From scheduling patients to interpreting images, artificial intelligence is poised to fundamentally change radiology, offering new potential applications and insights that will enhance scanning accuracy, workflow and patient outcomes.

But it’s still early, practices have been slow to adopt the technology and there are more questions than answers. From concerns over accuracy and biased algorithms, to skepticism over the specific value AI can provide on a day-to-day basis. From a wide-angle perspective, some healthcare practitioners, including medical students, worry the technology could someday render radiologists obsolete.

For Dr. Michael Atalay, professor in the department of diagnostic imaging and vice chair of imaging research in the Alpert Medical School at Brown University, addressing concerns about AI someday replacing radiologists is an important part of the job. He stresses, to students and colleagues alike, that AI will “assist” radiologists.

“We have to show how we’re embracing the new technology and the firm belief that it will enhance our roles as diagnosticians and improve patient safety and outcomes,” Atalay, who also serves as director of cardiac MR and CT at the Alpert Medical School, told HCB News. “Along the way it will also likely improve job satisfaction and quality of life and decrease the high physician burnout rates that we’re seeing now.”

Figuring out where and how AI fits into an individual radiology practice’s workflow, Atalay and others say, will be a critical step toward gaining physician trust and accelerating adoption.

Gaining the radiologist’s trust
Workflow at one practice is unlike another, based on different demographics, patient populations, data environments, prevalence of conditions and several other factors. As a result, AI algorithms that work well in one place may do little for or even hamper another’s workflow.

For instance, for a period, Rhode Island Hospital’s radiology department adopted an application for identifying intracranial hemorrhage on CT scans in the emergency department. While potentially beneficial in other facilities, the technology created delays in the busy ED’s workflow and was eventually removed.

For AI to be useful, providers must also understand how to use it in their specific workflow, according to Atalay. “Does the AI live at the scanner, on the PACS station, or on a standalone workstation? Does every pertinent imaging study have to go to that site or through an application somewhere else before it is eventually reviewed?” he said. “To improve efficiency and efficacy—and not cause disruption—these tools should be directly and transparently integrated into the workflow.”

The FDA, to date, has cleared over 200 AI applications for various uses in radiology, says Dr. Bibb Allen, FACR, chief medical officer for the American College of Radiology’s Data Science Institute and a general diagnostic radiologist in the Birmingham Radiological Group at Grandview Medical Center in Alabama. He warns that FDA clearance does not equate to universal efficacy.

“We just can’t rely on an FDA clearance to say, ‘Right, it’s FDA-cleared, so that means it can be used on every patient based on every example we have.’ That’s clearly not the case,” he said.

Bias incorporated in the development and training process can also interfere with AI’s clinical accuracy when deployed. “If there are racial differences or gender differences, and we haven’t properly characterized that because we don’t have enough of the various subpopulations, then we’re at risk for not providing optimal care for our patients,” said Dr. Bradley Erickson, professor of radiology and director of the Mayo AI lab.

Erickson says that larger, higher-quality data curated by standards is necessary for eliminating biases. In agreement is Dr. James Brink, enterprise chief of radiology for Mass General Brigham Health System, radiologist in chief at Mass General Hospital, and chair of radiology at Brigham and Women’s Hospital, who says that the best way to go about this is to train and develop AI sequentially in various facilities with diverse demographics, populations, data environments, and different equipment and systems.

“By circulating algorithms in development among multiple sites, a federated approach to algorithm development can help eliminate the biases that would come with simply training in one environment with one demographic,” he said.

Dr. Shandong Wu, a tenured associate professor of radiology and founding director of the Pittsburgh Center for AI Innovation in Medical Imaging, adds that more clinical trials are needed for AI evaluation, with only a handful currently underway worldwide.

He says that radiologists must be involved and perform rigorous assessments that result in evidence that not only shows how a specific model works but also helps understand why it is capable of performing a task.

“If the model can do some explainable diagnosis using radiologist-acquainted clues, that is helpful for physicians to understand how it works and to gain their trust and perspective,” he said.

Overcoming roadblocks
With inflation through the roof and tighter margins, radiologists, like other clinicians, are concerned about their return on investment when it comes to the technologies they use and procedures they perform. This has been a major roadblock for the adoption of AI, with no standardized reimbursement system set up to compensate clinicians utilizing the technology.

And while CMS funding and designations such as the new technology add-on payment are a starting point, Allen says that a proper payment system is needed.

"There's not a good model for reimbursement, and CMS and other third-party payers may be questioning if AI adds value for patients in addition to what they already do. When we look at taking care of populations of people rather than our current transactional payment system, then AI becomes a resource that could make institutions more efficient. They would then gladly fund its implementation," he said.

Additionally, most AI models are currently designed to detect only one finding, maybe two, in patients. This limits their ability to make accurate diagnoses, leading to delayed care.

“Does it just detect one thing, like pulmonary embolism, but not pneumonia if applied to the chest,” said Brink. “Is there a fracture? Is there a pulmonary embolism?” That doesn’t necessarily answer the more concrete question, “Is there an abnormality in this chest X-ray or is there an abnormality in this femur?”

Another problem, according to Erickson, is the limited ability of the technology to pull information from multiple sources. Most only use information from the pixels found in images they scan and do not incorporate background information from prior exams and medical records.

“The vast majority of radiology requires clinical information integration and/or other imaging examination information in order to really do a good job. That’s a much more challenging thing to do from an informatics perspective, and you need to train a more sophisticated AI model,” said Erickson.

Because of these limitations, many radiologists have been left scratching their heads on what true value AI offers radiology and how to incorporate it into their practice.

“All of these studies need to be looked at by a radiologist,” said Atalay. “It’s not clear to me how AI will alleviate the number of studies a radiologist has to look at or the time spent on them. But that’s the promise.”

Educating the next generation
In a study of 532 medical students, 23% said they would not pursue a career in radiology, because they believe AI will eventually replace radiologists and limit their future job prospects. One class was surveyed twice, with 50% in 2017 saying radiology was a no-go for them. This rose to 71% in 2021.

Atalay, who led the study, says this indicates a lack of awareness about the role that radiologists play and what the full scope of their work entails. In addition to interpreting images, a significant fraction of their day is spent on noninterpretive tasks such as protocoling exams; determining if a specific scan is safe, effective, and appropriate for individual patients; interacting with the technologists performing the exams; and discussing study indications and results with clinical colleagues.

He adds that it’s important that students understand that current AI tools are “single-minded one-trick ponies” that only detect one or two abnormalities, versus a myriad of potential causes to explain a patient’s condition, let alone a host of potentially important but unsuspected incidental findings.

“In essence, AI now and in the foreseeable future is an assist device helping radiologists perform their tasks. We in radiology already know this. But it is incumbent on us to share this with medical students and disabuse them of misconceptions. When medical students rotate through a radiology elective they are often surprised by the breadth and scope of our jobs,” said Atalay.

Wu says that medical associations are beginning to do this by developing educational programs for students and current physicians and radiology leaders about what they can expect and need to know about AI and its use in their work. His own institution is creating structured training
programs for educating students and clinical trainees, from the high school level
to graduate school.

“We need to make them think with a right attitude toward AI, ‘Ok. I understand what AI is, and I know what it does and how I can work with it. It will not usurp my career. In fact, this shows that radiology might be a more exciting career because of the strong component of AI in radiology,’” he said.

Interpreting the value of AI
Accelerating the adoption of AI in radiology and other medical specialties is a slow process and will take time. Refinements and the development of more sophisticated models, capable of performing multiple tasks and collecting more information will facilitate greater interest from radiologists by showing them the value it brings to their jobs and workflow.

Erickson says that as more advanced developments take shape, he is confident that radiologists will come to view AI as an assistant, and not a replacement. “There’s so much more information in the images than what we can extract today, but I think as long as we remain engaged in how the AI tools are developed and deployed, that we’re going to be at the center of that and not pushed aside.”

In addition to ensuring more accurate diagnoses and optimal care, the technology will improve the patient experience and safety, as well as communication between radiologists and other clinicians, says Atalay. “It will potentially offer us guidance in terms of management that we can then pass on to our clinical colleagues who are directly managing these patients.”

While refining these capabilities should be a central focus, ensuring equal access among different institutions is also important. “The big, well-endowed academic hospitals can’t be the sole users of this new technology because they have enough money to pay for it. If the new technology is such a game changer, then patients in places like rural Alabama ought to have equal access,” said Allen.

Wu also warns that as technology advances, solutions like AI will be more frequent targets of cyberattacks. As a result, radiologists and providers should have in place safety mechanisms for preventing these attacks from exploiting the technology and putting operations and patient care at risk.

“We want to make sure when we deploy something in the health informatics system or clinical workflows, that these softwares are safe to patient data in terms of what they do, and that there is awareness and solutions on preventing potential cyberattacks to the vulnerability of AI systems.”

Leveraging the full potential of AI in radiology and addressing any gaps in its implementation requires radiologists, manufacturers, AI researchers and healthcare leaders to work together and be transparent about what they need in the development and training process.

Doing so will, in turn, ensure that the value that the technology brings to radiology is understood and embraced, while assuaging concerns about it replacing radiologists.

“I’m confident that the depth of diagnostic potential and the breadth of opportunities for advancing the diagnostic frontier will forever allow us to develop new diagnostic methods not yet imagined, and move that frontier toward more precise and more accurate diagnoses from where we are today,” said Brink.