Covid-19 Detection Using Combined Imagery Datasets With Deep CNN’s

Abstract

COVID-19 is the currently on-going pandemic that has caused a global chaos and brought life into a verge of death. A significant impact of the COVID-19 may be seen in practically all aspects of life. Since it is identified just in the recent few years, and so there is a dearth of information on it, how to recognise it, and how to treat it. There is currently no known cause for this outbreak, but research is ongoing to iden tify a treatment. Due to time and money constraints, it is not viable to test for the coronavirus given the daily increase in cases. Accurate identification and diagnosis is crucial. The initial methods of detecting COVID-19 disease mostly relies on the expert research interpretation of Computer Tomography Scans or X-ray images. In this thesis, the major objective is to develop a deep learning model which can determine whether a patient has COVID-19. A literature review and experiment are planned to find a vi able algorithm for such a model. Evaluating the characteristics and main features that affect the prediction model. In the research, by using a unique convolutional neural network (CNN) identified COVID-19 in X-rays, CT Scans and ultrasounds. Combining the three different imagery data and forming a large dataset of X-Rays, CT Scans and ultrasounds such that a single model deals with the different type of images. Using fine-tuning of three CNN models including DenseNet121, ResNet101V2, NASNetMo bile and MobileNetV2 on COVID-19 detection. Performed two main experiments, in first experiment we have trained separate models by considering each dataset, and in second experiment we have trained a combined model by using combined imagery data for COVID-19 detection. This study employs data augmentation techniques to boost the artificial number of photos due to the limited quantity of COVID-19 images. The three different deep learning models Resnet 101v2, Mobinetv2 and Inceptionv3 were compared on this combined dataset and separately on each image data. The model is trained and then accessed using the transfer learning approach to classify between the normal and a COVID patient. The experimentation showed that the an accuracy level of DenseNet121, ResNet101V2, NASNetMobile and MobileNetV2 is 88.21%, 93.02% , 89% and 88.89%, respectively. It reveals that the CNN model with the ResNet101v2 has higher accuracy than DenseNet121, NASNetMobile and MobileNetV2 model on the data combination of X-rays, CT Scans and ultrasounds, and generates best predictions for COVID-19. This can not only boost the COVID-19 detection process but also aid in the instant treatment process. It is an effective method for classifying information, that humans might not be able to do.