Machine Learning and Deep Learning, Big Data, and Science in Radiology
Is machine learning as bad for radiology as they say?
Machine learning is no radiology apocalypse. In fact, the technology presents many opportunities for the specialty.
That was the take-home message of the ACR 2016 session on the topic.
“I, for one, am not incredibly worried about computers taking over radiologists’ jobs,” said Ross Filice, MD, assistant professor and chief of imaging informatics in the department of radiology at Medstar Georgetown University Hospital and chief of imaging informatics at MedStar Medical Group Radiology. “I look at this as a tool that could be really beneficial for us.”
Filice was one of four presenters during the session, which drew a standing-room-only crowd. The other presenters were Keith Dreyer, DO, PhD, FACR, associate professor of radiology at Harvard Medical School; Raym J. Geis, MD, FACR, a radiologist with Advanced Medical Imaging Consultants PC and vice chair of the ACR IT Informatics Commission; and Tarik Alkasab, MD, PhD, a radiologist in the division of emergency imaging in the department of radiology and service chief of informatics/IT and operations at Massachusetts General Hospital.
What Is Machine Learning?
To put it simply, machine learning is a statistical algorithm that learns and improves over time. Dreyer likened these systems to biological neural networks, which allow people to recognize patterns. One of the most popular examples of machine-learning technology is IBM’s Watson, a super computer that in 2011 beat Jeopardy champion Ken Jennings at his own game.
But technology companies aren’t the only ones leveraging machine learning. The health care industry is also beginning to engage the technology, and radiologists have an opportunity to drive its development. Dreyer noted that the ACR has created machine-learning solutions, including ACR Select®, a clinical decision support tool for appropriate image ordering at the point of care, and ACR Assist, a framework for helping radiologists produce structured reports.
Alkasab foresees such tools being particularly useful for things like comparison studies. “One of the most valuable things you can imagine is something that could identify a feature on a prior exam … find it for you on the current report, and then show you the comparison between them … saving you all the effort that goes into trying to create a valuable comparative report,” he said. “If we steer the development of these tools … we’re going to be able to make them things that serve us and make us have to do less of the tedious kind of stuff, the comparison kind of things. It’s actually going to make it possible for us to do more clinically useful work.”
How Can You Prepare?
To get involved in the development of machine learning tools, groups should begin working on the data “plumbing,” Geis said. That means finding and collecting the data to train the algorithms. “The limiting factor is not the algorithms, it’s the data,” he said. “If you don’t have access to the data and you don’t know what data you need, you can’t do any of it. The first thing that you want to start doing is find the data, then figure out how you’re going to get access to it. Your data are really valuable.”
Geis said radiology groups should hire imaging informaticists, data architects, and data scientists who have the expertise to mine data and develop machine learning systems. “You need to do this … because not only are we going to be looking for the data, but we are the data – somebody is looking at us right now,” Geis said. “It is vitally important that you start working on this because this is going to be related to whether you’re going to get paid or not. You’re going to have to come up with the appropriate data products so you can demonstrate your value.”
In conclusion, Geis said radiologists who embrace data science and machine learning have a bright future ahead.