Deep Learning of Radiology Reports for Pulmonary Embolus: Is a Computer Reading My Report?
Radiology. 2018 Mar;286(3):853-855. doi: 10.1148/radiol.2017172728. Krupinski EA1.
Mention the term deep learning and reactions tend to range from horror to ardent enthusiasm. Is deep learning, also known as artificial intelligence, really the harbinger of doom for the practicing radiologist, simply the latest gee-whiz computer-geek fad, or is it possible that it may actually have a positive role in radiology and health care in general (1)? In this issue of Radiology, Chen et al (2) report on the utilization of a deep learning convolutional neural network (CNN) natural language processing model to extract pulmonary embolism (PE) findings from thoracic computed tomography reports (ie, determine presence vs absence of PE). An important aspect of their study, from the perspective of the results and implications, was the fact that they used free-text reports (ie, not template-based), adding to the complexity of the problem. One might question whether the use of free-text reports could impact the reproducibility of the results, but the authors addressed this as well. They tested both tools in an internal (same institution) and external (another institution) data set, and although performance dropped a little in the external data set (as nearly always happens), the scheme still performed well.