Evaluation of Deep Learning Architectures for Classification in CXR Images

In early 2020, the Coronavirus triggered a global pandemic, with roughly 28.7 million infections and 55.7 lakh deaths reported worldwide. Since the less restrictions by the government across the globe for the economic improvement for the government as well as people and less obeying COVID regulations by the people and lethargic attitude number of cases may rise again. The existing test for COVID is RT-PCR however it takes in the range of 6 to 9 hours' time for getting the result and owing to less sensitivity it offers superior false negative results. X-ray imaging is a popular technique to visualize the lung disease impact and manual diagnosis is more error prone even with experts so a cascaded system using deep learning model like pre-trained and CNN model and many other deep learning model. Although they have a higher accuracy, some questions remain unanswered, such as whether the CNN based model performs better when the CXR input on the lung region.In this paper, we employ ROI to segment the lung using U-net and RESNET-50 to train the segmented images, and then use grad-CAM (Class Activation Mapping) to emphasize which regions the neural network considers essential for classification in each segmented example.