A systematic survey on deep learning techniques for chest disease detection using chest radiographs

Abstract

The lung is one of the most sensitive internal organs of the human body that gets infected by constant exposure to chemicals, particles, and infectious organisms in the atmospheric air. Due to this, the most dangerous chest diseases take place which are the leading cause of human disability and death throughout the world. These pulmonary diseases can be recognized by medical imaging techniques i.e. Chest radiography, Computed tomography, Pulmonary and bronchial angiography, Magnetic resonance imaging, Ultrasonography, and Nuclear medicine techniques. A massive amount of chest reports are generated every day that contain a large amount of anatomical and potentially pathological information, but manual detection and classification of chest abnormalities are considered a tedious and time-taking task. In addition, it also requires skilled radiologists as these reports are often difficult to read and differentiate the pathologies of the chest. To eradicate this issue and give a value-added solution; artificial intelligence and deep learning-based algorithms show excellent performance and have proven their effectiveness for object recognition and image segmentation. The primary aim of this study is to present a comprehensive analysis of the deep learning approaches used to identify various types of pulmonary pathologies on CXRs. In addition, we provide a detailed analysis of the most popular open-access CXR datasets, taxonomy of the state-of-the-art works to assist the researchers in preparing the plan for their research contribution, and discuss potential future research directions in this field. More than 350 research papers from various indexing services, including Web of Science, Scopus, PubMed, and IEEE, were considered for this review article. After several selection parameters, it is observed that most of the literature focuses on the DL approach with CXRs. However, few publications have focused on CT scans and Ultrasound images.