It’s been a common prophesy in radiology for nearly two decades — beware the advent of diagnostic computerized tools, for their expanded use could leave you jobless.
That was the undercurrent when computer-aided detection (CAD) emerged 20 years ago. But, today it’s a critical tool, especially in mammography, and the industry didn’t kick its mammographers to the curb. Instead, they use CAD to more accurately identify potential breast cancers earlier, improving treatments and outcomes.
Yet, this same refrain has re-surfaced now that technology has advanced with smarter tools, specifically ones dubbed artificial intelligence (AI). Whether its machine learning or the more complex deep learning, the excitement around new tools that present the potential for faster, more efficient patient care and work flow is tempered with cautious anxiety around just how many responsibilities and activities these technologies will assume.
But, according to Mark Michalski, MD, executive director of the Massachusetts General Hospital (MGH) and Brigham & Women’s Hospital Center for Clinical Data Science, that fear far outpaces what AI can do now or in the future.
“It’s a prevailing sentiment that people hope AI will help radiologists, not replace them. And, that will be true in the short and medium term,” he said. “But, in terms of replacing radiologists — if you could actually create a machine that end-to-end replaced radiologists, you would also have to replace so many other jobs that the economy would have fundamentally changed, and the world would be changing with it.”
Instead, he said, AI is proving itself to be a strong companion for the radiologist, oftentimes taking over tasks that can bog down a workday.
The MGH Experience
And, Michalski has first-hand knowledge with how AI is complementing radiologists’ work.
Last year, MGH partnered with global computing technology company NVIDIA to employ the DGX-1, a server designed for AI applications. NVIDIA engineers and MGH data scientists developed deep learning algorithms to program the server and used MGH’s phenotypic, genetic, and approximately 10 billion images to train it.
Although more than a dozen other algorithms are still in the works, including neurology and oncology, MGH pediatric radiologists currently use one to estimate bone age. After exposing the algorithm to thousands of images and teaching it to accurately assess bone age, radiologists now use the tool to determine a patient’s bone age based on X-rays.
This feature launched in July, and radiologists use it daily. So far, Michalski said, his colleagues find it useful, and reactions have been enthusiastic.
“It’s worked, and there’s been a lot of excitement about it,” he said. “Now, we’re just getting requests to tweak things on the front end — to make changes to the interface. These are typical requests that come with incrementally improving a model.”
Unequivocally, increasing patient data at your fingertips improves your diagnoses. If you know patient complaints, co-morbidities, and histories, said Eran Rubens, chief technology officer for enterprise imaging at Philips Healthcare IT, you’ll make more informed decisions.
For example, he said, if you know a patient receiving a CT scan for abdominal pain is also HIV-positive, you’ll know to check the lungs, as well.
To meet this need, Philips developed Illumeo, an AI tool similar to a smartphone map application, Rubens said, where dropping a pin on a business reveals contact information and hours of operation. Illumeo, which is still under clinical testing and integrates into Philips’ IntelliSpace PACS, pinpoints locations of clinical interest, and clicking on the marker reveals relevant contextual information.
Illumeo also provides access to prior images, he said. This feature helps radiologists, as well as referring physicians, determine longitudinal lung nodule growth for patients who return for follow-up without having to search the electronic medical record.
“It offers the opportunity to deliver more value as a radiologist,” Rubens said. “At the same time, you can improve on your own efficiency because many of the mundane tasks are automated. It’s a way to balance speed and quality.”
In fact, automation is key for many AI tools, said Joerg Aumueller, global product manager for Artificial Intelligence and Decision Support Solutions at Siemens Healthineers, calling AI the perfect assistant technology. Many products assume duties that make a radiologist’s job easier.
“The average radiologist is forced to interpret images quickly, potentially reducing diagnostic accuracy. When radiologists are rushed, their error rate rises,” he said. “AI combines human and machine to be more powerful than the human alone.”
Additionally, AI automation can help radiologists focus on tasks yielding higher reimbursement, he said. For example, Siemens’ AI technology under development could automatically review low-reimbursement chest X-rays, highlighting abnormalities in the lungs, heart, bones, spine, and vascular system, and make that information available when the radiologist opens the study. Consequently, he said, radiologists can read these scans faster, saving time for more complex cases.
While most AI tools are designed to help the individual radiologist, one company is working to bring the collective knowledge of radiologists together. Unanimous AI, a company founded by technologist Louis Rosenberg, uses technology to bring together industry experts and create a hive mentality.
It’s a different approach already at work in different fields, he said, but it has the same quality-of-care improvement goal. The plan is to bring this technology to the work station within two years.
“In the AI world, the typical approach is to build an intelligence system that can replace humans for a lot of applications, and there’s been a lot of advancement with this in imaging,” Rosenberg said. “But, with swarm AI, we amplify the intelligence of the human. It’s based on the fact that groups can be smarter than individuals.”
For example, in a recent study, the company investigated how well individuals identify a forced versus genuine smile based on still photographs. Alone, participants were wrong one-third of the time, he said. As a group, their error rate dropped to one-sixth.
Radiology is similar, he said, because much of the profession is based upon making subjective judgements. And, many diagnoses have high false-positive or negative rates. By giving radiologists the opportunity to log into a system to simultaneously view and discuss studies, their accuracy rates could improve.
The biggest stumbling block to successfully implementing an AI solution, MGH’s Michalski said, is changing how the tools are perceived. Instead of viewing it as an adversarial measure, think of it as an empowerment.
“People don’t feel threatened by Siri or by Google auto-populating their searches. But, that’s AI people use daily,” he said. “It’s the same with radiology. The actual experience with AI is much more something people are willing to adopt.”
And, to make it easier for radiologists to use AI, Rubens said, Philips worked with clinical partners from day one to design the best possible solutions to meet their needs.
“We didn’t develop a product in the lab,” he said. “Instead, we got feedback from them before we wrote the code. We showed them mock-ups and prototypes beforehand.”
Ultimately, Unanimous’s Rosenberg said, AI can bolster, augment, and improve diagnostic radiology in a myriad of ways. But, without the radiologist, patient care will suffer. To promote AI as an infallible diagnostic tool overlooks the unspoken expertise and knowledge providers amass through years of study and practice.
“It’s easy to think you can replace people with algorithms, and maybe for the easy cases, you can,” he said. “But, for the difficult ones, it’s easy to take for granted the power of human intuition, knowledge, and experience. The goal of AI is to leverage that and amplify it.”