MoBMGAN: Modified GAN-Based Transfer Learning for Automatic Detection of COVID-19 Cases Using Chest X-ray Images
Rajashree Nayak, Bunil Ku. Balabantaray and Dipti Patra
INTRODUCTION
The novel coronavirus 2019 (COVID-19) disease caused by Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) has been acknowledged as a wide-reaching pandemic by WHO. This pandemic has buckled the health-care systems of the whole world and has also drastically lowered the world economy. The exponentially increasing mortality rate of this dreadful disease across the world demands its early-stage detection. Early diagnosis of COVID-19 disease enables us to efficiently plan treatments and to take proper diagnostic decisions. This will preclude the rapid spread of this epidemic. Popularly used RT-PCR tool kits to detect the presence of the COVID-19 virus are sophisticated to handle, are costly, and have lower sensitivity rates which introduce false-negative results. Nevertheless, the detection process is further affected due to the dearth of expert physicians in remote areas. To resolve these issues, radio imaging (RI) techniques (X-ray and CT scan) are integrated with advanced artificial intelligence (AAI) techniques (deep learning models (DLMs) or transfer learning models (TLMs)) to provide superior performance in detecting COVID-19-infected patients (Lin Li et al. 2020; Salman et al. 2020). As compared to DLMs, TLMs outperform in terms of detection accuracy, generalization error, and training time at the cost of a vast amount of labeled training data (Salman et al. 2020).
This chapter focuses on the development of a computationally efficient automated detection scheme by utilizing modified Generative Adversarial Network (MGAN)-based transfer learning (TL) for the successful detection of COVID-19 cases using chest X-ray images. Initially, the MGAN model is used for the generation of a synthetic dataset with varied feature attributes for the initial training of DLMs. MGAN is more computationally efficient than the usual GAN model, as it utilizes a lighter weight network architecture instead of a deep network, and is free from the problem of instability by utilizing a weighted combination of content and structural loss functions. Consequently, the MGAN model produces realistic generated samples at a faster convergence speed. Afterward, various benchmark DLMs such as VGG19, ResNet50, InspectionV3, InspectionResnetV2, DenseNetl21, DenseNetl69, DenseNet201 and MobileNet are fine-tuned with the generated dataset. In the testing phase, the COVID-19 infected X-ray images are fed to these DLMs for the purpose of classification and an outperforming model is chosen for real-time use. Extensive experimental analyses have been performed on popularly available image datasets. The detection accuracy of our work is compared with similar state-of-the-art detection methods. The suggested MGAN model along with the pre-trained MobileNet model outperforms the other models.
The discussion flow of this chapter is structured in four sections. Section 3.2 delivers a brief discussion on some of the deep learning (DL)-based COVID-19 disease detection methods by using X-ray images in literature. Section 3.3 discusses the MGAN model for the generation of a dataset for the training of DLMs. Section 3.4 focuses on the performance analysis of our suggested work compared with some of the similar classification methods in the literature. Section 3.5 concludes the chapter, along with some future scope for this work.
Literature Survey
AAI-based methods utilizing various radiometric images (chest CT scan and X-ray images) provide an excellent breakthrough in detecting COVID-19-infected persons. This section briefly summarizes some of the existing DL-based models (Rahul Kumar et al. 2020; Linda Wang and Alexander Wong 2020; R K. Sethy and S. K. Behera 2020; Ali Narin et al. 2020; Tulin Ozturk et al. 2020; Karim Hammoudi et al. 2020; I.D. Apostolopoulos et al. 2020; E.E.D. Hemdan et al. 2020; Abdul Waheed et al. 2020; N.E.M. Khalifa et al. 2020; Mohamed Loey et al. 2020; Parnian Afshar et al. 2020; Ferhat Ucar and Korkmaz Deniz 2020; M.Z. Islam et al. 2020; T. Mahmud et al. 2020) by utilizing chest X-ray images for the detection of COVID- 19 infections from uninfected cases. Table 3.1 reviews the brief description of these
TABLE 3.1
COVID-19 Infection Diagnosis via Existing DLMs Using Chest X-Ray Images
References |
Number of Cases Used in the Experimentation |
Used DL Model |
Observations (Accuracy) |
Rahul Kumar et al. 2020 |
Training set: (42 COVID-19 cases + 894 normal + 897 pneumonia cases). Testing set: (20 COVID-19+447 normal+448 pneumonia cases) |
ResNet 152 model+ SMOTE |
Accuracy using Random Forest: 0.973 Using XGBoost predictive classifier: 0.977 |
Linda Wang and Alexander Wong 2020 |
53 COVID-19 + 5.526 Non-COVID-19 cases+ 8,066 healthy cases |
COVID-Net |
Model helps the clinicians to screen patients with an accuracy of 0.924 |
P.K. Setliy and S.K. Behera 2020 |
50 COVID-19 cases+ 50 Non-COVID-19 cases |
ResNet50+SVM |
Classification accuracy via SVM is 0.953 |
Ali Narin el al. 2020 |
50 COVID-19 cases+ 50 Non-COVID-19 cases |
Deep CNN+ResNet-50 |
Accuracy for binary class classification is 0.98 |
Tulin Ozturk et al. 2020 |
125 COVID-19 cases+500 Non-COVID + 500 pneumonia cases |
DarkNet+ 17 layered model |
For binary class, accuracy is 0.0980 and, for multi-class detection, accuracy is 0.870 |
Karim Hammoudi et al. 2020 |
Training set: 5,863 images Test set: 145 COVID-19 cases |
Tailored DL + ResNet34, VGG16 ResNet50 and DenseNetl69 |
Classification accuracies for COVID vs bacteria is 0.979 |
I.D. Apostolopoulos et al. 2020 |
224 COVID-19+ 700 pneumonia cases+504 |
TL+VGG19 model |
Model achieves accuracy of 0.9875 for |
healthy cases |
binary classification |
||
E.E.D. Hemdan et al. 2020 |
25 COVID-19 cases+ 25 normal cases |
COVIDX-Net |
The model detects at an accuracy of 0.90 |
Abdul Waheed et al. 2020 |
Generated via GAN: 1,399 normal cases +1,669 COVID-19 cases |
GAN+VGG16 model |
The model achieves a classification accuracy of around 0.95 |
N.E.M. Khalifa et al. 2020 |
Generated via GAN: Normal cases: 5,863 images COVID-19: 6,240 cases |
GAN+AlexNet, GoogLeNet, SqueezeNet, Resnetl8 |
Model using Resnetl8 achieved a classification accuracy of 0.99 |
Mohamed Loey et al. 2020 |
Generated via GAN: Normal cases: 2,100 images COVID-19 cases: 1.800 images |
GAN+AlexNet |
Classification accuracy is around 100% for the binary class problem |
(Continued)
TABLE 3.1 (CONTINUED)
COVID-19 Infection Diagnosis via Existing DLMs Using Chest X-Ray Images
References |
Number of Cases Used in the Experimentation |
Used DL Model |
Observations (Accuracy) |
Parnian Afshar et al. 2020 |
COVID-19 cases: 53 images Non-COVID-19 cases: 5,526 |
CapsNet |
CapsNet model obtains classification accuracy of 0.9830 |
Ferhat Ucar and Korkmaz Deniz 2020 |
Total augmented cases: 4,608 COVID-19: pneumonia: normal: 1,536: 1.536: 1,536 images |
Bayes-SqueezeNet |
Classification accuracy for the three-class problem is 0.9830 |
M.Z. Islam et al. 2020 |
Total cases: 4,575 images COVID-19: normal: pneumonia cases are 1,525: 1,525: 1,525 |
CNN+LSTM |
Detection accuracy for two-class problem is around 0.994 |
T. Mahmud et al. 2020 |
Total cases: 5,856 images Normal cases: 1,583 Non-COVID: 1.493 Bacterial pneumonia: 2,780 |
CovXNet |
Pre-trained CovXNet model provides detection accuracy of 0.974 for two-class problem |
classification models along with the size of the dataset used and achieved classification accuracies for different class problems.
Rahul Kumar et al. (2020) used both ML and DL techniques for the accurate prediction of COVID-19 patients. Here, a ResNetl52 model with synthetic minority over-sampling technique (SMOTE) was used to extract deep features, whereas random forest and XGBoost classifiers are used for the binary classification. Linda Wang and Alexander Wong (2020) have proposed the COVID-Net model to classify the COVID-19 images with an accuracy of 92.4%. R K. Sethy and S. K. Behera (2020) utilized various CNN models to extract features of the X-ray images, and these feature vectors are classified via support vector machine (SVM) classifier. They found that ResNet50 model with SVM classifier provides better accuracy. Ali Narin et al. (2020) used the pre-trained ResNet50 model for the classification purpose. Tulin Ozturk et al. (2020) utilized the pre-trained DarkNet model for the binary and multi-class classification of COVID-19 images. Karim Hammoudi et al. (2020) have proposed a tailored deep learning model via various CNNs architectures such as ResNet34, ResNet50. DenseNetl69, VGG19 and InceptionResNetV2-RNN to detect multi-class pneumonia-infected cases using chest X-ray images. I.D. Apostolopoulos et al. (2020) utilized a TL-based model with pre-trained VGG19 model for the classification purpose. E.E.D. Hemdan et al. (2020) utilized a DL-based model named COVIDX-Net, comprising of seven CNN models to diagnose COVID-19 X-ray images. Abdul Waheed et al. (2020), N.E.M. Khalifa et al. (2020), and Mohamed Loey et al. (2020) utilized a Generative Adversarial Network (GAN) model in the data augmentation process to generate a sufficient amount of training images, followed by the utilization of various TL models for the classification process. TL models pre-trained via the generated images boost up the classification accuracy. Parnian Afshar et al. (2020) and Ferhat Ucar and Korkmaz Deniz (2020) utilized capsule network-based and deep Bayes-SqueezeNet-based frameworks, respectively, for the identification of COVID-19 cases. In contrast, M.Z. Islam et al. (2020) utilized the combination of deep CNN framework with the long short-term memory network for the detection of the novel coronavirus. Deep features were extracted via CNN, and these extracted features were fed to the later model for the classification process. T. Mahmud et al. (2020) utilized multi-receptive features by utilizing multi-dilation CNN models for the automatic detection of COVID-19 infections.