FDA gives generative AI in radiology two breakthrough designation nods

STAT News – Thursday, June 25, 2026
By Katie Palmer

The Food and Drug Administration has granted breakthrough designation to two devices that use generative AI to interpret chest X-rays and draft the radiology reports typically written by human radiologists.

Machine learning systems have long analyzed images like X-rays and CT scans. But more recently, large vision language models have ushered in a new capability. Instead of highlighting a spot for a radiologist to review and write up, generative AI can process the entire image and draft many of its findings for a radiologist to review — a technological advancement that is challenging traditional validation and regulatory frameworks.

In March, one breakthrough designation went to Cognita, a Stanford researcher-founded startup acquired late last year by the large radiology practice Radiology Partners. Radiology AI company Aidoc announced its own breakthrough designation Thursday for a tool called First Read, specifically when it is used to detect and describe four life-threatening findings.

With nearly 1,300 products, the FDA’s breakthrough program is intended to support the development of medical devices that could better treat or diagnose serious conditions by providing more regular agency communication, flexible clinical trial design when appropriate, and prioritized review. The FDA does not typically share — with sponsors or the public — why their devices have qualified for breakthrough designation.

Cognita CEO Louis Blankemeier thinks generative AI-based draft reporting will solve both workload and performance challenges in radiology.

“The FDA realizes we need some breakthrough here to add capacity to the radiology workforce,” he said. Radiological imaging is exploding, but the number of radiologists isn’t growing fast enough to keep up, leaving some patients with significant delays getting their images interpreted. At the same time, Blankemeier imagines the technology will be more comprehensive than existing AI tools in radiology. “Some of these findings are really critical, more rare, and not necessarily covered by point solutions,” he said, “which will have an enormous impact on patients.”

Cognita and Aidoc are far from the only companies chasing the challenge of automated draft reporting in radiology, each leveraging access to unique databases of clinical images from a combination of customers, business partners, and academic medical centers. A startup called Voio launched in November with an open-source foundation model trained on CT scans and MRIs from the University of California, San Francisco. In early June, radiology AI company Harrison.ai released its own draft-report-generating foundation model for research purposes. Both Cognita and Aidoc are developing foundation models for CT scans in addition to chest X-rays.

While none of these tools have been authorized by the FDA, they are already in use at radiology practices today as part of Institutional Review Board-approved research studies, including at Radiology Partners practices and with Aidoc customers.

In the next month, Aidoc customer Hartford Healthcare plans to begin trialing First Read in its emergency departments and ICUs. “It’s the place where it’s going to have the biggest impact right away in the hands of clinicians that are probably the most comfortable looking at X-rays already,” said Hartford radiologist and chief clinical innovation officer Barry Stein.

“The industry is going incredibly fast,” said Sham Sokka, chief operating and technology officer at RadNet’s DeepHealth. “Not just build, but actual deployment,” in particular in teleradiology and outpatient imaging practices.

But the expansive technology presents significant challenges for existing AI validation and regulatory frameworks. “The most important thing for this next generation of foundation model is going to be quality and accuracy,” said Elad Walach, CEO and co-founder of Aidoc. “One of the toughest challenges we have is how to get a model to be accurate enough for comprehensive coverage.”

Most AI-based tools in radiology are validated based on their ability to detect or diagnose an individual finding with high sensitivity. Because draft radiology reports created with generative AI have the potential to identify thousands of findings, “we’re now shifting with report generation from that type of application to one where you have to be highly specific, as well,” Sokka said.

That scale means draft-report findings can’t be tested the same way as one-off AI devices. “We can’t do every one,” Blankemeier said. “That would be infeasible.” A draft finding also includes many more elements to validate than a typical AI detection or triage device. Add in elements such as location descriptors, severity, and changes between images, and the validation challenge gets even more difficult.

“On top of accuracy, are they acceptable? Are there any significant findings? Are they missing anything? Are they hallucinating? Do I agree with them? How do I look at the quality of their outputs? And then how do I rank them? These are all subjective,” said Eun Kyoung (Amy) Hong, an assistant professor of radiology at Stanford who has evaluated more than 10 draft-generating foundation models.

In general, Hong has found that the models are weak at measuring findings and locating them precisely, and they can also still introduce hallucinations and bias — though there is significant variation between models. “They have very different personalities and different niches,” she said, in large part depending on where their training data originates, including outpatient centers, emergency rooms, and ICUs.

While she sees draft report generation as an inevitable step for radiology, she wants to ensure that the technology ensures patients’ safety. “With this report generation system, with all these risks of automation bias, aren’t we not diluting our expertise by just accepting some mediocre AI generated report?” she asked. “I’m not the judge, but I want everyone to at one point think about this question and see for themselves.”

Breakthrough device designation gives both companies regular access to the FDA as it defines its regulatory approach for this and other generative AI-based medical devices. Blankemeier said the designation allows each development cycle to take weeks instead of months. “We all share the concerns that this device is bringing new risks to the world,” said Walach, including the FDA. “We feel there’s meaningful regulatory innovation to do it in a way that is fast, while upholding the same or even greater standards of safety.”

As companies continue to drive toward a more automated future of radiology reporting, they know the bar for safe, accurate, and effective AI-based reports is different for the FDA and real-world radiologists. In order to meaningfully tackle radiology’s labor crunch, a draft report doesn’t just need to be correct: It needs to help radiologists work faster and better than they currently do. Many radiology detection and triage tools have failed to be implemented, because they throw up too many alerts that radiologists need to double-check, ultimately costing them time.

“The accuracy bar for effective augmentation is actually very, very high,” Walach said. “If you throw in something that just generates simple cases, or you don’t trust it, then you actually are not reaping any of the benefits.”

The tools will also need to be meaningfully incorporated into radiologists’ workflows. With AI-drafted reports, “the job of the radiologist goes from dictating the full report to just QA-ing, making edits, and reviewing statements,” said Blankemeier. Given the length of some of those reports, “this is a very challenging problem.”

Generative AI-based report-drafting is being developed at a time when many radiology practices are reassessing their technologies. Radiologists have long used transcription tools to dictate their reports, but this summer, Microsoft will end support for its popular tool PowerScribe 360, leaving many users with an opportunity to choose new AI-based tools for reporting.

Though companies cannot commercialize image-to-text draft-reporting tools until they are authorized by the FDA, practices that are considering new tech vendors are keeping an eye on that future, said DeepHealth’s Sokka. “Even though many of them are not there yet,” he said, “they’re trying to figure out who is ready, who’s going to get there.”

Access this article at its original source.