Artificial Intelligence has been a topic of discussion for some time.
Outside of science fiction, the public’s first real-life exposure may have been IBM’s Deep Blue, which defeated chess world champion Gary Kasparov in a match that took place in May of 1997. In recent years, the evolution of AI has resulted in more than just a computer that can play chess really well. Instead, AI is beginning to play a more active role in radiology. To get a better understanding of what’s happening, we spoke with Dr. Keith Dreyer, a well-known expert and presenter on AI and vice chairman of radiology at Massachusetts General Hospital.
Dreyer has spoken about “AI winters,” or times when AI development essentially comes to a standstill. There have been two, but he doesn’t see a third on the horizon. He says there’s low chance for a winter for a couple of reasons. First, algorithms are performing better and with more accuracy than ever before. In other words, AI works this time around. In the past, there simply wasn’t enough data or the computers needed to process the vast amounts of data needed to make AI viable. Which brings us to the second reason — computing power. Today’s machines can handle huge amounts of data and more organizations are able to afford these machines today.
While more organizations are delving into AI, Dreyer says it’s still a technology that hasn’t really penetrated the market. In order to clear that hurdle, questions will need to be answered. Do the algorithms being created work well on all data? How does it improve the workflow? What’s it going to do for the user? How will it make their day better? What’s the tipping point to get payors on board around reimbursement?
On the topic of payors, Dreyer has yet to see any reimbursement that pays for the use of AI in medical imaging above and beyond what would be paid for service provided by humans not using AI. However, he believes that will change. “I think because there’s obviously so much tight scrutiny under payor reform in general, if you look at the process in which payment is determined by the largest payor — the government, with Medicare and Medicaid — they use a committee process where a group of subspecialty physicians determine where the dollars go. So you’d have to show an algorithm is well-worth its value in order for them to create a new code to be reimbursed for it. It’s not something that happens overnight, it has to be proven in the field,” he says.
Although the data required to fuel AI’s algorithms has improved dramatically over the years, there are practical and political constraints still playing a part in tethering its rise to the stratosphere. That’s because the data is locked behind hospital firewalls. Hospitals are still grappling with the process of making that data available in a safe, secure manner to companies making algorithms. Presumably, some are also still struggling with the thought of providing information that could benefit their competitors, even if those benefits extend to all participants.
With both the data and the people who would benefit from the data all contained within a hospital, it stands to reason that algorithms should be made within the hospital. “The problem with that is there are no data scientists inside the hospital,” Dreyer says.
Enter the American College of Radiology’s Data Science Institute. The DSI’s role is to promote the deployment of safe AI into facilities for medical imaging to better care for patients. The institute is piloting a project with the FDA’s MDDT program to certify algorithms, and later it assesses them in the marketplace after deployment.
The ACR has a program called Assess-AI that also monitors algorithm performance in order to satisfy FDA requirements and provide information to developers in order to improve and enhance algorithms.
For facilities (nearly all) that don’t have data scientists, Dreyer says the ACR launched AI-LAB in May. AI-LAB offers a vendor-neutral solution that goes into hospitals for free to create AI models. Dreyer believes this is a step in the right direction. He thinks everyone should be able to create who wants to. “The tagline we use for this is ‘democratization of artificial intelligence.’ We also hope vendors will provide robust, purchasable solutions for AI creation; if you want a Ferrari instead of a VW, you should be able to get to that if you pay,” he says. “If you want a car, though, everyone should be able to get it.”
To improve and evolve AI further, public relations efforts may need to be brought to bear. “AI algorithms need large amounts of data. “Fortunately, there are technical solutions such as transfer learning, federated learning and ensembles that will help provide the power of combined data without the need to move it off premises to a single location.” he says.
It’s still a long road. Most systems today still require human decision. The ultimate arbiter of truth is human. While Dreyer says you could create a system more accurate than humans, it’s still up to the diagnostician what to say in the report. For a radiologist, there are going to be times the equipment and algorithms are so sensitive that things may me found that can be disregarded with reason. Understanding the broad picture and making judgment calls like that is something AI is probably years from managing.