Coronavirus first appeared in Wuhan at the end of 2019, and it is a new strain of the Coronavirus family. It was called severe acute respiratory syndrome before the World Health Organization (WHO) named it as COVID-19 [1, 2]. Initially a public emergency was declared, and the WHO declared a pandemic for the new virus due to the increasing number of deaths in different countries [3]. Covid-19 transmit from one person to another via coughing, sneezing, or talking to others [4]. Once infected by the virus, various symptoms begin to appear, such as high fever, dry coughing, headache, respiratory symptoms similar as the effect of influenza infection. Moreover, in severe cases, difficulty in breathing with organ failure, which may lead to death [5–7]. Whereas in some people, none of these symptoms appear (asymptomatic) and this causes the spread over worldwide. This led to a large number of deaths and the lack of control over the spread of the virus in many countries, the health system has reached a collapse, forcing governments to carry out a complete closure and asking people to commit to stay at home [8].
The critical step in combating the virus is by checking people infected with the virus for isolation and treatment. At present, the foremost approach utilized so far in detecting the covid-19 virus is real-time reverse transcription poly-merase chain reaction test (rRT-PCR) [9, 10].
The abundance of PCR is its accuracy which is around 90%. However, there are limitations of using the Covid-19 test with PCR such as expense, time of duration, and insufficient number of kits [11, 12]. Due to the limitations, scientists proposed alternative methods based on radiographic chest Images (Computed tomography (CT scan) and X-ray), which can distinguish covid-19 infected from uninfected people without error [13, 14]. Since CT imaging modality is easily achievable at the hospitals, and the use of CT images of the chest has many advantages when compared to traditional methods (PCR) [15].
Because of the rapid growth in positive Covid-19 people, many researchers are working to develop several types of artificial intelligence to detect Covid 19 using CT scan images, but these proposals still need to be tested and improved [7, 16]. For this, deep learning technology is one of the most important systems used recently, especially in the medical field such as breast, cardiac, abdominal, pulmonary, pneumonia, and chest radiological images [9, 17–19]. The reason behind this success is that the deep learning technology does not depend on personal or manual use, but it depends on algorithms that can be trained by using labelled images. Recently, researchers started to use deep learning for the detection and classification of COVID-19 through the use of CT scans images [20].
In the recent literature, different types of pre-trained deep learning models are formed to be used in the classification of Covid-19 such as (GoogleNet[21], Xception[22], U-Net[23], AlexNet [24], VGG19[25], RestNet50[26], MobileNets[27], DenseNet[28], ResNet18 [26], and SqueezeNet[29]), whereas each one has different mechanics but at the end, the main idea is divided into two parts; to extract the features from the images and then apply the classification. Among these classification studies, Shrivastava et al. used CT and X-Ray images in different types of deep transfer learning (Resnet50, InceptionV4, and EfficientNetB0) to extract features followed by an Ensemble Learning for classification, with an accuracy of 97% [30]. A similar accuracy value was obtained by Halder and Datta [31], where more than 2000 CT images were used in DenseNet201, VGG16, ResNet50V2, and MobileNet for a binary classification problem for Covid-19 and healthy cases. Likewise, Eduardo Soares and et al. [32], proposed a deep learning model called xDNN to classify CT scan images that are infected with Covid-19 and non-infected. In addition to the previous studies, Panwar and colleagues [33], proposed a deep transfer learning model using Grad-CAM techniques which monitor the performance of the network that achieve a 95% accuracy. Furthermore, Arora and et al. [34], used deep transfer learning models such as XceptionNet, MobileNet, Inception V3, DenseNet, ResNet50, and VGG16 to classify CT scans where the dataset consists of two classes of 2481 Covid-19 and Non-Covid-19 images on which super-resolution techniques were applied.
Since the accuracy values obtained in the previous studies are fairly acceptable, we tried to increase the accuracy further by modifying the convolutional neural network by incorporating different classifiers after the feature extraction procedure. One of the advantage part of our study is the increased number CT images used in the training set. As a novelty, in the classification part, we implemented K-Nearest Neighbour (KNN) to identify the images. Moreover, we implemented Principle Component Analysis (PCA) on the features deduced from the last Convolutional layer to decrease the dimension of the input to the classifier. Additionally, we adopted SVM to compare the classification performance of the KNN. Thus, we propose to use a hybrid method that companies Deep transfer learning models with SVM and KNN classifiers to identify the binary classes as Covid-19 and Normal.