Keith J. Dreyer, DO, PhD, FACR, American College of Radiology (ACR) Data Science Institute Chief Science Officer, explains the state of artificial intelligence (AI) in radiology in 2022.
Although there are about 200 AI algorithms for medical imaging now cleared by the U.S. Food and Drug Administration (FDA), a recent ACR survey of its members showed AI only has about a 2% market penetration rate.
"So, there is about another 98% that fall into the category of potential addressable market," Dreyer said. "Now why is that when there is a lot of enthusiasm and we are past the days from six years ago when radiologists were fearful of losing their jobs to AI because Geoffrey Hinton said we should stop training radiologists because AI will take over in another 5 years. That was in 2016, and are now past the five-year mark and it's ridiculous, because today there is an incredible shortage of radiologists."
There is no fear of radiologists losing their jobs today, so the question has changed to how can this AI technology help them do their job more efficiently, Dreyer explained. It is also clear there is immense interest in AI, shown by the packed AI sessions at meeting like RSNA and HIMSS. At RSNA, there is now a separate area of the expo floor dedicated to radiology AI.
"So, I think there is a much more positive perception on AI than what the market shows," Dreyer said.
But he also said the decision to adopt AI is not really up to the radiologist, it is up to administrators or entire group practices to figure out if AI can help them and if it is worth the cost to adopt it. Often, that decision is at a much higher level in a health system and they do not understand the day-to-day work of a radiologist.
"Is AI saving time for the radiologist such that you make them more efficient and it makes the AI worthy of purchasing?" Dreyer said. "I don't see any published evidence that would demonstrate that as 100% yes. I hear anecdotal stories and I have seen some small studies that have shown this, but I have not seen by and larger that this is the case. But, this will be on a case-by-case basis, because it is not just the use case as to the accuracy of the algorithm, it is also how it was implement that it can make you faster or slower."
How AI is deployed makes a big difference for its utility and time savings. Dreyer used the example that if you have an autonomous self-driving car app, but you had to use it on your cell phone and then connect it to the car, it would be incredibly inefficient. The same is true if an AI application requires an additional login or is not part of the usual workflow, or needs s separate workstation that requires extra time to use it.
"Even the best algorithm in the world that is not connected in the best way is not going to help a lot," Dreyer said.
Should AI gain reimbursement or is it just a practice expense?
Dreyer said the lack of reimbursement for AI might be a driver for why health systems are not spending more to purchase this technology. However, health systems need to consider if the technology can help enhance the workflow of radiologists to make them more efficient, or if AI could help improve patient outcomes.
"There are not a lot of use cases today that have proven to a payer that AI is of value such that they are willing to reimburse the clinician who uses it. There is no reimbursement for PACS either, or for radiology information systems (RIS) or electronic health records (EHR), it's just a practice expense," Dreyer said. "So thinking of AI as a full category of its own, but just as a technology, then the question is does this technology make you more efficient? Does it bring is some new action that can improve care or make it faster so it is worth paying for it?
AI that can identify incidental findings that result in followups for additional testing and treatment could help patients and help catch some diseases that were not being looked for initially. AI that can help facilitate that followup process also could be worth the expense if it brings in additional business for the healthcare system.
AI that can automatically identify a pneumothorax on a mobile digital X-ray system, or alert clinical care teams of a suspected stroke, pulmonary embolism or other emergency conditions also could potentially improve patient outcomes.
Even AI that can help detect cancers, lung nodules or other conditions in the non-acute setting may have benefits to prevent missing things in scans, or to act as a second opinion. As a second set of eyes for radiologists, AI can double check datasets to make sure nothing was overlooked, or to get a second opinion about a questionable area of an image. AI might also help detect rare conditions a radiologist might only come across a few times in their career.
Many AI algorithms also perform complex measurements that are reproducible, image reformatting, anatomical labeling, contouring anatomy, and other tasks that are time consuming. AI in these instances can help reduce the tedious manual processes and may help improve radiology report information and accuracy, while reducing the time it takes to read an exam, especially amid falling reimbursements to read exams.
How are ACR and other medical societies helping with AI adoption?
"The ACR is constantly asking its members what is it that you would like us to do and really trying to understand the voice of the customers," Dreyer said. "We also do a similar thing with the vendors, because they are going to provide us with this technology."
He said education on the use cases for AI is still critical. "We need to make sure our radiologists are really attuned to what can and what cannot happen with AI and when things go off the rails with AI," he said. "We also need to help create standards in the industry so things can move faster. And also making sure as we work on projects with the FDA to make sure this post-market analysis is done in the correct way."
Dreyer said the FDA can direct vendors in the development and regulation of AI, but the agency does not have that same control over the providers of care. He said the ACR really wants to be able to step into that role and help on a national level. "So after this is deployed, how do you know you are using this appropriately and what do you do when it goes off the rails. What happens when you switch from algorithm A to algorithm B? How do you make a decision up front to decide between 10 algorithms and which one is going to work for your population? These are areas where large medical societies can cone in an really lend a hand," Dreyer said.
Dreyer also holds the positions of vice chairman of radiology at Massachusetts General Hospital, chief data science and information officer for the departments of radiology for both Massachusetts General Hospital and Brigham and Women's Hospital, and associate professor of radiology at the Harvard Medical School.
This video is part of a 4-part series of interviews with Dreyer. Here are links to the others:
VIDEO: Where will radiology AI be in 5 years?
VIDEO: Development of AI app stores to enable easier access
VIDEO: Segmenting the Radiology Artificial Intelligence Market by Function
Dave Fornell has covered healthcare for more than 17 years, with a focus in cardiology and radiology. Fornell is a 5-time winner of a Jesse H. Neal Award, the most prestigious editorial honors in the field of specialized journalism. The wins included best technical content, best use of social media and best COVID-19 coverage. Fornell was also a three-time Neal finalist for best range of work by a single author. He produces more than 100 editorial videos each year, most of them interviews with key opinion leaders in medicine. He also writes technical articles, covers key trends, conducts video hospital site visits, and is very involved with social media. E-mail: dfornell@innovatehealthcare.com