TH-SOF-1923-CXR Image-Based Classification of COVID-19 using CNN Model

Abstract

Novel COVID-19 has dispersed all over the world and had taken many lives. Many of them are still fighting against this infectious virus. This virus has not only exaggerated the health of people but has also let down everyone financially, economically and has changed day-to-day individual and social lifestyles. Coronavirus testing through RT-PCR is not affordable by everyone and has a lengthy process. Chest X-ray screening is comparatively easily accessible, cheaper, and acquire less amount of time. Our virologists and health care centers need Computer-based systems for the initial and exact classification of COVID-19 using CXR images. Machine learning and artificial intelligence have grown immensely. This is the best time to come forward and help the health care centers with machine intelligence. This research has been divided into two parts. In the first part we did classification, we have worked on the largest available CXR dataset of COVID-19, Normal, Viral Pneumonia, and Lung Opacity disease. Data augmentation has been applied to those classes that have relatively fewer samples of CXR images. The Efficient-Net model is the latest proposed model of convolutional neural networks. B0 model of Efficient-Net has been implemented in this research for the recognition and classification of COVID19 and other lung diseases using CXR images. Different learning rates have been experimented with and have observed the model accuracy and loss values. The highest training accuracy 99.99% with a minimum loss of 00.68% using the Efficient-Net B0 model has been attained. We have accomplished the best highest test accuracy 99.47% with a loss of 00.67%. The trained Efficient-Net B0 model has been also tested on few available CXR data sets and their accuracy results are quite impressive. The second part is about the localization of opacity. SIIM-FISABIO-RSNA COVID-19 Detection dataset has been used for localization of opacity in CXR images. The latest Yolov5s model has been used for the training of the SIIM-FISABIO-RSNA COVID-19 Detection dataset for two classes Opacity vs nonOpacity. Testing has been performed on 4 different datasets with 5 different ‘Confidence’ values. The accuracies results are quite good when we have dropdown the ‘confidence’ value. Complete localization methodology and their results are discussed in the respective chapters. Transfer learning & fine-tuning methods have also been used for the training of the SIIM-FISABIO-RSNA COVID-19 Detection dataset for the two-class & 4-class classification problem. The achieved testing accuracy with finetuning for a binary class is 89% for the SIIM-FISABIO-RSNA COVID-19 Detection dataset. We have achieved the best-trained model weights that will also work marvelously in the future on the largest available CXR datasets