Why radiologists prefer domain-specific AI over generic AI

Domain-specific AI models can enhance impressions

Radiologists use AI algorithms to help them analyze scans faster and with greater accuracy. AI also helps radiologists summarize images in the impressions section of radiology reports.

The findings section of a radiology report is written for radiologists, but fine-tuning the impressions section of the report with domain-specific AI helps a wider audience of clinicians and patients understand it, noted Nicholas Galante, M.D., musculoskeletal radiologist and medical director for informatics at Radiology Associates of North Texas, which has more than 300 radiologists. Galante was unaffiliated with the npj Digital Medicine study, but has been using Rad AI for five years.

In addition, AI helps radiologists develop differential diagnoses, such as determining if a patient has cancer or an infection. Identifying the subtleties in medical images is important to making differential diagnoses, and the historical context that domain-specific AI has over generic models helps making these differential diagnoses easier, Del Gaizo explained.

Domain-specific AI delivers concise text for radiologists

Conciseness in imaging reports is a key consideration driving whether busy radiologists would consider using an AI application, Del Gaizo noted.

"They really want that impression to get right to the point," he said. "No one has time to read a novel, so the information needs to be summarized pretty well."

In the study, 10 clinicians rated radiology impressions from 200 oncologic CT reports for "completeness, correctness, conciseness, clarity, and clinical utility as well as patient harm." The evaluators then conducted a blinded analysis, consisting of a mix of impressions from human radiologists, custom domain-specific AI and a general-purpose LLM.

 The domain-specific AI model generated impressions that were similar to those of human radiologists in terms of completeness, correctness and conciseness. Evaluators rated the impressions by generic AI models as "significantly less concise" than those from both human radiologists and domain-specific AI models.

"When we ran the study and compared the two, what we saw was that the generic AI models tended to be very verbose, even when specifically prompted to be as concise as possible," Del Gaizo said.

Generic models averaged about 75.1 words per medical image impression compared with around 34.2 words generated by custom AI models and 41.2 words for "original" radiologists without AI.

But if reports are too long and not summarized adequately, the radiologists, clinicians and patients reading them later could be confused, Del Gaizo explained.

Echoing Del Gaizo, Galante added that "radiologists prefer more short and concise presentations of the data and the output."

Galante described the extraneous words in reports as "noise" and a distraction for radiologists.

"In radiology, where it's high-velocity, high-intensity knowledge work, where you know information is coming at you almost like a tsunami throughout your workday, being able to filter it and have something that's concise are very important," Galante said.

For radiologists, saving time when creating reports is a key consideration when deciding to use AI models. Therefore, model conciseness is critical.

"The key to getting radiologists to embrace AI in their clinical workflow is that it's going to save them time," Del Gaizo said. "And to do that, it has to summarize the information in the format that they need it and prefer it."

Further, custom models delivered information to radiologists in a couple of seconds compared with generic models, which took 10-11 seconds because they search a larger database, Del Gaizo said.

In addition to the speed benefits, a custom, or domain-specific, model helps create radiology reports with medical terms that are familiar to surgeons, clinicians and radiologists. It's fine-tuned with a health system's previous data, Del Gaizo said.

He explained that when radiology reports are worded according to clinician preferences, providers can more easily decide what steps to take next with a patient.

Why generic AI models hamper clinicians' workflows

The new study follows 2023 findings in the journal Radiology that showed radiologists without AI produced impressions that were more coherent, comprehensive and factually consistent than impressions generated by GPT-4. The npj study used GPT 4.1 for the generic model.

Del Gaizo noted that generic AI models ultimately created more work for radiologists in terms of editing and modifying summaries. Meanwhile, domain-specific AI creates summaries with more concise language to speed up radiologist workflows. He described the impressions section of a radiology report as a "sweet spot" for AI.

A radiology impression created by a domain-specific AI model could read "there's a liver observation most in keeping with hemangioma," Del Gaizo said, referring to a benign tumor comprising an abnormal cluster of extra blood vessels. Domain-specific models are fine-tuned to word impressions according to clinicians' preferences. However, generic models are wordier and not industry or surgeon-specific, he added.

In addition, generic models can introduce misinformation from sources that are not peer-reviewed, Del Gaizo noted. Although the generic models still provide definitive findings, they may lack historical context from individual radiologists. The missing information could lead to ambiguity as well as reduced clinical utility, Del Gaizo said.

He added that generic AI models could produce hallucinations unrelated to the question a user asked because they are trained on data from the internet. These AI models are prone to guessing, as OpenAI explained in 2025.

However, custom models are fine-tuned with specific information in accordance with the individual radiologist using them, enabling them to develop knowledge of how the radiologist words their reports and impressions, Del Gaizo explained.

The fact that radiologists in the study preferred more specific language generated by domain-specific AI models was unsurprising to Galante.

"The more specialized [the AI model] is with the radiology lexicon, and the more it's trained on radiology data sets, the better it's going to be able to execute those tasks," Galante said. "And I think you would see that with any other highly trained and skilled set of knowledge workers, be it law, engineering or any of the hard sciences or soft sciences."

How domain-specific AI reduces burnout for radiologists

Amid a radiology shortage, companies are looking to make the workday more palatable for radiologists. Recent research from the American Medical Association shows that nearly half of radiologists (45%) are exhausted in the workplace. AI is one technology that could help, Galante said.

The study authors wrote that using domain-specific AI to create impressions helps address clinician burnout.  

"These summaries distill complex imaging findings into actionable clinical insights that guide patient management decisions," the study stated. "However, generating high-quality impressions requires significant cognitive effort and time, contributing to radiologist burnout amid increasing case volumes and workforce shortages."

Galante echoed the researchers, noting that clinicians would lose productivity time if they had to edit and summarize radiology reports on their own without AI.

"So that was really the impetus for [using AI], is staving off burnout, so that you can continue to be in the workforce and deliver high-quality care for patients," Galante said.

A next step for researchers could be to study impressions for conditions beyond oncology, including how the AI models perform for neurosurgeons or neurologists, Del Gaizo said. He also sees domain-specific AI being applied to the work of ER physicians and primary care providers and perhaps even helping patients directly by creating easy-to-understand summaries without jargon.

Today, radiologists use AI models to summarize their text descriptions of what they see on an image, but in the future, the AI models could summarize the radiology images themselves, according to Del Gaizo.

"The technology is going in the direction with foundation models where AI will be able to look at that full CT, oncologic CT, and provide a summary or suggested findings that should go into the report," Del Gaizo said.

In the end, being able to manage the volume of radiology impressions more easily will help address delayed patient care, longer waiting lines and delays in diagnosis, Galante further noted.

"You want to be able to handle this volume in a way that you know is still safe and effective, and in a way you're not sacrificing quality," Galante said. "And ideally, you're actually improving your quality along with your efficiency and productivity."

Brian T. Horowitz started covering health IT news in 2010 and the tech beat overall in 1996.