Exploring the Unknown
New innovations are poised to revolutionize radiology. What will these changes mean for your patients?
Autopilot planes. Self-driving cars. Chess-playing computers. Robotic radiologists? According to leading informatics experts in the radiology specialty, the answer is … not just yet.
While computers aren’t quite ready to take over radiologists’ jobs, machine learning does present both challenges and opportunities for the specialty. And radiologists should prepare now for fast-moving technological evolutions that will shape their practices and the profession itself.
Pushing Past the Buzzwords
“Machine learning, artificial intelligence, deep neural networks, deep learning — these techniques are more plausible today than ever before due to faster hardware and investments in new algorithms and software techniques,” says Keith J. Dreyer, DO, PhD, FACR, associate professor of radiology at Harvard Medical School and chair of the ACR Commission on Informatics. “Until recently, the leaps in machine learning have come largely in the consumer world thanks to advancements in image recognition. As a result, there is now a natural translation of that technology to medical imaging.”
But what exactly is machine learning and how can these powerful technologies revolutionize radiology? Dreyer explains that machine learning is a statistical algorithm that recognizes patterns, learns appropriate responses, and improves over time. Think Watson: IBM’s super computer that beat Jeopardy champion Ken Jennings in 2011.
“Machine learning is on the cusp of being a valuable tool for clinicians,” says Ross W. 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. “The computing power and the quality of the data sets make this a good time to watch what machine learning can do for us. There should be some interesting tools coming along for radiologists and other specialties to help us do a better job. But it remains to be seen how soon these tools will be production ready.”
Overcoming the FUD Factor
In health care, the rise of machine learning has given way to considerable apprehension about what changes it might bring, including concerns that technology may soon replace certain providers, including radiologists. That fear, uncertainty, and doubt (or, as a marketing strategist might say, FUD), however, is unfounded according to Don K. Dennison, president of Don K. Dennison Solutions, Inc., and chair of the ACR Informatics Industry Activities Committee.
“Don’t panic!” he says, “And don’t let the FUD factor turn you against new technology. We’re still at the peak of the hype cycle of machine learning in terms of it being able to replace radiologists in the near future. In reality, we are a long way from that happening.”
Filice agrees that anxiety over technology in the radiology community is not realistic. “I have zero fear that an algorithm is going to replace a radiologist anytime soon,” he emphasizes. “Instead, I look at this as a tool that could be highly beneficial to the practice of radiology.”
Here’s how Filice envisions machine learning impacting a radiologist’s ability to provide better patient care. The typical radiologist is overwhelmed with numerous studies to interpret each day, he says. And each study relates to massive amounts of data that should be used to form interpretations, including clinical data, best-practice recommendations, follow-up guidelines, and institutional processes. “Machine learning will relieve the overwhelming nature of all that data,” he predicts. “It will free radiologists to do what they do best, because they won’t have to plow through mountains of data to provide an intelligent diagnosis or treatment plan.”
Navigating the Human Body
Changing the way we look at data is exactly what an early pioneer in machine learning and radiology has in mind. “When I first got into this arena in 2006, my vision was to look at how GPS technology revolutionized traffic data,” says Khan M. Siddiqui, MBBS, chief technology and chief medical officer of Chicago-based higi (the nation's largest network of health-monitoring stations) and visiting associate professor at Johns Hopkins University School of Medicine. “When engineers first built the GPS to target missiles, they had no idea we’d now be using it for shopping, searching for nearby restaurants, helping us commute, and connecting with each other. GPS coordinates on images of Earth allow us to understand all of the intricacies of traffic and routing, and instantly use that information to accomplish a task. In short, our work focuses on enabling an image-based GPS for the human body.”
Siddiqui’s goal for machine learning is to enable the computer to understand all of the anatomical structures in the body, as well as to sort through all of the other relevant data — including symptomatic, clinical, pharmacy, billing, and much more — to gain similar insights into the body as when using an automated map. “Machine learning technology is moving so fast, especially with self-driving cars. Medical imaging is just catching up to all the advancements, but the work has already started,” says Siddiqui.
He considers machine learning’s low-hanging fruit to be binary tasks: questions with a yes or no answer. Take, for example, a scenario where a nurse put in a PICC line. Is its tip in the right location or not? “That kind of image doesn’t require a lot of knowledge to interpret,” Siddiqui says. “It’s either yes or no, but it needs an immediate answer. That’s an area where machine learning should take over and streamline our workload, optimize our processes, and reduce our costs. It’s a key point where technology can help us change how we’re doing things.”
Another of the most promising applications of machine learning lies in finding objects, describing their features, taking measurements, and identifying patterns, says Dreyer. “Imagine a system that could pre-detect findings on MR while the patient is still in the scanner. Then the radiologist could modify the protocol in real-time,” he says. “Or consider a machine that determines which cases among a queue of thousands are the most important for you to read first. You can train the computer to do that. Then the radiologist can spend less time searching through images and more time aggregating the information from the images.”
Filice also foresees such tools being particularly useful for precision health care. “The algorithms have the potential to take massive data sets associated with a particular patient and their disease or condition,” he says. “Software can then compare the case to best practices and synthesize that into a small amount of high-quality, field-consumable information for the radiologist — all within a matter of minutes. That technology could help us focus on exactly the parts of the exam where we should be applying more attention or using our judgment.”
Bridging the Gap
While many of these advancements are still on the horizon, Dreyer outlines three steps the radiology community needs to take to bridge the gap between where we are today and the best possible future. First is to create the data sets necessary for the algorithms. Second is to perform the machine learning effort itself, where data scientists take those data sets and use them to train the algorithms. The third step is implementing the systems inside clinical applications. Massachusetts General Hospital, where Dreyer works, has started a clinical data science center to work on all three of these areas.
“To date, the focus has been mostly on the middle step: taking existing data sets and training the algorithms for higher accuracy,” he notes. “But we haven’t spent enough time on steps one and three. For machine learning to truly hit the mainstream, we also need to create the data sets in large enough quantity for various clinical challenges and then deploy the technology. I encourage all diagnostic medical centers and others with data and clinical environments to get involved, because it’s going to take thousands and thousands of algorithms with massive data sets to tackle these challenges.”
The ACR has a role to play as well, and the ACR Commission on Informatics (which Dreyer heads) recently formed a Committee on Clinical Data Science that will bring together radiologists and industry partners to formulate nomenclature, definitions, and directions for machine learning in radiology. Dennison says it is imperative for both providers and technology vendors to engage in these discussions to ensure that the tools being developed will be effective in helping radiologists improve how they do their jobs and deliver better patient care.
“There’s a lot of hype and jargon right now,” Dreyer says, “so it’s difficult for people to even have a meaningful conversation. The committee’s first effort will be to establish a common terminology that we can adopt to make it easier to move the ball forward.” The ACR could also provide direction on informatics research, workflow, patient record integration, and other areas in which machine learning can provide the greatest benefit to patients, suggests Dreyer. With its extensive network of clinical trials, the ACR might also be able to provide a large-scale set of images with known findings or patient outcomes to help guide machine learning approaches.
Embracing the Future
How can radiologists prepare themselves for the advent of machine learning? According to informatics experts, the answer is simple: Start getting informed — now!
Says Dreyer, “While the technology itself is still a bit unformed, there is going to be a steady increase of stepwise innovation. So don’t believe all the hype today, but start getting familiar with machine learning so you are ready as our environments begin to change incrementally.”
Filice takes that a step further, saying radiologists should embrace machine learning. “If a vendor or researcher approaches you with a solution or a request to participate, join them on the journey,” he says. “To develop good tools, you need experts to guide the algorithms and to understand what questions need to be answered. Vendors should know what radiologists need at the workstation when they’re going through their cases and what would help them at the point of care. The more clinicians get involved, the more the technology will actually help them.”
Dennison urges radiologists to be patient. “For radiology, there will be hundreds of failures before machine learning finally succeeds,” he predicts, “just like self-learning cars. The first time a technology-driven car crashed, it went viral. But humans cause hundreds of thousands of crashes a week. People are far from perfect, and we should temper our expectations for computers in the early days. But they will learn — just like we do.”
By Linda Sowers, freelance writer for ACR Press