Deep Learning, Clinical Data Science and Radiology
At a packed ACR 2017 session on machine learning that delved into artificial intelligence (AI) and deep-learning algorithms, co-moderator Raym Geis, MD, FACR, vice chair of the ACR Informatics Commission, posed the question: What should radiologists think about machines that think?
The purpose of Deep Learning, Clinical Data Science and Radiology was to help identify how radiologists can adapt in research and in practice to a host of machine learning tools and technology that is becoming increasingly pervasive in the field. There was a particular emphasis on separating hype from reality and on being “prepared, not scared” when faced with machine learning.
Geis told the audience that, in fact, computers can find patterns in complex data that human beings simply cannot find. They can more easily identify things that are normal or abnormal, he said, adding that everyone in the room was likely already using a machine-learning algorithm in their smart phones. In deep learning, computers are excellent at feature extraction, finding what he called “sub visual features,” and utilizing “super CAD,” which is a souped-up version of computer-aided detection.
Involved Radiologists Won’t Be Replaced
The world, however, will still need radiologists in the presence of machine learning greatness, Geis said. And it will make radiologists even more valuable, he added. But it will also absolutely change the way they work in their practices. “Are we doctors or are we image interpreters?” Geis asked the crowd. “Because if you just want to sit in a dark room, those jobs aren’t going to be around. If you want to take care of patients, then you definitely still have a job.”
He then introduced Garry Choy, MD, a cardiovascular and thoracic radiologist, to talk about his interest in medical informatics and data science. Choy opened by saying that radiologists need to get involved in the deep learning conversation, adding that the field is moving far too fast to learn anything from textbooks.
Instead, he suggested, they should turn to academic institutions and corporations offering degrees and certifications in things like neuroscience data and deep learning fellowships. They should reach out to industry leaders like IBM, Google, GE, Amazon, and others for expert guidance and partnering opportunities. They should turn to online sources and social media outlets for information. And they should rely on the new ACR Data Science Institute™ (DSI), he said, which will work with government and industry to facilitate the development and implementation of AI in medical imaging to improve patient care.
Dollars and Quality Care Are at Stake
People today should think about AI in the same way that people had to embrace electricity years ago, Choy said. Meaning that it will eventually be everywhere in everything and used by everyone.
The next speaker and co-moderator, Keith Dreyer, DO, PhD, FACR, and chair of the Commission on Informatics, continued that theme saying that the future will mean harnessing AI in radiology. There will be a lot of money spent on AI by companies, Dreyer said, as companies shift from industrial to digital entities. But health care knowledge will be invaluable during that transition, he said. “You can’t just throw money at it,” Dreyer said. “You can’t just walk into a hospital, for example, and put in a bunch of new machines and say ‘I’m going to try out some new software here.’”
Internet domains will also be a big part of AI growth, Dreyer noted. AI knowledge will spread via the Internet and new domains. The more radiologists know about it, the easier it will be to implement the tools that use that technology, he said. Think of it as diagnosing a patient, he said. “If you don’t know much about the patient, you’re not likely to get a good diagnosis.”
Dreyer also encouraged audience members to use DSI to implement AI in their practices. It’s not a simple process, he said. You will have to deal with new content in new cases, embrace new computing methods, understand the economics of it, create new standards, commercialize, educate, and stare down a host of legal and ethical issues. The economic success of radiology depends on this entire process, he said.
AI in clinical diagnostics could have a huge impact on patient care as well. AI is applied first to thousands upon thousands of findings and quantifications, Dreyer said. Then it is applied to conditions and their interpretation. It’s the second part where radiologists are needed most. A radiologist, as a human being, might order a different treatment course than one recommended based on a learning algorithm because he or she understands the patient’s particular real-life situation.
Ultimately, he said, AI may be able to shift the potential need for care to a pre-symptom starting point. Right now, Dreyer said, people show up when they have symptoms. “But what if we use AI to detect disease before there are symptoms?” he asked. In a population health management scenario, for instance, wrist sensors could alert people to signs that symptoms may emerge. And pre-PHM, he said, they are looking at genetics to figure out who is predisposed to certain diseases. “The goal is precision medicine,” he said.
All of the presenters agreed that machine learning technology will serve radiologists – through the gradual implementation of AI and deep learning algorithms – and become commonplace in clinical practice. No matter what you think about machines that think, there is a bright future ahead for radiologists willing to put in the work and embrace the changes that are upon them. And ACR’s guidance in overcoming the practical challenges that lie ahead will be a driving force for economic success and increased patient-centered, value-based care.