A novel deep learning based method for COVID-19 detection from CT image

1. Introduction

In the year 2020, a new coronavirus, called COVID-19, was marked as the cause of a pandemic disease by the WHO that first spread in China [1]. COVID-19 is a highly contagious Severe Acute Respiratory Syndrome (SARS). The number of confirmed coronavirus cases by the 27th of October 2020 was 43,341,451, of which 29,300,000 were recovered, and 1,157,509 of the infected died [1]. The standard method to detect COVID-19 cases is the Reverse Transcription-Polymerase Chain Reaction (RT-PCR). However, this detection method has some limitations such as shortage of kits, relentlessness, manual methodology, and is time-consuming. Moreover, the positive rate of the RT-PCR test is only 63% [2].

Recently, many researchers indicate that the abnormalities in the chest radiography images of patients and the computerized tomography (CT) scan are different diagnostic approaches for COVID-19 detection [3], [4], [5], [6]. These studies show that visual symptoms in the lugs are different in patients. CT images have some advantages, such as rapid triage of suspected COVID-19 patients, lower risk of transmission, and high availability. Some researchers try to develop a method for COVID-19 detection using CT images [5], [6].

Deep learning-based methods can extract the features of images that are not obvious in the original image [7]. Many studies have applied a Convolution Neural Network (CNN) to detect COVID-19 [8], [9], [10]. Ohata [11] et al. combined the CNNs with some techniques based on machine learning, such as Bayes. They showed the proposed architecture using a support vector machine (SVM) and a linear kernel with an F1-score of 98.5%. Researchers [12] suggested a method using the CNN model and class decomposition. Abraham [13] proposed a model using a multi-CNN and correlation feature selection method. They achieved an accuracy of 97.44%. Several studies have been suggested inserting a preprocessing phase using the Visual Geometry Group (VGG) model to modality the images [14], [15], [16]. The precision of these methods is up to 86%. Azemin et al. [17] presented an approach based on ResNet-101 CNN architecture. The accuracy of the method is 77.3%. The model of Mangal [18] includes a pre-train layer with a 121-layer dense convolutional network and a full-connected layer with a 90.5% accuracy.

Many approaches use transfer learning with CNN architecture [19], [20], [21]. Loey introduced a method based on Generative Adversarial Network (GAN) with deep transfer learning. He used three deep transfer techniques and achieved 85.2% in testing accuracy [22].

Elaziz et al. [23] introduced a parallel framework for automatic COVID-19 diagnosis. This model uses a fraction multi-channel for feature selection.

Some researchers proposed COIVD-19 detection methods for multi-class classification [24]. Khan [25] suggested a model using the deep CNN structure to COVID-19 infection diagnosis. This method uses an architecture based on Xception. The classification accuracy of the approach is 95% for the 3-class model. Karim [26] proposed to use human explanations and a deep learning model for prediction.

Yoo et al. [27] have combined three binary decision trees with a CNN architecture based on the PyTorch frame. The maximum accuracy was 98% for the first decision tree.

In common deep learning classification structures, the feature extraction process is performed only in the convolution layer. The polishing layer only plays the role of minimizing the dimensions of the feature maps and the network. This paper proposed a new pooling layer that performs reducing the network dimensions and feature extraction simultaneously. The feature extraction is performed using Haar wavelet [28]. Moreover, the model introduces a network structure based on Batch Normalization (BN) and Mish Function[29] to reduce the convergence time and achieve better performance.

2. Material and methods

3. Experimental results

The experiments have been run on a PC with GeForce Turbo RTX-2080 GPU and Corei3-9100f CPU running at 3600 MHz. The implementation of the model was in Python 3.7 using the Pytorch library. All experiments have been done in GPU. Because of the difference in class numbers in any dataset, the weighted random sampler [34] was used as a Sampler in the training step. Table 2 shows all training parameters.

The proposed Deep Neural network was trained using Covid-19 CT Scan images. The training phase used eight different optimizers. Table3 shows the metric results of the proposed approach. This table expresses the best results in metrics achieved using the RAdam optimizer.

4. Discussion

This work proposes a deep model based on batch normalization and Mish function to detect COVID-19 cases from CT images. Two real datasets were used to evaluate the proposed model. The metric results of the proposed network showed an accuracy of 99.03.

This study tries to introduce a network with kernel-based machine learning applications [47]. Therefore, some networks have been selected for the test phase suitable for real-time and online applications. According to Table 4, the worst result in popular DNNs belongs to VGG11 with a 92.73% accuracy. This network is too deep and has 128 million parameters. The best performing DNN was Inception-v3 with a 98.11% accuracy. As shown in Table 4, the best recognition time in popular DNNs is achieved using Resnet18 and is 4.262 ms. This network has more than 11 million trainable parameters and has about 93.39% accuracy on the test dataset. According to the Cohen Kappa score rate for Resnet18, it is about 88.98%; the classification is not good.

Although all of the popular DNNs in our work are powerful networks in classification problems for RGB images [48], they do not have good CT image behavior. Comparing to these networks, the proposed model has four novel innovations; using wavelet [49] for pooling the feature maps, Mish [33] activation Function, Mini-Batch Normalization [32], and SE-blocks [50]. Using wavelet transform into network structure is a reason for extracting the correct and appropriate information for recognition tasks. Mish Function is a novel activation function used in powerful detection and classification networks, such as yolov4 [51], which has an essential role in the appropriate information extraction. One of the essential parts of the proposed model is using mini-batch normalization in any layer. Squeeze-and-Excitation blocks (SE block) is an architectural unit that can be plugged into a CNN structure to improve performance with only a slight increase in the total number of parameters. Squeeze-and-Excitation blocks explicitly model channel relationships and channel interdependencies and include a form of self-attention on channels. Consequently, it results in the best performance in classification metrics parameters and recognition time.

Furthermore, the number of trainable parameters is more suitable for CT images than more deep structures that suffer from overfitting. The proposed network is the best in classification metric parameters comparing to the popular DNNs. Moreover, WCNN4 has the least prediction time. Therefore, it is a helpful network for real-time application for the recognition of covid-19 using CT images.

Fig. 5, Fig. 6, Fig. 7, Fig. 8 show the networks are similar in the final Epochs in the training phase. However, there are differences between them in the network test phase. This difference is generally due to the complexity of the network structure. Bigger structures need more data in the training phase. As shown in Fig. 7 and Fig. 8, vgg11 has the most considerable accuracy entropy in the training phase, and the proposed network has a smooth curve. Fig. 9 and Fig. 10 are the accuracy and loss for test data after each epoch. As shown, the suggested model has decreasing trend entropy in loss and accuracy.

Finally, an important achievement of the proposed network's evaluation is that the test mode's recognition time was 0.069 ms in GPU.

5. Conclusion

Rapid and accurate diagnosis is necessary as the number of patients with COVID-19 infection is increasing. This work proposes a deep learning-based model to detect COVID-19 disease from CT images. The model is an automated method without any feature extraction phase. The proposed model has four novel innovations, including wavelet for pooling the feature maps, Mish activation function, mini-batch normalization, and SE-blocks. For evaluation, the model was trained and tested on the datasets of two public hospitals. The results show the suggested model achieved excellent metric parameters in classifying COVID-19 cases, such as accuracy of 99.03. Moreover, it decreases trend entropy in loss and accuracy. Finally, the recognition time in test mode was 0.069 ms in GPU. The experimental results show the proposed model is a helpful network for real-time application for recognition of covid-19 using CT images.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.