Introduction
Machine-learning (ML) techniques have been used in medical imaging and infectious disease diagnosis (Lundervold and Lundervold, 2019; Chen et al., 2016; Ardabili et al., 2020). The coronavirus disease 2019 (COVID-19), which started to spread from Wuhan (China), as of the end of December 2019 (Zhu et al., 2020; Huang et al., 2020), has affected the whole world. According to the updated data, there have been more than 134 957 021 confirmed cases and 2 918 752 confirmed deaths because of COVID-19 in 223 countries as of 11 April 2021 (WHO, 2021). Coronaviruses (CoVs) are related to zoonotic viruses that can cause disease in mammal or bird species (Tezer and Bedir Demirdag, 2020). Various medical approaches are available to diagnose and detect COVID-19 in patients, such as the transcription-polymerase chain reaction (RT-PCR) test kits (Ai et al., 2020) and chest computed tomography (CT) images. Chest CT scans have played a vital role in diagnosis during this pandemic (Akcay et al., 2020; Bao et al., 2020; Chung et al., 2020; Lei et al., 2020). Early detection, diagnosis, isolation, and treatment are critical to preventing further spread of the disease (Guner et al., 2020). In some cases, real-time polymerase chain reactions can give incorrect or inadequate information (Ai et al., 2020). It is critical to develop cost-effective and accurate detection methods that all countries can benefit from. Deep learning (DL) is a part of machine learning (Deng, 2014). Recently, this technique has shown effective performance in the field of medical image processing. DL-based research has been presented for the detection of COVID-19. This includes artificial neural networks (ANN), convolutional neural networks (CNN), recurrent neural networks (RNN, a hybrid classifier architecture) (Goreke etal., 2021); ResNet50, Incep-tionV3, and InceptionResNetV2 (Narin etal., 2020); nCOVnet (Panwar et al., 2020); DarkCovidNet (Ozturk et al., 2020); VGG-19 (Ioannis and Bessiana, 2020); COVID-NET (Wang and Wong, 2020); and Xception and ResNeXt (Jain et al., 2021) models using X-ray images. At the same time, studies that detect COVID-19 using CT images have been presented in the literature. AlexNet, VGG-16, VGG-19, SqueezeNet, GoogLeNet (Ardakani et al., 2020), ResNet-50 as a specific model (Li et al., 2020), DenseNet (Yang et al., 2020), and DenseNet201 (Jaiswal et al., 2020) algorithms have used CT images for COVID-19 detection and diagnosis. Although there are studies on the subject, there is still insufficient literature. Deep learning algorithms can help develop a new useful diagnosis and management system for COVID-19 cases. In this study, we have proposed three models for an automatic prediction of COVID-19 using DL-based using chest CT images. The proposed models have an end-to-end architecture that uses feature extraction methods and raw chest-CT images for analysis. These models are a customized CNN, ZFNet, and DenseNet121. We reached this accuracy by means of a large-size dataset and multi-layer models. Along with the customized CNN model, we proposed a small-size nonstandard dataset.
Radiologists have to be pioneers in medical imaging and interpretation during the COVID-19 pandemic, but the medical staff are currently under heavy workloads. Therefore, DL-based approaches can help contribute to the medical system and offer a secondary perspective.
This study is organized as follows: in section 2 (Materials and methods), we give a short overview of the literature in deep learning and the proposed models; we describe how we obtained the dataset, and we present architecture charts and plots. Then, we provide a statistical analysis. In section 3 (Results), we show the results of the experiment. After that, we discuss and interpret the obtained results and conclusions.
Materials and methods
In this section, we define the dataset used in the study for DL. The second part is followed by the proposed CNN, ZFNet, and DenseNet121. It compares the performance of these three models. We built DL-based platforms for automatic detection and prediction of COVID-19 (Figure 1).
Deep learning is a subclass of machine learning and is the most popular approach in artificial intelligence applications (LeCun et al., 2015). DL is a method that imitates the human brain in the use of information and aims for new approaches in complex data solutions. The most important feature that distinguishes deep learning from traditional neural networks is that it has more than one hidden layer (Sejnowski and Tesauro, 1989). Generally, the architecture of convolutional neural networks consists of input, convolution, pooling, convolution, pooling, fully connected, fully connected, and output prediction (Pouyanfar et al., 2018).
The data were downloaded and used from the GitHub public database. The COVID-CT-Dataset has 360 CT images containing clinical findings of COVID-19 from 216 patients and 397 non-COVID-19 CT images (Github/UCSD-AI4H, 2020). No human and no animal rights were violated. The research was performed according to the principles of the Declaration of Helsinki. We used the Keras deep learning library with the TensorFlow backend to implement deep learning models (Figure 3). This study was done on a personal laptop equipped with an Intel 15 processor, 6 GB of RAM, and a GTX 940MX NVidia GPU with 2GB of VRAM. Table 1 shows the DL methods used and statistics of the dataset for chest CT images.
A chest CT dataset was used (Figure 2), which was obtained with different techniques and did not have standard features. Therefore, all images were pre-processed.
Problem solving with the help of deep learning is equivalent to optimally designing a multi-layered network structure. The raw CT dataset and 100 epochs were used as the input layer in our study (Figures 4 to 6). BatchNormalization caused the model to learn better during training, and it also positively affected the stability of the network.
CNNs have been used in imaging-based classification in various medical areas (Lakhani and Sundaram, 2017; Esteva et al., 2017). They have been significantly practiced in medical image processing to develop health research (Choe et al., 2019). CNNs are a kind of artificial neural network with multiple layers contributing to high accuracy and cost reduction in its large datasets (Panwar et al., 2020). The customized CNN's architecture contains 3 layers instead of 2 in the 2-3-4 layers (Figure 4). Therefore, we increased the model's performance in terms of accuracy.
Performance measures
This study used the receiver operating characteristic (ROC) curve to evaluate classifier output quality. ROC curve analysis is generally used in medical studies for evaluating the diagnostic accuracy of a continuous class (Kamarudin et al., 2017). The confusion matrix is based on four parameters, identified as True Positive (TP), False Positive (FP), False Negative (FN), and True Negative (TN), as shown in Table 2. Accuracy, precision, recall, F1-score, and ROC-AUC metrics were used for performance measurement (Togacar et al., 2020).
Experimental results
We used CNN, ZFNet, and DenseNet121 DL algorithms to detect COVID-19 for training, validation, and testing purposes. Visual performance plots are given in Figures 7 to 9, and the evaluation results of the methods are shown in Table 3.
We obtained the best performance with the customized CNN model, with an accuracy of 94,7%, a recall of 90%, a precision of 100%, and an Fl-score of 94,7%. The lowest performance values were obtained by DenseNet121 with an accuracy of 84,2%, a recall of 76,5%, a precision of 85%, and an Fl-score of 81,2%. The confusion matrix for the detection of COVID-19 obtained from the study is given in Figure 7. Accuracy, precision, recall, Fl-score, and ROC-AUC were used for performance evaluation (Table 3).
The loss function layer is used to calculate the expected results predicted by the vital features (Sriporn et al., 2020) (Figure 8).
According to the results of the ROC curve (Figure 9), we obtained the best result from the CNN model with 95%, the second best result from the ZFNet model with 86,8%, and the lowest result with the DenseNet121 model (83,5%).
Conclusions
In this study, we focused on the detection and prediction of COVID-19 using chest CT imaging. As a result, the CNN algorithm gave more satisfactory and higher accuracy than ZFNet and DenseNet121. The additional layers we applied for the CNN model increased the study's performance (while the standard CNN ROC-AUC value was 94%, it increased to 95% with the customized method). We presented a different perspective to the standard CNN approach. According to the results, the customized CNN model can be used to automatically predict the COVID-19 disease.
A major limitation of this study is the use of a limited number of COVID-19 chest CT images; it can offer radiologists and medical staff a second perspective. COVID-19 diagnosis performed using DL-based algorithms can help medical staff with reporting and interpreting. In future work, studies with more datasets and different machine learning methods will be presented.