"We have seen from the U.S. FDA the number of cleared algorithms available for commercial use are escalating," explained Allen. "The tools radiologists have available, whether they are tools for segmentation or diagnostic tools for use in triage and decision support, they are available now in a lot of subspecialty areas of radiology."

But, he said the adoption rate has been very slow. This is partly due to hospitals wanting assurances that the AI will work as intended. There also is a lack of reimbursement to cover the cost of adopting AI.   

"We have not seen good payment models, so one of the fears that we have is that larger academic research centers will be able to adopt AI into clinical use, while smaller practices, particularly in underserved areas, are going to struggle without a reimbursement plan and create health disparities that I don't think anyone wants to see," Allen explained. 

"I think finding the right AI tools and the right value proposition for the institutions is the main thing," Allen said. "On the payment and policy side, there are two reasonable arguments. As a radiologist, I believe will provide us with safer and more effective care, whether that comes from decreases in turn-around times or whether that comes from decision support capabilities, or the identification of critical findings. On the other hand, the payer might say 'well wait a minute, we are already paying you, the radiologist and the expert, to make those same conclusions, so why should we on a fee for service basis provide additional payment?' And I think that is going to be a struggle for health policy makers."

But, he said there are several key AI value propositions that should be considered. 

"AI models can find things that radiologists cannot find, or we as radiologists just can't see," he said.

This can include figuring out a phenotype of a brain tumor using radiomics so the appropriate therapy can be chosen. A tool like that could face faster adoption and possibly come with some sort of payment, Allen said.

Population health is another area where AI can sift through vast amounts of imaging data to identify patients with key incidental findings for things like pulmonary embolism, coronary disease, pulmonary emphysema and hepatic steatosis. 

"All of these things no one is going to care about when a patient is in the emergency department for diverticulitis. They are going to get antibiotics, and that is really the end of their episode of care," Allen described. "And the fact that they have hepatic steatosis, or they have coronary artery calcification, and they are only 40-years-old gets lost, even if we say it it just gets buried in the report or it does not make it to the problem list for followup. So this opportunistic AI screening for population health has a great chance for ROI."

AI can help close areas of health disparities, which is of interest for health system administrations and possibly to payers. 

For AI that can act as a second set of eyes for radiologists, that might be an area where the radiologists figure out what the value is on that type of algorithm for themselves. 

"If we have that second set of eyes that helps us find more breast cancers, maybe that is valuable enough to us to give us that extra piece of mind," he said. Valuable enough to invest in the technology, even if there is no reimbursement. 

But, opportunity for AI may also rest in conversion from a fee-for-service model to a value-based payment model.

"You can imagine in a value-based payment system, any tool that makes you more efficient, if you are getting paid the same, you can pay a little more for AI to become more efficient and it helps your bottom line," Allen explained.
 
Allen also explained trends in how radiology AI algorithms are being brought to market. In addition to app-store models where some vendors offer a menu of various AI options to choose from, another trend lately has been direct integration with PACS vendors. He said another trend that we might see is AI integration directly into the imaging scanners.

He said scanner-based AI might be a better method of AI delivery than via PACS. This is because each time you want to make an upgrade or add software functionality to PACS, there usually is a fair amount of IT involvement, vendor involvement and usually extra money required. However, including AI as part of a bundle of software in a new scanner would make the technology more attractive, he believes.