麻豆果冻传媒影音

Is artificial intelligence (AI) a friend or foe to radiologists when it comes to image interpretation?

Debates about potentially helpful or disruptive technologies are not new to radiology, said Paul Chang, MD, of the University of Chicago, and have centered on PACS (picture archiving and communication system), structured reporting, and speech recognition, which are all now integral parts of the field.

"Without exception, radiologists -- and healthcare in general -- are very susceptible to the 'Gartner Hype Cycle,' a visualization that represents the stages through which new technologies travel from conception to large-scale mainstream adoption," Chang noted. "It is a misconception to think that physicians -- radiologists, in particular -- are early adopters of technology. We are not. We are susceptible to the hype, but because of systemic issues in healthcare we are incapable of consuming it early. We over-promise, initially under-deliver, then, after a while, eventually properly consume technology."

"It is the reason why we are 10 years behind other business verticals when it comes to big data, analytics, and AI," he added. "AI is taken for granted when it comes to things like Amazon, insurance, your automobile, and yet we radiologists are still talking about it. However, we get there eventually."

Chang said that "Amara's Law" -- the phenomenon in which the effect of a new technology is overestimated in the short term and underestimated in the long term -- may be more appropriate in describing how a technology like AI will be consumed in radiology.

For example, the early literature on PACS indicated that there were significant reservations among radiologists regarding the new technology, Chang noted. "Now if you asked everyone if they could function without PACS, they would say, 'of course not.'"

The question of whether the use of AI for image interpretation is good or bad for radiologists is the focus of dueling articles published in the American Journal of Roentgenology, in which several prominent radiologists took on the task of debating the issue.

"The most apparent benefit of using AI tools for image interpretation is to address the challenge of identifying true disease, while minimizing harms associated with false-positive results," wrote Randy Miles, MD, MPH, and Constance Lehman, MD, PhD, both of Massachusetts General Hospital in Boston, in their .

Radiologists must be "intimately involved in guiding the rapid yet careful application of AI for image interpretation in support of access to high-quality affordable diagnostic services for the diverse range of patients in need of radiologists' care," they argued.

Chang said that when talking about the role that AI will play in radiology, its benefit in detection and diagnosis may be overemphasized. "That is not necessarily where the benefit of AI is going to be," he said. "The real benefit is going to be on upstream workflow optimization."

For example, the use of AI should help shorten the time it takes to acquire an MRI. Or, instead of making a patient come back for additional imaging because of an incorrect technical acquisition, AI could notify a technologist before a patient finishes the exam and leaves the office.

"You'll have more optimized scheduling, acquisition, and protocoling," Chang explained. "There will be scores of upstream applications of AI that will make significant measurable improvements in productivity, throughput, and quality -- and if it is being done properly, the radiologist won't even know they're being applied."

"Our future is every other industry's past," Chang observed. "And when you look at other industries, that is the way AI is used, to minimize variability and to improve productivity and efficiency in upstream processes."

"Despite our hopes, AI is unlikely to remain relegated to a supporting role," wrote Frank Lexa, MD, MBA, of UPMC Presbyterian Hospital in Pittsburgh, and Saurabh Jha, MD, of the Hospital of the University of Pennsylvania in Philadelphia, in their .

They noted that radiologists are basically operating at full efficiency as they attempt to deal with an increased demand for their services. Consequently, they feel overwhelmed by the mundane tasks associated with their jobs, such as measuring lymph nodes and lung nodules, and will embrace any technology that will perform these tasks, save them time, and improve their performance and job satisfaction.

However, if AI can perform that percentage of their work (they used a figure of 10%), "then there is no reason why it will not be able to do 20%, then 30%, and then 40%. This probably will not happen overnight. Rather, expect a slow incremental takeover," they suggested.

Furthermore, they argued, if AI can perform such a large percentage of a radiologist's tasks, it stands to reason that radiologists may become less needed.

"A rational firm, whether for-profit or not-for-profit, minimizes costs," they warned. "Radiologists are expensive, inconsistent, and fallible. Reading too fast, reading at the end of a long shift, or reading under other adverse circumstances such as sleep deprivation increases errors. AI is unencumbered by these deficiencies. Thus, both revenue generation and cost reduction motives could reduce the employed radiologist workforce."

Much of that argument is spot on, said Chang. "We do need help, and it is absolutely true -- and inevitable -- that AI will, and should, replace human radiologists in performing certain mundane tasks."

He recalled that when he was a resident, he received a glowing recommendation from a mentor on his ability to hang films very quickly -- a lost art with the technological advances that have occurred in the field of radiology over the years.

"Now, I'm not valued because I can hang films quickly, but for my cognitive skills," Chang added. "And that is the value of AI. We should be looking at replacing all of these mundane things that AI can do better so we can move up the value chain."

As for the argument that increasing adoption of AI means fewer radiologists, Chang said this was perhaps "overly simplistic."

Demand for radiology services "is historically difficult to predict," he noted, particularly in the age of precision medicine. Imaging and image interpretation are not getting any simpler, and precision radiology is necessary to meet the requirements of referring physicians who practice precision medicine, which means more quantification and much more complex phenotypic description, he pointed out.

For example, one of the increasing demands that Chang is hearing from oncologists when imaging the abdomen is to perform sarcopenia measurements. "It would take me an hour to do the segmentation and analysis to perform a sarcopenic measurement," he said. "And oncologists believe it should be done on every cancer patient, and there is no way a human radiologist can do that. We need AI, or else we're not going to do it."

"Our workload and the expectations of ordering physicians is increasing exponentially, and we are going to need help," he concluded.

Comment