This difficult era of SARS COVID, a deadly viral disease, which first got spread in Wuhan, a city in China and after that to the whole nation, has become a topic of great concern due to less efficiency in the detection tools which are used in hospitals. In this chapter, the authors have used deep learning frameworks to detect Covid-19, pneumonia, tuberculosis, and no-findings using chest x-ray images. Till this time, no work has been done for classifying the above four classes using a single deep learning model. As this study contains four classes which all are quite similar, so the authors tried various neural network architectures which can deeply analyze to separate the features. The authors have used pre-trained DenseNet-121 for classification purpose and have extended some dropout layers and normalization layers at last to reduce overfitting. As a result of this experiment, the authors achieved a validation accuracy of 97.38%. In this study they have tried to differentiate between four different classes.
TopIntroduction
In early 2020, a new virus began generating headlines all over the world because of unprecedented speed of its transmission. Its origins have been traced to a food market in Wuhan, China in December 2019 (Jiang et al., 2020). From there, it has reached countries which are as distant as the United States and Philippines. The virus which is officially named as SARS-CoV-2 has been responsible for over 100 million infections globally, causing around 2.5 million deaths. The United States is the country most affected. The disease caused by this is COVID-19, which stands for coronavirus disease 2019.
World Health Organization (WHO) declared corona as global pandemic. The disease was spreading so fast, exponentially day-by-day leading to the shortage of testing kits and medical services. Some of the common symptoms of the virus was headache, fever and tiredness and some other symptoms are aches and pain, diarrhea, sore throat, conjunctivitis, loss of smell and taste, rashes on skin or discoloration of toes and fingers (MD et al., 2020). This virus causes breathing problems. It directly affects the lungs therefore it requires urgent detection and medication.
RT-PCR which stands for Reverse Transcript Polymerase Chain Reaction (Fang et al., 2020), is the most widely used method for detection of corona virus. But one of the major disadvantages of this method is its complexity, less accuracy and less sensitivity (Xie et al., 2020). Therefore, we need some different method which is less complex and which left the least room for errors.
For disease which are related to lungs such as Covid-19, Pneumonia and Tuberculosis, radio-diagnostic techniques such as X-ray, CT scan (Computer Tomography) and MRI scans (Magnetic Resonance Imaging) acts as painless and non-invasive methods of disease examination (Ai et al., 2020). These days Artificial Intelligence is being used for automatic detection of disease but these techniques still face challenges of reliability and interpretability.
A large amount of work already has been done in this field to increase the accuracy of detection with least complexity. Methods such as CovNet, ResXnet, DarkNet and Capsule Networks for binary and multiclass detection has been used (Moura et al., 2020). Further (Zhao et al., 2020) found mixed GGO (The lung imaging shows patchy areas in lungs called Ground Glass Opacities) in found mixed GGO in most of the patients, but they also observed a vascular and consolidation dilation in the lesion. GGO consolidation and interlobular septal thickening were also reported by Li and Xia (Roentgenol, Li, & Xia, 2020). These are common CT features of COVID-19 patients. Deep learning techniques have been successfully applied in many problems such as arrhythmia detection (Rajput, Wibowo, Hao, & Majmudar, 2019), skin cancer classification (Nahata & Singh, 2020), breast cancer detection (Shen, et al., 2019), brain disease classification (Menikdiwela, Nguyen, & Shaw, 2018), pneumonia detection from chest X-ray images. M. F. HASHMI ET AL (Hashmi, Katiyar, Hashmi, & Keskar, 2021) introduced a new framework named as Compound Scaled Deep Learning Model, which is an upscaled model of ResNet-50. An accuracy of 98.14% has been obtained in binary classification of pneumonia. and lung segmentation (Rahman, et al., 2020). Due to limited number of radiologists, fast AI techniques will be helpful to come over this situation.