Generic Foreign Object Detection in Chest X-rays

KC Santosh, Shotabdi Roy, Siva Allu

Abstract: In an automated Chest X-Ray (CXR) screening process, foreign objects such as coins, buttons, medical tubes, and devices and jewelry can adversely influence the performance of abnormality screening tools. As machine learning algorithms did not separately consider them into account, they result in false-positive cases. In our work, we employ You Only Look Once (YOLOv4) algorithm - a Deep Neural Network - to detect foreign objects in CXR images. Considering its genericity, on a dataset of 400 publicly available CXR images hosted by LHNCBC, U.S National Library of Medicine (NLM), National Institutes of Health (NIH), we achieve the following performance scores: accuracy of 91.00%, precision of 85.00%, recall of 93.00% and f1-score of 89.00%. Unlike state-of-art works, where they are limited to specific type of foreign object (e.g., circle-like objects), this is the first time we report experimental results on all possible types of foreign object.