Pneumonia is a widespread and potentially deadly disease. In the United States alone, about 1 million hospital visits each year are due to pneumonia, according to the Centers for Disease Control and Prevention. Often, a chest x-ray is used to diagnose this infection but even the best radiologists have difficulty interpreting these complex images, said Stanford radiologist Matthew Lungren, MD.
Hoping to give radiologists a boost in their ability to diagnose and treat this and other chest diseases, Lungren and colleagues from the Stanford Machine Learning Group have developed a deep learning algorithm, called CheXNet, to make diagnoses using chest x-rays. This algorithm can diagnose 14 different pathologies and is the most accurate of any algorithm working off the National Institute of Health’s recently released ChestX-ray14 dataset, which is currently the largest publicly available chest x-ray dataset. A paper about CheXNet was published this week on the scientific preprint website arXiv.
Lungren is shown in the photo above discussing the results of tests of the algorithm with the paper’s lead authors, Jeremy Irvin and Pranav Rajpurkar, both graduate students in the Stanford Machine Learning Group.
When it comes to diagnosing pneumonia specifically, CheXNet outperformed four Stanford radiologists working independently – which means it agreed with a majority vote of radiologists more often than did the individual radiologists. Said Lungren, who is co-author of the paper, in a Stanford News article about this research:
The motivation behind this work is to have a deep learning model to aid in the interpretation task that could overcome the intrinsic limitations of human perception and bias, and reduce errors. More broadly, we believe that a deep learning model for this purpose could improve health care delivery across a wide range of settings.
The arXiv paper also discusses a computer-based tool the researchers developed that could help radiologists prioritize their workloads and potentially reduce the incidence of missed pneumonia diagnoses. This tool produces heat map-like images of the x-rays to highlight the areas that are most indicative of pneumonia according to CheXNet. In the future, the team hopes to extend their work to increase access to high-quality, low-cost radiological services in remote and resource-poor areas of the world.
The senior author is Andrew Ng, all of the Stanford Machine Learning Group.
Previously: AI for imaging: Experts delve into its promise and Not just an image: Radiologists boost communication skills
Photo by L.A. Cicero