets the following criteria, it represents an outlier point: 3.2. Deep Learning (DL). DL is a branch of machine learning that uses techniques to mimic the activity of the human brain neocortex. Te neocortex is made up of layers of neurons that make up 80% of the wrinkly brain and is where most thinking occurs [32]. Tese layers are learned in a hierarchical order in which degrees of abstract representation are learned to interpret data patterns. Low and many levels of data abstraction represent high-level abstractions. Te ability of deep learning to automatically learn how to extract features at various abstract levels is one of its most potent capabilities. Tis property allows us to develop systems that are able to locate complex transfer functions without the requirement for handcrafted features [32]. Building DNNs involves stacking layers vertically and learning from a big dataset. During the learning of a DNN, the weights and biases are updated via the backpropagation algorithm to minimize the prediction error. DNN layers learn feature extraction in a hierarchical manner, where the lower layer learns low-level features such as edges, while the higher layer learns high-level features such as eyes, noses, and faces [33].

Deep Support Vector Data Description (DSVDD).
Most researchers are drawn to DL because of its recent success in almost all areas of computer science. Ruf [31] proposed the DSVDD model to detect abnormal images. DSVDD is a hybrid of the deep convolutional neural network (DCNN) and SVDD. Te DSVDD looks for the minimum-volume hypersphere that encloses the transferred data generated by the DCNN. Te authors proposed two models: soft-margin and one-class models. Te one-class model reduces the hypersphere's volume by minimizing the mean squared distance between the hypersphere's center and all training samples. It is intended for this hypersphere to include only normal images, as shown in Figure 1, so images outside of it are categorized as abnormal.

Transfer
Learning. Big datasets are often required for DNN training, but building large datasets can be timeconsuming and occasionally impossible because of a data shortage. Transfer learning (TL) is a strategy for transferring knowledge obtained by DNNs such as DCNNs to tackle another related challenge. Since COVID-19 is a new disease, the lack of adequate data is the biggest obstacle. Terefore, TL and DCNN are considered signifcant techniques for building intelligent systems that aid in diagnosing the disease. Te TL process starts by initializing a pretrained DCNN model with weights obtained previously by training on a large dataset. Tese initial weights allow the DCNN  [34]. Te frst employs the pretrained model as a feature extractor. Te model architecture is unchanged, except for the classifcation components, which are replaced by a diferent classifcation structure. In this strategy, only the classifcation parts are trained using the current task's dataset, whereas the feature extractor part is still the same [35]. Te second strategy involves adapting the pretrained DCNN architecture to our current problem [34].

Dataset
Building a deep-learning system to identify COVID-19 in the absence of sufcient data is a difcult task. As a result, most of the studies used clinical data to train their deep models. We constructed a CT dataset by gathering data from three resources on the Internet. Tis dataset contains images categorized into COVID-19, non-COVID-19 pneumonia, and intact. Te CT scan fles are made up of a series of slices. If the fles belong to an infected patient, healthy slices must be removed, and nonparenchymal slices must be removed from all CT fles [36]. As a result, an experienced radiologist reviewed these datasets to eliminate unwanted slices. Te COVID-CT dataset [37] has two sets of slices. Te frst set contains 349 slices from COVID-19 patients, while the second set contains 397 slices from non-COVID-19 lung infections. To supplement the data, two additional datasets are used. COVID_DATASET [38] has three sets: COVID-19 (719 slices), intact (2495 slices), and pneumonia (1825 slices). Finally, the SARS-CoV-2 CT dataset [39] is split into two categories: 1252 COVID-19 CT slices and 1230 non-COVID-19 CTslices. We removed the non-COVID CTfrom the previous two datasets because they could contain a wide range of lung disorders, and we wanted to create a dataset with only three categories. Figure 2 shows some examples from our dataset. Furthermore, we also used COVID-19 [40], which consists of 20 COVID-19 CT scan fles (NII.gz fles) with ground truth fles. Te ground truth fles are divided into three groups of masks: one for the entire parenchyma, including the lesion regions; another for the lesion areas only; and a third for the parenchyma. We used this dataset to train U-Net [41] for parenchyma extraction. Details about these datasets are shown in Table 2.

Proposed Approach
Te proposed approach, including the preprocessing stage and deep classifcation models, is described in detail in this section. Figure 3 depicts the proposed approach.
Te diagnostic process in the proposed approach starts with the CT being classifed as healthy or infected using the intact model; if the CT is infected, it is passed to the COVID-19 model for classifcation as COVID-19, or it is passed to the pneumonia model for classifcation as pneumonia or another type of infection. To develop our approach, we proposed two methods: one that used pretrained VGG16 with DSVDD and the other that used pretrained ResNet50 with DSVDD, where each method was built separately. For all models (intact, COVID-19, and pneumonia), the frst method uses the VGG16 with DSVDD as the classifer, whereas the second uses ResNet50 with DSVDD. All stages and models of the proposed approach will be explained in detail in the following sections:

Preprocessing Stage.
Preprocessing is a crucial step in the development of DL systems. Te constructed dataset's slices came from a variety of sources. Terefore, the proposed method may not work well on most of these slices due to various edges close to the lung tissue. We trained the U-Net model to extract the parenchymal mask. Morphological operations are carried out on the mask to remove any possible small areas close to the lung parenchyma. To obtain the parenchyma, we utilized the extracted mask. Finally, we scaled all slices to 256 × 256 pixels, and the color was adjusted to fall within the range [0, 1].

Proposed Method.
We proposed two methods for identifying COVID-19 based on DSVDD [31], Te frst method combines DSVDD with the pretrained ResNet50 [42], and the second with the pretrained visual geometry group network VGG16 [43]. We will go over the proposed methods in detail in this section: Computational Intelligence and Neuroscience

Te Modifed Pretrained VGG16. Te original VGG16
is built with fxed-size 3 × 3 convolution flters and trained on millions of images with 1000 classes (ImageNet). It has a simple architecture but is very efective at the same time. Te small 3 × 3 feld flter can learn discriminative features of biomedical images at a fne-grained level and can gain information in a small area around the center [43,44]. VGG16 uses thirteen convolutional layers to extract features and three dense layers for classifcation [43,45]. Te name beyond "VGG16" is that it has sixteen layers with learnable parameters [45]. VGG16 can extract a wide range of features due to the variety of classes it was trained on. As a result, it is applicable to a wide range of classifcation tasks.
In our classifcation task, we used the pretrained VGG16 after replacing the last three dense layers with a global average pooling layer and a dense layer with only 128 neurons, as illustrated in Figure 4.

Te Modifed Pretrained
ResNet50. Te design of very deep networks causes a problem of gradient vanishing and degradation. Tis issue is resolved by residual learning using the ResNet developed by He et al. [42]. In the proposed method, we used the pretrained ResNet50 after replacing the last dense layers with a global average pooling layer and a dense layer with only 128 neurons. ResNet50 is made up of multiple stacked residual blocks, conventional convolution, and pooling layers. Each residual block combines three convolution layers of 1 × 1, 3 × 3, and 1 × 1 kernels, respectively, with a skip connection that adds the inputs of the frst layer to the outputs of the last layer in the block, as  Computational Intelligence and Neuroscience depicted in Figure 5. Each convolution layer in the residual block is followed by batch normalization (BN) and the rectifed linear activation function (ReLU), respectively [46].

Deep Support Vector Data Description (DSVDD) with Pretrained Models.
Tere are abnormal areas when comparing CT slices of COVID-19 patients with those of healthy individuals. We can consider this abnormal area an anomaly compared to healthy areas. Te severity of the infection enlarges the anomalous regions [10]. In this study, DSVDD with pretrained models (VGG16 and ResNet50) is proposed to detect anomalous CT. Te proposed DSVDD addresses data shortages while also improving COVID-19 diagnosis. We proposed two methods. DSVDD-VGG, which combines the pretrained VGG16 with DSVDD, and DSVDD-ResNet, which combines the pretrained ResNet50 with DSVDD. Te pretrained models serve as feature extractors, while DSVDD seeks the minimumvolume hypersphere containing the extracted features of the normal slices. { }, L represents the number of layers. Te proposed DSVDD aims to initialize VGG16 and ResNet50 with the trained weights and retrain them on our dataset to locate the minimum-volume hypersphere surrounding the samples that belong to the normal class. We formulate below the objectives of DSVDD-VGG and DSVDD-ResNet: Te length between the extracted features φ(x i ; W) and c ∈ F is penalized using the quadratic loss function. Te second term is a network weight decay regularizer with    Computational Intelligence and Neuroscience a hyperparameter t > 0. In order to reduce the hypersphere's volume, DSVDD minimizes the mean distance between the hypersphere's center and the extracted features. A given x ∈ Χ is considered an anomalous slice if its length exceeds the last radius R * from the hypersphere's center c s.t where W * is the fnal trained network's weights x is anomoly, score(x) > R * .

Implementation Outline.
Because our dataset has three categories, we trained three models of DSVDD-VGG and three models of DSVDD-ResNet independently. During each model's training, one category was labeled "normal" and the other "anomalous." Due to restricted resources, we use a batch size of four images to train our models. We developed our approach using the Keras and TensorFlow libraries. We trained each model for 50 epochs using the NVIDIA Tesla K80 GPU. Te Adam optimizer was used, with an initial learning rate of (10 −5 ) and β1, β2 were set to 0.9 and 0.999, respectively. In terms of training the U-Net model, we used 90% of the data for 50 epochs to train the standard architecture U-Net model. Te remaining data was used to evaluate the model during training to select the best-trained model. We used the Adam optimizer with a batch size of 10 and an initial learning rate of (10 −4 ). Augmentation with (rotation � 0.2, width and height shift � 0.05, shear � 0.05, zoom � 0.05) was used to increase the training data.

Training and Testing
Strategies. We trained one model of DSVDD-VGG and DSVDD-ResNet independently for each class in our dataset. We trained each model to classify one category as normal and the other as anomalous. To further investigate the efcacy of our proposed models, three different split ratios were used to train all models: 70%, 50%, and 30% of all categories were used for training, 10% of all categories were used for validation, and the remaining samples were used for testing.
Te intact model considers the healthy CT slice as normal, while the pneumonia and COVID-19 slices are anomalous. Similarly, the pneumonia and COVID-19 models consider their slices as normal while the others are anomalous. We used only positive samples for training because we used a one-class classifer, but we used both positive and negative samples for validation and testing. Let's take a look at the intact model training process with 70% of the data as an example.
We trained the model with intact samples and validated it after every epoch with samples from healthy, COVID-19, and pneumonia slices. Te trained model was tested on data from all classes that had not been included in training and validation. Te checkpoint was saved when the current validation's evaluation metrics values exceeded the previous highest validation value.

Evaluation
Metrics. Specifcity, sensitivity, area under the receiver operator curve (AUC), and the F1 score were utilized to assess the proposed models. Specifcity indicates how well the model distinguishes negative cases, whereas sensitivity indicates how well the model distinguishes positive cases. Because we were dealing with highly unbalanced data for binary classifcation tasks, the AUC metric and F1 score were utilized to assess each model's overall performance. Te following are the formulas for the metrics used: , ,  Figure 6. Figure 7 depicts the receiving operating characteristic (ROC) curves for the intact models.
Te ROCs indicate that the models are stable and perform well. According to the confusion matrices, DSVDD-VGG trained on 70% of the intact data, correctly classifed all intact slices except fve, and had the highest sensitivity metric value of any intact model. Te specifcity values indicate that the DSVDD-ResNet, which was trained on 70% of the intact data, produced better results regarding separating anomalous CT. In terms of overall evaluation, F1 score values show that DSVDD-ResNet has fewer false positives than DSVDD-VGG, but DSVDD-VGG has a higher AUC value, indicating that it is more sensitive to intact slices. Figure 8 exhibits examples of CT slices that were misclassifed using intact models. Table 4 summarizes the values of the COVID-19 model metrics for validation and testing according to the diferent training ratios. Te confusion matrices in Figure 9 indicate that the models efciently discriminate COVID-19 slices from other categories, with specifcity and sensitivity values close to 99%. Figure 10 depicts the ROC curves of the COVID models. Figure 11 exhibits examples of misclassifed slices. Figure 11(a) depicts samples of COVID slices that are classifed as non-COVID. In Figure 11(b), non-COVID slices are classifed as COVID-19 slices. Table 5 summarizes the pneumonia models' validation and testing evaluation metrics: specifcity, sensitivity, F1 score, and AUC. As shown in the confusion matrices in Figure 12, DSVDD-ResNet outperformed DSVDD-VGG, such that DSVDD-ResNet was very sensitive to pneumonia, with only eight slices misclassifed using the model trained on 70% of the data and only 20 slices misclassifed using the model trained on 30% of the data. Furthermore, DSVDD-VGG correctly classifed the majority of pneumonia CTs, with only (8, 21, and 28) misclassifed among the three DSVDD-VGG models. Figure 13 depicts the pneumonia models' ROC curves, which show the models' stability and ability to achieve high performance in the absence of a large dataset. Figure 14 depicts examples of misclassifed slices.

Pneumonia Models.
We provided a quantitative comparison of the average metrics values for DSVDD-VGG and DSVDD-ResNet models based on all the training scenarios with the stateof-the-art methods in Table 6.
As shown in Table 6, DSVDD-ResNet50 outperformed the majority of the state-of-the-art methods used in the comparison and achieved the highest metric values among the majority of metrics used. We excluded from the comparison studies that used only two classes or that used tumor data, where distinguishing between tumor and COVID-19 is simple due to the distinct signs of tumor.

Statistical Analysis Using McNemar's Test.
We devoted this section to comparing the performance of the proposed models using McNemar's test. McNemar is an Statistical hypothesis test that can be used to evaluate two binary classifers. It performs the test using a 2 × 2 contingency table [47]. We have two hypotheses we are going to investigate using this test: H 0 � Te Null Hypothesis: both classifers have the same error rate.
H 1 � Te alternative hypothesis, a signifcant diference exists and both have a diferent error rate.
Such that: n 10 � Te number of correctly classifed samples by ResNet50 but misclassifed by VGG16.
n 01 � Te number of correctly classifed samples by VGG16 but misclassifed by ResNet50 Tables 7-9 contain the 2 × 2 contingency tables for all models' comparisons and the results of performing the McNemar test with signifcance threshold α � 0.05.

Discussion.
Although COVID-19 is a new strain of coronavirus and there is not enough data, many studies have proposed many deep models to diagnose the disease and aid in the fght against it. In our turn, we proposed a new approach to discriminate between COVID-19, non-COVID pneumonia, and intact lung CTs. Te proposed framework is diferent from the other studies in that a one-class deep model is devoted to diagnosing each class in our dataset, which overcome the imbalance-class problem.
Although the results of the proposed framework were promising, we noticed some important issues that needed to be discussed. In the COVID-19 models, the trained models using 30% of the data produced the same AUC values as the models trained on 70% of the data and were better than the models trained on 50%. Tis is because the CT scan fles are composed of a set of slices, and consecutive slices are very similar (see Figure 15). To overcome this problem, we should use nonconsecutive slices or use slices from diferent CT fles for diferent patients during the construction of the dataset, but in our situation, we did not have enough data.
In addition, the intact model's performance was lower than the COVID-19 and pneumonia models due to some area around the parenchyma are similar to the GGO, which is the main indicator of pneumonia (see Figure 16).
Moreover, the proposed models sometimes failed to identify the infected people when the lesion was at the edge of the lung's parenchyma, especially in the slices at the beginning or end of the CT fles (see Figure 17). Terefore, we suggest considering all the slices of each patient as one instance in order to solve this issue.    Computational Intelligence and Neuroscience

Conclusion
Tis study proposed a new approach for distinguishing COVID-19 from other infections in an efort to halt the disease's spread. Several models are included in the proposed approach, including one that classifes intact lung images as normal and the others as anomalous; another that identifes the COVID-19-infected lung and the other images as anomalous; and fnally, a model that identifes the lung with non-COIVD-19 pneumonia and the other images as anomalous.
Our approach models were built using DSVDD and the pretrained models VGG16 and ResNet50. Te VGG16 and ResNet50 models are used separately to be the deep network of DSVDD, such that for every category in our dataset, two models are built: one combining DSVDD with ResNet50 and the other combining VGG16 with DSVDD. We trained all models in an end-to-end fashion and used diferent split ratios of data for more verifcation of the efciency of our approach. Experiment results revealed that the majority of proposed models outperformed state-of-the-art models. In future work, we will try to combine the patient's symptoms, if they are available, with the extracted features to improve the outcome. Furthermore, we will employ semantic segmentation to detect lesion areas and assess the severity of the infection.

Data Availability
Te data used in this study are provided in the reference.

Conflicts of Interest
Te authors declare that they have no conficts of interest.