A Chest X-ray Image Retrieval System for COVID-19 Detection using Deep Transfer Learning and Denoising Auto Encoder

The COVID-19 pandemic is the defining global health crisis of our time which is currently challenging families, communities, health care systems, and government all over the world. It is critical to detect and isolate the positive cases as early as possible for timely treatment to prevent the further spread of the virus. It was found in few early studies that patients present abnormalities in chest radiography images that are characteristic of those infected with COVID-19. In the current context, a rapid, accessible and automated screening tool based on image processing of chest X-rays (CXRs) would be much needed as a quick alternative to PCR testing, especially with commonly available X-ray machines and without the dedicated test kits in labs and hospitals. Several classifications based approaches have been proposed recently with encouraging results to detect pneumonia based on CXRs using supervised deep transfer learning techniques based on Convolutional Neural Networks (CNNs). These black box approaches are mainly non-interactive in nature and their prediction represents just a cue to the radiologist. This work focuses on issues related to the development of such an automated system for CXRs by performing discriminative feature learning using deep neural networks with a purely data driven approach and retrieving images based on an unknown query image and performing retrieval evaluation on currently available benchmark datasets towards the goal of realistic comparison and real clinical integration. The system is trained and tested on an image collection of 1700 CXRs obtained from two different resources with encouraging results based on precision and recall measures in individual deep feature spaces. It is hoped that the proposed system as diagnostic aid would reduce the visual observation error of human operators and enhance sensitivity in testing for Covid-19 detection.