Pneumonia Detection and Classification on Chest Radiographs using Deep Learning

Abstract

Computer Aided Diagnosis (CAD) is progressively becoming a reliable tool in enhancing the productivity and accuracy of a radiologist in detecting abnormalities on chest radiographs. Detection of airspace disease such as pneumonia can be facilitated with the help of image processing and deep learning algorithms. In this study, we aim to develop and evaluate the performance of a deep learning model to detect pneumonia on chest radiographs. We have used RetinaNet with Resnet-101 as backbone architecture and trained on chest radiographs with pneumonia findings. The trained model was validated on open source datasets from the US National Institutes of Health (NIH), The Society of Thoracic Radiology (STR) and private data repositories. For the purpose of validation, bounding boxes enclosing the ground truth were used as the inference standard. The deep learning model correctly predicted pneumonia on chest radiographs with an accuracy of 96.33%, the sensitivity of 97.51%, specificity of 95.55% and Area Under Receiver Operating characteristic Curve (AUROC) of 97%.