Radiology in the age of AI

When Meghann Maiellano, 41, began experiencing sharp, stabbing abdominal pain, causing severe cramps, she had no idea what was causing the discomfort. 

“I thought it was my appendix,” she says.

Maiellano contacted her primary care doctor in the University of Rochester Medical Center network, who referred her to a gastrointestinal specialist. The specialist ordered an emergency CT scan, but the cause of her pain remained unclear.

Follow-up imaging that might have provided answers was not available for weeks, leaving her in pain and without a diagnosis. The delay only deepened her worry. 

“It was the time it took to get these tests,” she says. “Not knowing what’s wrong, the pain.”

Eventually, the pain became too much. Maiellano went to the emergency room at Strong Memorial Hospital, where the same cycle continued: waiting, pain, and frustration.

Her experience highlights a central question in contemporary medicine: Can artificial intelligence accelerate patient diagnoses, or does it introduce additional risks to patient care?

At URMC, this question has become tangible.

Serving about 1 million patients annually across Upstate New York and the Finger Lakes, URMC is a vast academic health system centered on the 897-bed Strong Memorial Hospital.

In the radiology department, AI-assisted software supports physicians in dictating reports, drafting summaries of findings, and retrieving pertinent details from patients’ prior imaging histories. The anticipated benefit is a more efficient and organized workflow within a strained system.

However, the inherent risk is significant, as medical errors can have substantial consequences.

Limited scope

The technology implemented at URMC is more limited in scope than the general public’s perception of “AI in health care.” Rather than independently diagnosing patients or replacing radiologists, it functions as software integrated into the existing radiology reporting workflow.

Sean Cleary M.D., a cardiothoracic radiologist at URMC and vice chair of informatics for the Imaging Sciences Department, says the department has long used Microsoft’s PowerScribe 360 as its dictation platform. A radiologist studies images on one screen and dictates the report into PowerScribe on another.

“PowerScribe is basically just the dictation platform,” Cleary says. “The AI part of it only started within the last two years.”

The newer layer, PowerScribe One, adds AI features to the radiology reporting process. One of those features is Smart Impression. In a radiology report, the “findings” section contains the fuller description of what the doctor sees on the scan. The “impression” is the shorter summary at the end that highlights the most clinically important conclusions.

Cleary says Smart Impression does not independently read the scan or make its own diagnosis. Instead, after the radiologist dictates the findings section, the software drafts a first impression based on what the radiologist has already said. 

“The impression section, for lack of a better expression, is like a TLDR (Too Long; Didn’t Read) of the report,” Cleary explains. “What they’re really good at is summarizing information. They’re not doing independent thought.”

URMC is also piloting Microsoft’s Dragon Copilot, which Cleary describes as a tool for summarizing a patient’s earlier radiology reports. For patients with years of prior imaging, Dragon Copilot condenses older reports and links each summarized point back to the original source. 

“It gives you a clickable link to each sentence that it’s getting the summary from,” Cleary says. “So I know what the provenance of that information is.”

The tools described by URMC are used to draft report language, summarize extensive findings, expedite retrieval of prior information, and assist radiologists in managing high workloads more efficiently. Radiologists continue to interpret images, review AI-generated content, edit reports, and provide final approval. 

As Cleary notes, “the ultimate guardrail is the radiologist.”

A strained system

The impetus for adopting AI stems from significant pressures in radiology. Routine MRI appointments may not have an available appointment for weeks, and additional days are required for interpretation.

Cleary says the efficiency gains from AI are incremental, not transformative, estimating improvements of approximately 5 percent. While this may seem minor in individual cases, the cumulative effect over a full day is potentially significant. 

However, radiologist workload efficiencies may not be a direct perceptible benefit to patients, who may continue to experience wait times due to the high volume of studies being performed at imaging centers (or due to imaging centers operating at max capacity).

Gregg Nicandri M.D., URMC’s chief medical information officer, says the institution approaches new AI tools skeptically before they are rolled out.

“We kind of take an evidence-based approach to anything we deploy,” he says. “We want to make sure that it works and it’s safe before we just turn it on.”

Nicandri says the review process begins before contracts are signed. Committees assess whether a tool does what it claims, meets privacy and legal standards, and has cybersecurity protections in place. When large language models are involved, he says, URMC does not want staff to put patient data into public AI tools. 

“We kind of came up with our AI policy, which basically prevents anybody from leveraging any publicly available AI tools with any kind of patient information,” he notes. Instead, the medical center has given employees access to secure internal versions. “Any information that gets put in there, we control and protect. It’s not used to train the larger model.”

Exploring other uses

Report drafting is not the only way AI can be used in radiology.

In one example, Nicandri illustrates how AI can help identify a possible stroke on a scan and move it higher in the reading queue, so it is interpreted sooner by the radiologist. 

“You basically identify opportunity for early intervention where it changes the outcome for the patient,” he says.

Another example is use in mammography, where AI may help identify the need for additional imaging views while the patient is still on site, rather than sending them home to wait for another appointment. 

“Save them the extra visits, save them the anxiety,” Nicandri says.

Then there’s the money. The financial impact of AI in health care is multifaceted.

Implementation requires substantial investment in software, security assessments, internal validation, pilot programs, staff training, and ongoing oversight. However, URMC leaders contend that increased efficiency and the prevention of costly downstream complications can offset these expenses.

For example, Nicandri cites stroke care, where expedited identification of urgent scans may enable earlier intervention and reduce the need for lengthy rehabilitation and extended care. Additionally, AI can decrease the duration of some MRI scans, improving throughput and reducing patient time burdens. 

Thus, the financial rationale extends beyond institutional savings to encompass more effective utilization of time, labor, and treatment resources within a strained health care system. These benefits include expedited prioritization, concise summaries, rapid access to prior information and, in certain cases, earlier clinical intervention. 

Rochester Regional Health is also using AI in image-based care, though in a different context. In February, the system announced the installation of the Ethos Adaptive Radiation Therapy system at Rochester General Hospital. 

RRH says the system uses daily CT imaging, AI and rapid replanning to generate customized radiation treatment plans in real time, adjusting for anatomical changes between sessions. The health system says that physicians still review and approve each plan, “maintaining full clinical decision-making authority,” and that the approach may improve precision while reducing radiation exposure to healthy tissue. 

While URMC described AI as a tool for reporting, summarization and workflow in radiology, RRH has highlighted its use in cancer treatment, where AI helps adapt radiation plans in real time to a patient’s anatomy. 

Risk factors

While the use of AI offers clear advantages, they are accompanied by significant risks. While the technology can summarize, flag, and draft information, it is also susceptible to errors.

The danger with AI, some experts warn, is not just that AI can be wrong. It is that it can be wrong in ways people do not yet fully understand.

Ashique Khudabukhsh, a professor and AI researcher at Rochester Institute of Technology, says that is what worries him most about deploying large language models in health care settings.

PowerScribe One is not a large language model in itself. It is a cloud-based radiology reporting platform that includes AI features, some of which are powered by large language models. Dragon Copilot, by contrast, relies more directly on LLM technology.

“The biggest risk that I see is that when you have systems that are very complex and not easy to understand … they can fail in very unexpected ways,” Khudabukhsh says. “We understand human failures much, much better because we have been doing this for ages.”

He offers a sharp example of how misleading AI fluency can be. 

“You can have LLMs that can write a fancy poem,” he says, “and then cannot count.” In other words, a system can sound polished and convincing while still breaking in strange ways. “What if it just inserts one sentence,” he says, “which maybe the doctor did not intend to say?”

Khudabukhsh points to his own research on how often large language models bend under pressure.

In a recent RIT-led study, researchers tested not just whether popular AI chatbots knew the facts, but whether they would stick to them under conversational pressure. Using a three-step framework called HAUNT, the team first asked chatbots to generate true and false statements about well-known movies and novels, then asked them to verify those statements, and finally nudged them toward the false ones in later exchanges. The result: none of the five models tested were fully self-consistent, and some—especially Gemini and DeepSeek—accepted and repeated falsehoods nearly half the time when subtly prompted.

“Our understanding of AI failures is like an evolving process,” Khudabukhsh says. “We do not have enough time to understand the current things, and new things are appearing.”

That uncertainty becomes unsettling in a field where even a small error can have real human consequences.

Khudabukhsh says large language models also create a subtler problem: overreliance. 

“At some point, there is kind of a cognitive unloading,” he says. “I am transferring the thinking load on the machine.” If that trust deepens because the system performs well most of the time, he adds, “there is a chance that you might miss something.”

Nicandri raised a parallel concern within the medical community itself. One of the larger questions, he says, is how to train younger doctors in an era when AI assistance may become normal.

AI is already part of everyday life for many teenagers, especially in school. A Pew Research Center survey published in February found that 64% of U.S. teens ages 13 to 17 have used AI chatbots, and about three in 10 say they use them daily. More than half say they have used chatbots to search for information, and 54% say they have used them for schoolwork. Pew also found that one in 10 teens say chatbots help with all or most of their schoolwork, while larger shares say they use them for at least some assignments.

“How do we train trainees in the era of AI to not be dependent on AI?” Nicandri says. Faculty at URMC, he says, often learn the tools before residents do, and residents use them under tighter oversight. “You really do need the expert human in the loop.”

Another caution comes from a 2024 study by researchers at Harvard Medical School, the Massachusetts Institute of Technology and Stanford University, which found that AI assistance did not affect radiologists evenly. In a large-scale analysis of 140 radiologists across 15 chest X-ray diagnostic tasks, the researchers found that some clinicians performed better with AI support while others performed worse. The study concluded that “individual clinician differences shape the interaction between human and machine in critical ways that researchers do not yet fully understand,” a finding that suggests AI’s value in radiology may depend not just on the tool itself, but on how a given doctor uses it.

The other risk

A few months after her emergency room visit, Maiellano finally received a colonoscopy. Weeks later, she had an ultrasound.

Her doctors have said the issue may be related to her gallbladder or fructose intolerance, but she still needs more testing before she has a clear answer.

Through it all, the pain has continued.

“I’m used to it, at this point,” she declares.

For doctors at URMC, the argument is not only about the risks of using AI. It is also about the risk of refusing to use it.

Cleary says the radiology department is dealing with rising volumes and a shortage of specialists. “The lack of radiologists is a major driver,” he says. 

Nicandri widens the frame further. “There is a shortage of doctors, period.” He notes that even if the country decided today that it needed many more physicians, it would still take years to train them.

At the same time, Cleary says, radiology has become more central to modern medicine because doctors can see more, catch more, and follow disease more closely than they once could. Cancer patients are scanned more often because treatments have improved. Screening programs, such as lung cancer screening, have increased demand for imaging. Better scanners have widened what radiology can detect. All of that has brought clinical benefit, but has also added pressure to the system.

Cleary argues that AI cannot be judged solely by the danger it poses. 

“Everyone always thinks about the risks of using it,” he says, “but you also have to think about the risks of not using it, because things are gonna get missed as well.” In his view, the technology’s best use is narrow and practical: helping radiologists summarize more accurately, retrieve important health history more quickly, and avoid omitting anything from a report. “It’s a powerful tool for efficiency increase, but ironically, it’s also kind of a patient safety tool.”

Nicandri frames the same idea in terms of process.

“My fear is automating and adopting before it’s proven to be valuable,” he says. “We try to look at, like, what could go wrong, and how do we measure the impact of that? The idea is the AI plus the human should make less errors.”

For patients at URMC, AI in health care is unlikely to arrive in the form of a dramatic machine taking over the exam room.

Cleary says that, from the patient’s perspective, much of the change remains invisible. 

“For the average patient right now, a lot of it’s behind the scenes,” he says.

He also sees a more hopeful future in the technology, one that goes beyond speed. While describing AI’s potential, he spoke about using it at home to create personalized Chinese learning materials for his 7-year-old daughter, tailoring lessons to her interests. In health care, he sees a similar possibility: not AI replacing care, but helping make care more personal and more relevant to the individual patient. “I think we’re going to get there with health care as well.”

Still, both doctors and researchers say the rise of AI in medicine should not be mistaken for a reason to trust AI on its own. Cleary acknowledges that patients are already asking chatbots health questions, and says those tools may sometimes be clearer than a basic internet search. But, he adds, “you still need to talk to your doctor.”

Nicandri warns that consumer AI can become “Dr. Google on steroids.”

Khudabukhsh says patients should be clearly informed when AI is used and should have a say in how much of it affects their care. 

“People should have informed consent,” he says. “The right of the person to choose how much AI they want … should be there.”

Nicandri says that choice is not always simple. In some cases, patients can be asked for permission. In others, AI is increasingly built into the systems that deliver care. “It’s becoming increasingly difficult” to opt out because some of the technology “is part of the way we deliver care.”

Maiellano, an inpatient psychiatric nurse at a local hospital, says she supports the use of AI if it can help patients like her get answers faster while they are in pain. But she worries about doctors becoming too dependent on it, especially because, to her, AI in general is “terrifying.”

“I would like to have faith in the health system,” she says. “Plus, being a nurse, I feel like I need to trust it and be willing to adapt to the changes.”

What all of this means for patients going forward is still being written.

AI may improve efficiency in radiology and help surface urgent findings more quickly, but its use still depends on human review, institutional oversight, and patient trust.

Rob Bell is a Rochester Beacon contributing writer and former Democrat and Chronicle reporter, photographer and editor. He also produces and hosts “Plants & Beats,” a podcast exploring mindfulness, music and culture. 

The Beacon welcomes comments and letters from readers who adhere to our comment policy including use of their full, real name. See “Leave a Reply” below to discuss on this post. Comments of a general nature may be submitted to the Letters page by emailing [email protected].