Optimized Deep Transfer Learning for Covid-19 Screening using Chest X-Ray Image

Abstract

The Covid-19 pandemic continues to spread at an alarming rate. One of the methods used to screen potential Covid-19 infected patients is analysis of chest x-ray images. However, the sheer number of patients may overwhelm the radiologists that have to perform such analysis. Therefore, it is desirable to perform automatic screening to provide Covid-19 positive cases as early as possible. In this study, transfer learning by using pre-trained deep residual network model was implemented to perform the binary classification between Covid-19 infected persons and normal (ie., non-infected) persons. We also use automatic cropping of the input images to focus on lung area to optimize the learning performance. Our experiments show that this approach yields better performance achieved, achieving accuracy rate of 99.35% compared to an accuracy rate of 98.08% without application of automatic cropping. The performance of the proposed system when used to classify images taken from a dataset that are completely different from those used in the training process is also satisfactory. The system only produces 1 false negative (out of a dataset containing 66 images) with automatic cropping compared to 3 false negatives and one false positive without cropping. These results show that the pre-trained model with automatic cropping gives superior performance and is suitable to be used in automated Covid-19 screening based on chest X-ray images.