Breaking New Ground with AI

Professor Eliot L, Siegel, MD

In many ways, AI today is reminiscent of the transition from film to digital in the early 1990s. Back then, medical professionals were extremely excited about having images available anywhere, anytime for diagnosis, detection and digital enhancement, but many radiologists also feared that digitization would open ‘Pandora's Box’ by providing everyone, including surgeons and ER doctors, full access to all the images from radiology. Of course, looking back now, those fears were misplaced, and ubiquitous access to images has greatly benefited everyone.

AI has similarly generated a tremendous amount of excitement and some level of fear about reducing the need for radiologists.

“At the American College of Radiology meeting in 2016, Ezekiel Emanuel gave a keynote speech that suggested radiologists may be replaced by computers within the next four to five years,” remarked Professor Elliot Siegel. “And yet, here we are five years later without any significant move in that direction.”

“Computers are faster and better than humans at determining what’s in an image, but a computer doesn’t know that the snowman in a picture shouldn't be there without snow, for example. Any five-year-old could probably beat the strongest supercomputer in the world in the task of finding out what's wrong with a picture?” he continued. “Similarly, radiographs are really challenging for computers to interpret. AI will become very accurate for very specific applications, like finding lung nodules, but I think humans and computers will be working together for a long, long time.”

Convolutional neural networks now allow computers to learn by example, the way humans do. “They can learn in a manner similar to the way my residents and fellows learn,” said Professor Siegel. “Moreover, they can be created in a matter of days to perform the same task as a CAD algorithm that took years to make. The problem is that it may take the same number of years to test and validate these algorithms and essentially get regulatory approval.”

The FDA has recently been talking about the potential of using adaptive automated AI algorithms that dynamically learn and change once they're released to users. “We don't have this in the market yet, but it's a really exciting possibility,” he added.

Many interesting AI advances, including non-pixel based innovations, have recently been achieved. For example, “Chong et al. attained a persistent 17 percent reduction in no-show rates after applying AI to look at reducing patient no-shows,” recalled Professor Siegel. “Another company’s algorithm looks at the body of a radiology report to generate the impression based on many previous reports that a radiologist has created.”

Other hopeful advances include US insurance companies seriously considering reimbursement for AI. There is now a technical fee for reimbursing an algorithm that looks at large vessel occlusion in patients with a suspected stroke.

While it may be many years before Artificial Intelligence becomes seamlessly adopted and incorporated into routine practice, AI has certainly come a long way since the early days. //

See more topics in this webinar.