A deep learning based approach for automatic radiology report generation from chest x ray images

Abstract

Chest radiograph, colloquially known as chest X-ray (CXR), is one of the most commonly used newlineimaging modalities for screening pulmonary diseases like tuberculosis (TB), pneumonia, lung newlinecancer, and COVID-19. Though CXR is inexpensive, nonand#711;invasive, easy to acquire, low in newlineradiation dose, widely available in emergency rooms and ambulatory vans, it is one of the most newlinedi cult imaging modalities to interpret. Various studies indicate that there are considerable newlinediscrepancies in the radiological reports written by physicians in emergency or outpatient newlinedepartments. Since the clinical decisions depend upon the sanctity of the radiological reports, it is highly desirable to automate interpretation and reporting of CXRs. This automation would newlinefacilitate early and timely reporting, smoothen the clinical workflow. This thesis presents four newlinemethods for generating significant radiology reports from CXR images. First method is based on newlineco-attention and reinforcement learning, second based on contextual word embedding. Though newlineboth the models achieved significant results, were over-parametrized. Thus third and fourth newlinemodel are the compressed model based on co-attention and contextual embedding, respectively. newlineIn third and fourth models redundant neurons are eliminated by using one-shot global pruning. newlineTherefore, this work presents novel lighter and faster deep neural model to generate radiological newlinereports from CXR images that achieves performance better than the state-of-the-art methods. newline newline newline